# Acaysia — full site corpus > Single-file Markdown snapshot of every page on https://acaysia.com for long-context agents. The per-page index is https://acaysia.com/llms.txt; each page is also served individually at the .md variant of its URL. --- # Control the Impossible Canonical: https://acaysia.com/ Optimization for physical processes. Higher yield, lower energy, failsafe in under 100 milliseconds. Proven first on chemical reactors. [Book a discovery call](https://acaysia.com/contact.md) ## Key metrics - **>2%** — Yield Improvement - **~10%** — Energy Reduction - **<100ms** — Failsafe Response ## The problem ### You've optimized this reactor before. You'll do it again next year. Every plant has been through advanced control projects. Most have working APC on at least some loops. The problem isn't that optimization doesn't exist. It's that it doesn't compose. Each project rebuilds the same primitives. The model is bespoke. The controller doesn't transfer. The improvement decays as the process drifts. Eighteen months and several million dollars later, you have a slightly better version of what you started with, and the next reactor needs its own project from scratch. A lot is being left on the table. ## How it works ### Optimization that reads the process Acaysia describes your process as data, then optimizes against it in real time. The same four steps run whether the unit operation is a reactor, a column, or a dryer. #### Step 01 — Model the process Acaysia builds a model of your process from your plant data and known physics. That model is what the controller optimizes against. #### Step 02 — Optimize in real time Every cycle, the controller simulates thousands of control trajectories on the model and picks the best. This is MPPI, model predictive path integral control. It optimizes for yield, energy, and throughput inside your limits. #### Step 03 — Fail safe to PID On any fault the controller hands back to your existing PID in under 100 milliseconds. You hold the safety envelope the whole time. #### Step 04 — Improve on your data Models retrain on your process history on-premises, with versioned rollback. Performance climbs without data ever leaving the plant. ## Proof ### Proven in chemical reactors Continuous stirred-tank reactors are where Acaysia runs today. The full control system runs on real reactor hardware, so the yield and energy numbers above come from a running CSTR, not from simulation. Control is handed over in stages. The system runs in shadow and advisory first, so it never gets more autonomy than it has earned. [See the deployment journey](https://acaysia.com/product.md) ## Why it generalizes ### One architecture, every unit operation PID controls every process in every industry because it runs on one generic idea, an error signal. Acaysia works the same way. It runs on a generic description of a process, so a new unit operation is not a new project. You describe it the same way as a reactor, and the same controller runs it. **Every unit operation, described the same way:** | Aspect | Example | | --- | --- | | What has to balance | Mass and energy, in and out | | Where it connects | Feeds, products, and utilities | | What you measure | Temperature, pressure, level | | What you control | Flows, valves, heating and cooling | - **Controller** — Reads the ontology. The same controller runs every unit operation. - **Ontology** — One shared description for every unit operation. - **Unit operation classes** — CSTR, Batch reactor, Plug flow reactor, Distillation column, Crystallizer, Evaporator. One ontology, one controller. Each new unit operation joins the same system. We do not rebuild the controller. ## Safety and integration ### Safe to put in a real plant Acaysia sits next to your PLC and respects every interlock it already has. The safety architecture is unit-op-agnostic by construction. The same Trust Arbiter that supervises a reactor will supervise a column or a crystallizer. #### Supervision — Trust Arbiter Every control move is checked against your limits before it reaches the plant. Anything outside the envelope is blocked. #### Fallback — PID under 100ms On any fault the controller reverts to your proven PID in under 100 milliseconds. No gap in control. #### Standards — ASIL-D inspired The safety architecture is modeled on automotive ASIL-D and is SIL compatible. It never interferes with your Safety Instrumented Systems. #### Integration — Brownfield drop-in OPC UA and EtherNet/IP connect to the PLCs you already run. IEC 62443 cybersecurity considerations throughout. No rip and replace. ## Works with the infrastructure you already have - [Siemens](https://www.siemens.com/global/en/products/automation/systems/industrial.html) - [Rockwell Automation](https://www.rockwellautomation.com/en-us.html) - [Beckhoff](https://www.beckhoff.com/en-us/) - [Schneider Electric](https://www.se.com/us/en/) - [ABB](https://new.abb.com/control-systems) - [OPC UA](https://opcfoundation.org/) ## Take on a real process with us We are taking on a small number of pilot partners through 2026 across chemicals, pharmaceuticals, and adjacent process industries. If you operate reactors, columns, or other continuous or batch unit operations and want to evaluate Acaysia on a real process, talk to us. [Book a discovery call](https://acaysia.com/contact.md) Partners with: [NVIDIA](https://www.nvidia.com/), [a16z speedrun](https://speedrun.a16z.com/) --- # The Acaysia Control System Canonical: https://acaysia.com/product Enterprise-grade edge compute appliance with MPPI optimization for physical processes. Proven first on chemical reactors. Drop-in deployment, millisecond failsafe, continuous learning. [Book a discovery call](https://acaysia.com/contact.md) ## NVIDIA Jetson Orin Edge Compute Enterprise-grade edge compute in a compact, industrial form factor. ## Built for Operations Technology Purpose-built for the unique requirements of physical process control. Proven first on chemical reactors. ### Compute — Edge Computing NVIDIA Jetson Orin delivers deterministic, low-latency inference. Complete air-gapping available for sensitive environments. ### Optimization — MPPI Control Model Predictive Path Integral (MPPI) brings advanced process control (APC) to unit operations, computing trajectories under constraints while PIDs execute steady-state control. ### Visibility — Operator Dashboard Transparent interface showing setpoints, constraints, and confidence. Historian replay and one-click rollback included. ### Safety — Failsafe Design Automatic reversion to plant PID on any fault. Multi-layer watchdogs and health monitors ensure continuous safe operation. ### Learning — Continuous Learning On-premises time-series storage with secure export for retraining. Reactor-specific models with versioned rollback, trained at scale on the [AcaysiaRT high-throughput GPU simulation runtime](https://acaysia.com/engines/acaysia-rt.md), the [AcaysiaDRT 3D spatial simulation engine](https://acaysia.com/engines/acaysia-drt.md), and [Rete](https://acaysia.com/engines/acaysia-rt.md#rete) for whole-plant flowsheets with recycle. ### Connectivity — Standard Protocols OPC UA and EtherNet/IP connectivity to plant DCS/PLC systems. Historian hooks for full auditability. ## Vendor Integration - [Siemens](https://www.siemens.com/global/en/products/automation/systems/industrial.html) - [Rockwell Automation](https://www.rockwellautomation.com/en-us.html) - [Beckhoff](https://www.beckhoff.com/en-us/) - [Schneider Electric](https://www.se.com/us/en/) - [ABB](https://new.abb.com/control-systems) - [OPC UA (OPC Foundation)](https://opcfoundation.org/) ## System Architecture Edge AI-powered control with multi-layer safety and continuous learning capabilities. Detailed specifications available under NDA. - [Interactive Architecture](https://acaysia.com/architecture/) - Safety Docs (access-gated — request access at founders@acaysia.com) ## Deployment Journey From observation to optimization in controlled stages. ### Phase 01 — Shadow Mode Install edge unit near PLC and connect via standard protocols. System observes reactor dynamics for 2-4 weeks, learning behavior without touching control. Compare AI predictions against actual PID performance. ### Phase 02 — Advisory Mode AI provides optimization recommendations while operators maintain full control. Review predictions, build confidence in the model, and refine safety constraints based on plant-specific requirements. ### Phase 03 — Closed-Loop Control Activate autonomous operation within your defined safety envelope. System co-manages control with PID, automatically reverting on any fault or operator request. Full audit trail and event logging maintained. ### Ongoing — Continuous Optimization Models improve with every batch cycle. On-premises retraining with versioned rollback ensures performance improves continuously while maintaining safety guarantees. ## See Acaysia in Action Schedule a personalized demo to see how our intelligent control system can optimize your specific process. [Book a discovery call](https://acaysia.com/contact.md) --- # About Acaysia Canonical: https://acaysia.com/about Building intelligent control systems for physical processes. Proven first on chemical reactors. Safe and interpretable by design. ## The Problem We Solve Physical processes drift. Feedstock changes, fouling, and ambient conditions push reactors, columns, and dryers away from their optimum. Keeping them there requires constant manual intervention. We founded Acaysia because we believe that problem is solvable with intelligent, adaptive control. Chemical plants are where we proved it first. ## Our Approach We build on a generic abstraction of physical processes. Conserved quantities, ports, state variables, and manipulated variables. A chemical reactor is one subclass. So is a distillation column, a crystallizer, or a fermenter. We start with hybrid physics-informed models to give operators transparency and confidence. Safety is always the top priority. Concretely, the controller is driven by an MPPI (Model Predictive Path Integral) core algorithm, while a Trust Arbiter supervises every control move and reverts to a deterministic PID fallback in under 100 milliseconds if anything looks wrong. The design is ASIL-D inspired and SIL compatible, and it never interferes with your existing safety instrumented system. It installs as a brownfield drop-in over OPC UA and EtherNet/IP, so it works alongside the controls you already run. ## Mission and Vision ### Mission — Unlock Consistent Performance To unlock consistent performance improvements in physical process industries through intelligent, safe, and adaptive control systems that work alongside existing infrastructure. No disruption, just results. ### Vision — Every Process at Its Optimum A future where every physical process operates at its theoretical optimum, continuously learning and adapting while maintaining the highest standards of safety and reliability. ## Core Values ### Safety First We never compromise on process safety. Every design decision prioritizes failsafe operation and seamless integration with existing safety systems. ### Engineering Rigor We combine deep domain expertise with cutting-edge ML. Our solutions are built by engineers who understand both the chemistry and the code. ### Customer Partnership We succeed only when our customers succeed. We build long-term partnerships focused on measurable outcomes and continuous improvement. ### Transparency Every decision our system makes is interpretable and auditable. We believe operators deserve to understand what their control system is doing and why. ## Why Acaysia The name Acaysia is inspired by the acacia tree, a symbol of precision, growth, and stability. Like the acacia that adapts to harsh environments, our control systems help industrial processes adapt to changing conditions while maintaining stability and efficiency. ## Industries we work with ### Specialty and Fine Chemicals Intermediates, APIs, custom synthesis, polymers. ### Pharmaceutical Manufacturing Continuous manufacturing, crystallization, API reactors. ### Mining and Minerals Dryers, washers, classifiers, screens. ### Industrial Biotech and Precision Fermentation Bioreactors, downstream processing. If your process has a conserved quantity and a manipulated variable, we can model it. ## Work with Us We are building the future of process control. If you run a physical process and want consistent performance without constant manual intervention, let us talk. [Book a discovery call](https://acaysia.com/contact.md) --- # Get in Touch Canonical: https://acaysia.com/contact Ready to optimize your physical processes? Let us discuss how Acaysia can help your operation. ## Email Us For demos, partnerships, careers, or general inquiries. [founders@acaysia.com](mailto:founders@acaysia.com) ## Send a Message The page provides a contact form (submitted via `POST /api/contact`) with the following fields: - **Name** (required) - **Email** (required) - **Company** (optional) - **Message** (required) --- # AcaysiaCORE Canonical: https://acaysia.com/core **Coming Soon** Simulation as a service. Four engines that today need a CUDA toolchain, a GPU and a working knowledge of each one's API — brought into a single browser application, with the hardware on our side of the wire. Built for process engineers, control engineers, plant operators, and academics who want to *run* these models, not build them. [Request Early Access](https://acaysia.com/contact.md) Diagram: four engine domains — AcaysiaRT (127 process models), Rete (plants with recycle), AcaysiaDRT (3D spatial fields) and AcaysiaGEM (vessel geometry) — converging through one contract family into one AcaysiaCORE surface ("one surface, one login"), which serves a single browser. Four engines, four very different kinds of compute — a batched GPU runtime, a flowsheet marcher, a lattice-Boltzmann solver and a CAD kernel — behind one API, one job format and one set of charts. ## Everything, From One Tab Each of these is real output from the engine that produces it — not a mockup of an interface. CORE is the thing that gets you to them without an install. - **[AcaysiaRT](https://acaysia.com/engines/acaysia-rt.md) — move a slider, watch an ensemble.** Results are reduced to quantile envelopes on the GPU before they cross the network, so the interactive path stays fast without shipping raw tensors to a browser. - **[Rete](https://acaysia.com/engines/acaysia-rt.md#rete) — wire a plant, turn its knobs.** Build a flowsheet on a canvas, inspect the definition as plain JSON, then watch per-unit responses and steady-state settle as you change a duty. - **[AcaysiaDRT](https://acaysia.com/engines/acaysia-drt.md) — submit a 3D run, get fields back.** Where the other engines treat a vessel as one well-mixed point, DRT resolves it as a grid and marches every cell. Jobs are rendered on the GPU box and come back as fields and frames rather than a file you have to post-process. - **[AcaysiaGEM](https://acaysia.com/engines/acaysia-gem.md) — shape the vessel, export it to the solver.** Set the parameters, see the geometry in 3D, check the boundary regions, and send it to a DRT run — without a CAD licence or a meshing step in between. ## Where the GPUs Actually Sit The browser talks to a control plane; the control plane mediates work through storage. The compute tier is **egress-only** — workers dial outward and nothing ever connects *to* a GPU box. Architecture schematic: a browser application talks to a control plane, which mediates jobs through object storage; GPU workers dial outward to the control plane and are never connected to directly. Batch work is claimed atomically from storage, so a worker dying mid-run means a peer takes over rather than a job lost. Interactive sessions hold an outbound socket and are stateless-recoverable — reconnect and re-run. ## CORE Is Not a Control Tool **Read this first.** This is worth stating before anything else, because the distinction is a safety boundary rather than a positioning choice. AcaysiaCORE is an **analysis and exploration** application. It runs simulations, shows you what they did, and lets you compare them. It does not touch a plant. The planned desktop connector is **read-only data acquisition** — connect, browse, subscribe — so you can lay real plant behavior over a simulation and see where they diverge. There is no write surface, and its absence is enforced by test rather than by policy. Closed-loop control is the [Acaysia control system](https://acaysia.com/product.md), which is a separate product with a separate safety architecture. ## How It's Built ### Provenance — Trust tiers come from the engine Every model in the catalog carries its validation standing — emitted directly from [AcaysiaRT's](https://acaysia.com/engines/acaysia-rt.md) own registry and T0–T5 scorecard, never hand-copied into a marketing table. If a model's evidence changes in the engine, it changes in the app. ### Isolation — Compute never accepts connections GPU workers dial *outward* to the control plane and claim batch work from storage. Nothing ever connects inbound to a compute box, and job artifacts are tenant-isolated structurally rather than by convention. ### Durability — Crash anywhere, resume Batch jobs are claimed atomically, with queue pointers that live until the job reaches a terminal state — so a worker dying mid-run means a peer takes over, not a job lost. Interactive sessions are stateless-recoverable: reconnect and re-run. ### Contracts — Versioned schemas, drift-tested The wire contracts are the source of truth, with JSON Schemas emitted from them and tested against drift. Interactive results and archived batch results share one shape, which is why the same chart component renders both. ## Who It's For ### Process engineers — Ask harder questions Sweep an operating envelope instead of running one case at a time, and see whether a lumped assumption survives contact with a spatially resolved model. ### Control engineers — Test against a fast plant A model fast enough to sample thousands of futures per tick is a model fast enough to develop a controller against, long before anything touches hardware. ### Operators — Compare plant to model Overlay recorded plant behavior on a simulated run — read-only — and see where the real unit has drifted from the one on paper. ### Academics — Reproducible by construction A scenario is a versioned artifact. Runs carry the engine build they executed on, so a result is something a reader can re-run rather than take on faith. ## Want In Early? AcaysiaCORE is in build and not yet generally available. If you have a process you would want to run through it — or a workflow it should support — we would like to hear about it while the shape is still being decided. Email [aias@acaysia.com](mailto:aias@acaysia.com), or use the [contact form](https://acaysia.com/contact.md). --- # Resources Canonical: https://acaysia.com/resources Technical documentation, case studies, and insights on intelligent process control. ## Guides - [What is advanced process control?](https://acaysia.com/resources/advanced-process-control.md) — Start here if APC is new to you: what it is, what it does for yield, energy, and throughput, and where MPPI fits in. - [MPPI vs PID vs MPC](https://acaysia.com/resources/mppi-vs-pid-mpc.md) — How Model Predictive Path Integral control compares to PID and classical MPC, and why sampling handles the nonlinear chemical reactor dynamics that linearized solvers struggle with. ## Technical Documentation - PLC Integration Guide: OPC UA (access-gated) — Step-by-step guide for connecting Acaysia to your existing PLC infrastructure via OPC UA and EtherNet/IP protocols. - MPPI Control: Theory and Practice (access-gated) — Deep dive into Model Predictive Path Integral control algorithms and their application to physical process optimization. - Safety Architecture Overview (access-gated) — ASIL-D inspired safety design, SIL compatibility, and multi-layer failsafe architecture documentation. - Deployment Checklist (access-gated) — Complete pre-deployment, installation, and validation checklist for Shadow Mode through Closed-Loop Control. - API Documentation (access-gated) — Full REST API reference for the Acaysia control system, including historian access and configuration endpoints. More documentation coming soon. ## Case Studies (Modeled Examples) - Specialty Chemicals: 1.8% Yield Improvement (access-gated) — A modeled specialty-chemicals reaction, and the yield MPPI control recovers from it. - Petrochemical Refinery: Energy Reduction (access-gated) — What adaptive MPPI control does to energy consumption in a modeled refinery operation. - Pharmaceutical: Batch Consistency (access-gated) — How physics-informed ML tightens batch-to-batch consistency in a modeled pharma process. More case studies in development. ## Whitepapers - Gray-Box vs Black-Box Models in Process Control (access-gated) — Comparing physics-informed and pure data-driven approaches to chemical reactor modeling. - The Economics of 1% Yield Improvements (access-gated) — Quantifying the financial impact of small yield improvements at scale in chemical manufacturing. - Federated Learning for Chemical Processes (access-gated) — How federated learning enables cross-plant model improvement without sharing proprietary process data. - ISA/IEC 62443 Compliance Guide (access-gated) — Meeting industrial cybersecurity standards for AI-powered control systems in chemical manufacturing. ## Blog & Updates - Introducing Acaysia (access-gated) — Our founding story and the vision for intelligent physical process control. - Why Chemical Plants Need Adaptive Control (access-gated) — The case for moving beyond fixed PID parameters to ML-guided adaptive control. - Drop-In PLC Integration via OPC UA (access-gated) — Technical deep-dive into how Acaysia connects to industrial PLCs without disruption. Access note: the documentation, case study, whitepaper, and blog links above contain patent-pending technology and are available by access code only — requests without a code receive HTTP 401. To request access, [contact us](https://acaysia.com/contact.md) or email founders@acaysia.com; enter your code at https://acaysia.com/resources. ## Frequently Asked Questions ### What does Acaysia do? Acaysia is a control system that optimizes physical processes for yield, energy, and throughput. It models each unit operation as a typed object and computes control moves in real time. It is proven on real continuous stirred-tank reactor hardware, and chemical manufacturing is its first vertical. ### Why does Acaysia work across different processes? Acaysia runs on a typed compositional ontology of unit operations. Each process is described by conserved quantities, ports, state variables, and manipulated variables. The controller reads the ontology rather than hard-coding one process, so the same architecture extends from reactors to columns to crystallizers. ### What types of unit operations does Acaysia support? Acaysia runs on continuous stirred-tank reactors today. The same architecture extends to batch reactors, plug flow reactors, distillation columns, crystallizers, and evaporators. Chemical manufacturing is the first vertical. ### How does MPPI control differ from traditional PID control? PID reacts to the error it sees right now. MPPI (Model Predictive Path Integral) control couples machine learning with predictive optimization to anticipate where the reactor is heading and pick the best move before a disturbance lands. The payoff is higher yield, lower energy use, and faster recovery from upsets, with PID still underneath as the safety net. ### Is Acaysia compatible with existing PLC systems like Rockwell and Siemens? Yes. Acaysia connects to existing PLCs from Rockwell, Siemens, and other major vendors over the standard protocols, OPC UA and EtherNet/IP. No rip-and-replace; it drops in alongside what you already run. ### What is Shadow Mode and how does deployment work? Shadow Mode is the first phase of deployment where Acaysia observes your reactor operations without making control decisions. This allows the ML model to learn your specific process. The system then progresses through Advisory Mode (recommendations only) before optional Closed-Loop Control with full automation. ### How quickly can the failsafe system respond to issues? Acaysia's failsafe responds in under 100 milliseconds, automatically reverting to proven PID control or a safe shutdown when it detects an anomaly. ### Is it safe to run in a real plant? Yes. A Trust Arbiter checks every control move against plant limits, and the system falls back to existing PID control in under 100 milliseconds on any fault. The safety architecture is ASIL-D inspired and SIL compatible and never interferes with Safety Instrumented Systems. ### What kind of ROI can we expect from implementing Acaysia? Modeled scenarios show 1%+ yield improvement and 2-5% energy reduction. For a mid-size chemical plant, this translates to hundreds of thousands of dollars in annual savings, with payback periods often under 12 months. ## Need more? Can't find what you're looking for? Talk to the team. We'll answer technical questions or point you to the right doc. [Contact Us](https://acaysia.com/contact.md) --- # Terms of Use Canonical: https://acaysia.com/terms Effective August 21, 2026 These Terms of Use (the "**Terms**") form a legally binding agreement between you and Acaysia, Inc. ("**Acaysia**", "we", "us", or "our") and govern your access to and use of [acaysia.com](https://acaysia.com/index.md) and any related services we provide (the "**Site**"). Please read these Terms carefully. By accessing or using the Site, you agree to be bound by these Terms and by our [Privacy Policy](https://acaysia.com/privacy.md). If you do not agree, do not access or use the Site. ## 1. Eligibility You must be at least eighteen (18) years of age and have the legal capacity to enter into a binding contract to use the Site. By using the Site, you represent and warrant that you meet these requirements. If you access the Site on behalf of an entity, you represent that you are authorized to bind that entity to these Terms, and "you" refers to that entity. ## 2. License to Use the Site Subject to your compliance with these Terms, Acaysia grants you a limited, revocable, non-exclusive, non-transferable, non-sublicensable license to access and use the Site solely for your personal information and lawful business inquiry purposes. Any other use requires our prior written consent. ## 3. Prohibited Conduct You agree not to, and not to attempt to: 1. Access or use the Site in violation of any applicable law or these Terms. 2. Use any automated means, including bots, scrapers, crawlers, or spiders, to access the Site, except as expressly permitted by our robots.txt and only in compliance with applicable law. 3. Reverse engineer, decompile, disassemble, or otherwise attempt to derive the source code, underlying ideas, or algorithms of the Site or any part of it, except to the limited extent permitted by mandatory applicable law. 4. Probe, scan, test the vulnerability of, or attempt to breach the authentication, authorization, or security measures of the Site or any related system. 5. Interfere with or disrupt the integrity or performance of the Site, including by overloading, flooding, spamming, or otherwise overwhelming Site infrastructure. 6. Impersonate any person or entity, or misrepresent your affiliation with a person or entity. 7. Submit content that is unlawful, infringing, defamatory, harassing, fraudulent, malicious, or that contains personal information of third parties without authorization. 8. Upload or transmit viruses, malware, or any other malicious code. 9. Frame, mirror, or otherwise reproduce or display the Site or any portion of it without our prior written consent. 10. Collect, harvest, or use the personal information of others through the Site. 11. Use the Site to send unsolicited communications, promotions, or advertisements. 12. Encourage, facilitate, or instruct any third party to engage in any conduct prohibited by these Terms. ## 4. Intellectual Property ### 4.1 Site content The Site and its content, including without limitation text, graphics, images, software, code, designs, logos, trademarks, audio, video, data, and the selection and arrangement thereof (collectively, "**Site Content**"), are owned by Acaysia or its licensors and are protected by United States and international copyright, trademark, trade secret, and other intellectual property laws. Except for the limited license expressly granted in these Terms, no rights are granted to you with respect to the Site Content. ### 4.2 Trademarks "Acaysia", the Acaysia logo, and other Acaysia product names, service names, and slogans are trademarks of Acaysia. You may not use these marks without our prior written permission. Third-party trademarks displayed on the Site are the property of their respective owners and do not imply endorsement. ## 5. User Submissions When you submit information, materials, or other content to us through the Site, including via the contact form ("**Submissions**"), you represent and warrant that: 1. you own or otherwise have the right to grant the rights and licenses described in these Terms, and your Submission does not infringe or violate the rights of any third party; 2. your Submission does not contain confidential information of any third party that has not authorized its disclosure to us; and 3. your Submission complies with these Terms and all applicable laws. You grant Acaysia a worldwide, royalty-free, perpetual, irrevocable, non-exclusive, sublicensable license to use, reproduce, modify, adapt, publish, translate, distribute, and display your Submission solely for the purpose of operating the Site, responding to your inquiry, and conducting our business in accordance with our Privacy Policy. We do not claim ownership rights in your Submission beyond the license described in this Section. **Do not send us confidential or trade-secret information through the Site.** Any information you submit will be handled in accordance with our Privacy Policy, and we accept no obligation of confidentiality with respect to your Submission unless you have a separate written agreement with us that establishes such an obligation. ## 6. Privacy Our [Privacy Policy](https://acaysia.com/privacy.md) describes how we collect, use, and disclose personal information. By using the Site, you acknowledge our Privacy Policy. ## 7. Third-Party Links and Services The Site may contain links to, or integrate with, third-party websites, services, or resources that are not owned or controlled by Acaysia. We do not endorse and are not responsible for the content, products, services, or practices of any third party. Your use of any third-party site or service is subject to that third party's terms and policies. ## 8. Disclaimers To the maximum extent permitted by applicable law, the site and all site content are provided "as is" and "as available" without warranty of any kind, whether express, implied, statutory, or otherwise. Acaysia expressly disclaims all warranties, including warranties of merchantability, fitness for a particular purpose, title, and non-infringement, and any warranties arising from a course of dealing, usage, or trade practice. Acaysia does not warrant that the site will be uninterrupted, timely, secure, or error-free, that defects will be corrected, or that the site is free of viruses or other harmful components. Any information on the site is provided for general informational purposes only and does not constitute engineering, legal, financial, or other professional advice. Nothing in these terms excludes or limits any warranty or liability that may not, as a matter of applicable law, be excluded or limited. Some jurisdictions do not allow the exclusion of certain warranties, so some of the above exclusions may not apply to you. ## 9. Limitation of Liability To the maximum extent permitted by applicable law, in no event will Acaysia, its affiliates, or their respective directors, officers, employees, agents, or licensors be liable to you for any indirect, incidental, special, consequential, exemplary, or punitive damages, including damages for lost profits, lost revenue, lost data, or business interruption, arising out of or in connection with these terms or your use of the site, whether based on contract, tort (including negligence), strict liability, or any other legal theory, and whether or not Acaysia has been advised of the possibility of such damages. To the maximum extent permitted by applicable law, Acaysia's total cumulative liability arising out of or relating to these terms or your use of the site will not exceed one hundred United States dollars (USD 100). Multiple claims will not enlarge this limit. The limitations in this section apply even if any limited remedy fails of its essential purpose. Nothing in these terms excludes or limits liability for fraud, fraudulent misrepresentation, gross negligence, willful misconduct, or any other liability that cannot be excluded or limited under applicable law, including, where applicable, liability for death or personal injury caused by negligence. ## 10. Indemnification To the maximum extent permitted by applicable law, you agree to defend, indemnify, and hold harmless Acaysia, its affiliates, and their respective directors, officers, employees, agents, and licensors from and against any and all claims, damages, losses, liabilities, costs, and expenses (including reasonable attorneys' fees) arising out of or relating to (a) your use of or access to the Site, (b) your violation of these Terms, (c) your violation of any law or the rights of any third party, or (d) any Submission you make to the Site. ## 11. Termination We may suspend, terminate, or restrict your access to the Site at any time, with or without notice and with or without cause, including for any violation of these Terms. Upon termination, all licenses and rights granted to you under these Terms will immediately cease. Sections 4, 5, 6, 8, 9, 10, 12, 13, 14, and 15 will survive any termination or expiration of these Terms. ## 12. Governing Law These Terms, and any dispute arising out of or relating to these Terms or your use of the Site, will be governed by and construed in accordance with the laws of the State of Delaware, United States, without regard to its conflict-of-laws principles. The United Nations Convention on Contracts for the International Sale of Goods does not apply. ## 13. Dispute Resolution; Arbitration; Class Action Waiver ### 13.1 Informal resolution Before filing any claim against Acaysia, you agree to attempt to resolve the dispute informally by contacting us at [founders@acaysia.com](mailto:founders@acaysia.com) and providing a written description of the dispute and the relief sought. The parties will negotiate in good faith for at least thirty (30) days before either party may initiate a formal proceeding. ### 13.2 Binding arbitration Except as set forth below, any dispute, claim, or controversy arising out of or relating to these Terms or your use of the Site that cannot be resolved through informal resolution will be resolved by binding arbitration administered by JAMS in accordance with its Streamlined Arbitration Rules and Procedures, or, if the amount in controversy exceeds USD 250,000, JAMS' Comprehensive Arbitration Rules and Procedures. Arbitration will be conducted by a single arbitrator in Wilmington, Delaware, or by videoconference at the option of the party initiating arbitration. The arbitrator's award will be final and binding, and judgment on the award may be entered in any court of competent jurisdiction. The Federal Arbitration Act governs the interpretation and enforcement of this Section. ### 13.3 Class action waiver To the maximum extent permitted by applicable law, you and Acaysia each agree that each party may bring claims against the other only in an individual capacity, and not as a plaintiff or class member in any purported class, consolidated, or representative proceeding. Unless both parties agree otherwise, the arbitrator may not consolidate claims brought by or against more than one person. ### 13.4 Opt-out You may opt out of the arbitration and class action waiver in Sections 13.2 and 13.3 by sending written notice to [founders@acaysia.com](mailto:founders@acaysia.com) within thirty (30) days of first agreeing to these Terms. Your notice must include your name and a clear statement that you wish to opt out of arbitration. If you opt out, disputes will be resolved in the courts identified in Section 13.5. ### 13.5 Exceptions and judicial forum Notwithstanding the foregoing, either party may bring an individual action in small-claims court for claims within that court's jurisdiction, and either party may seek injunctive or equitable relief in a court of competent jurisdiction to protect its intellectual property rights or to enforce confidentiality obligations. For claims that are not subject to arbitration or for which arbitration is unavailable, the parties consent to the exclusive jurisdiction and venue of the state and federal courts located in Wilmington, Delaware. ### 13.6 EU, UK, Canadian, and consumer users The arbitration and class action waiver provisions of this Section 13 do not apply to the extent prohibited by mandatory law of your country of residence in the EEA, the UK, Switzerland, or Canada (including provincial consumer-protection law, such as Quebec's Consumer Protection Act), or by other mandatory consumer-protection law. In such cases, disputes will be resolved in the courts of competent jurisdiction in accordance with applicable mandatory law. ## 14. Changes to These Terms We may modify these Terms at any time by posting a revised version on the Site. The revised Terms will become effective on the date posted or such later date as we specify. Your continued use of the Site after the effective date of any revised Terms constitutes your acceptance of those Terms. If you do not agree to the revised Terms, you must stop using the Site. ## 15. Miscellaneous 1. **Entire agreement.** These Terms, together with the Privacy Policy and any other documents incorporated by reference, constitute the entire agreement between you and Acaysia regarding the Site and supersede all prior agreements and understandings on this subject. 2. **Severability.** If any provision of these Terms is held invalid or unenforceable, that provision will be enforced to the maximum extent permitted, and the remaining provisions will remain in full force and effect. 3. **No waiver.** Acaysia's failure to enforce any right or provision of these Terms will not be deemed a waiver of that right or provision. 4. **Assignment.** You may not assign or transfer these Terms or any rights or obligations hereunder without our prior written consent. We may assign these Terms in connection with a merger, acquisition, financing, reorganization, sale of assets, or by operation of law without notice to you. 5. **Force majeure.** Neither party will be liable for any failure or delay in performance under these Terms caused by events beyond its reasonable control, including acts of God, war, terrorism, civil disturbance, labor disputes, governmental action, or infrastructure failure. 6. **Notices.** We may provide notices to you by posting on the Site or by sending an email to the address you provided. Notices to Acaysia must be sent to [founders@acaysia.com](mailto:founders@acaysia.com). 7. **Headings.** Section headings are for convenience only and have no legal effect. 8. **Relationship of the parties.** These Terms do not create any agency, partnership, joint venture, employment, or franchise relationship between you and Acaysia. 9. **Government users.** If you are a U.S. federal government end user, the Site Content is a "commercial item" as that term is defined at 48 C.F.R. 2.101, and any technical data is "commercial computer software documentation" as that term is defined at 48 C.F.R. 12.212. Use is subject to the rights and restrictions set forth in these Terms. ## 16. Contact Acaysia, Inc.\ Attention: Legal\ Oakland, California, United States\ [founders@acaysia.com](mailto:founders@acaysia.com) --- # Privacy Notice Canonical: https://acaysia.com/privacy Effective August 21, 2026 This Privacy Notice explains how Acaysia, Inc. ("Acaysia," "we," or "us") collects, uses, and shares personal information through acaysia.com (the "Site"). It applies to personal information of individuals located anywhere in the world, including the United States, Canada, the European Economic Area, the United Kingdom, and Switzerland. By using the Site you acknowledge this Notice, and where the law requires consent for the collection, use, or disclosure described here (as Canada's PIPEDA does), you provide it through the choices described in this Notice. If you have questions about this Notice or want to exercise the rights described below, email us at [founders@acaysia.com](mailto:founders@acaysia.com). ## What this Notice covers This Notice covers personal information we collect from visitors to acaysia.com, primarily through the contact form, through our own first-party page-view analytics, and through standard web logs maintained by our hosting provider. It does not cover personal information we receive from customers in the course of a commercial engagement, which is handled under separate contractual terms. ## What we collect **Information you give us.** When you submit the contact form or otherwise correspond with us, you give us your name, your email address, the company you work for, and the contents of your message. Contact form submissions are stored together with the IP address and browser identification (user agent) they were sent from, for security and abuse prevention. **Information we collect automatically.** We run our own first-party page-view analytics. When you visit the Site, it records the page you viewed, the page that referred you, your browser language, a bucketed viewport-size label (never raw screen dimensions), and any campaign (UTM) tags in the link you followed. It sets no cookie and stores no identifier in your browser. On our servers, your IP address is used to derive an approximate country and a visitor hash whose salt rotates daily — so visits can't be linked across days — and the raw IP address is not kept in our analytics; it's retained only transiently for rate limiting and abuse prevention. If your browser sends a Do Not Track signal, the analytics beacon doesn't run at all. Separately, our hosting provider maintains standard web logs of requests to the Site. **Cookies.** The Site sets no cookies. A few functional flags are kept in your browser's own session storage (for example, that you've unlocked gated resources during a visit); they never leave your browser and are cleared when the tab closes. We don't use cookies for advertising, profiling, cross-site tracking, or behavioral targeting. We don't deploy third-party analytics, ad trackers, social media pixels, fingerprinting, or session-replay tools. Because we don't engage in any of that, you won't see a cookie banner on this Site. **What we don't collect.** We don't knowingly collect sensitive personal information (the categories defined by Article 9 of the GDPR or by California's CPRA, for example government identifiers, precise location, health data, religious beliefs, sexual orientation, or biometrics). Please don't send us any of that through the contact form. ## How we use information We use the information we collect for a small number of straightforward purposes: - To reply to your inquiry and provide information you've asked for. - To operate, maintain, and secure the Site, and to detect and prevent fraud or abuse. - To follow up with you about your inquiry and discuss related business matters where appropriate. - To meet our legal obligations, respond to lawful requests, enforce our terms, and defend our legal rights. - To improve the Site and our products. We do not sell your personal information. We do not share your personal information for cross-context behavioral advertising as those terms are defined under California law. We do not make automated decisions about you that produce legal or similarly significant effects. ## Legal bases (EEA, UK, and Switzerland) If you're in the EEA, the UK, or Switzerland, we process your personal information on one of these legal bases under Article 6 of the GDPR (and the equivalent provisions of the UK and Swiss laws): - **Consent.** When you submit the contact form, you consent to our processing of what you've submitted so we can respond. You can withdraw consent at any time. - **Legitimate interests.** For security, fraud prevention, internal record-keeping, and following up on your inquiry, where those interests are not overridden by your rights. - **Legal obligation.** Where the law requires us to retain, disclose, or otherwise process information. - **Pre-contractual steps.** Where you're inquiring about a potential commercial relationship, processing what you've submitted may be necessary to take steps at your request. ## How we share information We share information sparingly. **Service providers.** We use a small number of vendors to operate the Site and respond to inquiries. They process information on our behalf under contractual confidentiality and data protection obligations. They are: - **Amazon Web Services (AWS)** for hosting, content delivery, storage, and email delivery (Amazon S3, CloudFront, Lambda, DynamoDB, and SES). AWS is bound by a data processing addendum that incorporates the European Commission's Standard Contractual Clauses for international transfers. - **Google** (Google Workspace), which hosts our company mailbox, and therefore processes email you exchange with us. The Site's typefaces are self-hosted; no font provider receives your requests. If we add new providers, we'll update this Notice. **Legal disclosures.** We may disclose information when we believe in good faith that disclosure is required or permitted by law, for example in response to a subpoena, court order, or other legal process, to enforce our terms, to protect Acaysia or others from harm, or to investigate fraud or security incidents. **Business transactions.** If we go through a merger, acquisition, financing, reorganization, bankruptcy, or sale of all or part of our assets, personal information may be transferred or disclosed as part of that transaction with appropriate confidentiality protections and in line with applicable law. **With your consent.** We share information with other parties when you direct us to or when you consent to it. ## International data transfers We're based in the United States, and we store and process information here and in the other places our service providers operate. Other countries' data protection laws may differ from yours and may not provide the same level of protection. If we transfer personal information from the EEA, the UK, or Switzerland to a country without an adequacy decision, we rely on appropriate safeguards, typically the European Commission's Standard Contractual Clauses, the UK International Data Transfer Addendum, or another mechanism permitted under Chapter V of the GDPR. You can ask us for a copy of the safeguards we rely on. If you're in Canada, your personal information is transferred outside your province and outside Canada: it is stored and processed on our systems in the United States, where it is subject to United States law and may be accessible to US courts, law enforcement, and national security authorities. We apply the protections described in this Notice wherever the information is processed, and we require comparable protection from the service providers that process it for us. ## How long we keep information We keep personal information for as long as we reasonably need it for the purpose it was collected, to follow up on your inquiry, to meet legal and record-keeping obligations, and to resolve disputes. Contact form submissions are retained for the duration of any active discussion and for a reasonable period after, consistent with our internal record-keeping. You can ask us to delete information earlier (see the next section). ## Security We use reasonable administrative, technical, and physical safeguards to protect personal information. No system is perfectly secure, and we don't claim otherwise. If a breach of our security safeguards involving your personal information creates a real risk of significant harm to you, we will notify you and the appropriate regulators — including the Office of the Privacy Commissioner of Canada where required — as applicable law requires, and we keep records of any such breach. ## Your rights Depending on where you live, you have rights with respect to your personal information. To exercise any of them, email [founders@acaysia.com](mailto:founders@acaysia.com). We may need to verify who you are before we respond. We won't retaliate against you for exercising your rights, and you can use an authorized agent. ### If you're in the EEA, the UK, or Switzerland You have the right to: - Access the personal information we hold about you. - Have inaccurate information corrected. - Have information deleted (the "right to be forgotten"). - Restrict or object to certain processing. - Receive a portable copy of information you gave us, in a structured, commonly used, machine-readable format. - Withdraw consent at any time, without affecting the lawfulness of processing before you withdrew. - Lodge a complaint with your local data protection authority. For the EEA, see [edpb.europa.eu](https://edpb.europa.eu/about-edpb/about-edpb/members_en). For the UK, see [ico.org.uk](https://ico.org.uk). ### If you're a California resident You have rights under the CCPA, as amended by the CPRA: - Know what categories and specific pieces of personal information we collect, use, disclose, or share about you, the sources, the purposes, and the categories of recipients. - Delete personal information. - Correct inaccurate personal information. - Opt out of the sale or sharing of personal information for cross-context behavioral advertising. We don't do either. - Limit the use and disclosure of sensitive personal information. We don't collect sensitive information for purposes that would trigger this right. - Not be retaliated against for exercising these rights. In the past twelve months we have collected these CCPA categories: identifiers (name, email), commercial information (in limited form, where you mention it in a contact-form message), internet activity (IP address, browser metadata), and approximate geolocation (derived from IP address). We have not sold or shared personal information for cross-context behavioral advertising, and we don't intend to. We don't have actual knowledge that we sell or share personal information of consumers under sixteen (16). ### If you live in another US state with a comprehensive privacy law This includes Virginia, Colorado, Connecticut, Utah, Iowa, Texas, Oregon, Montana, Tennessee, Delaware, New Hampshire, New Jersey, and others. You generally have similar rights to know, correct, delete, port, and opt out of certain processing. We honor verifiable requests under your state's law. ### If you're in Canada We handle personal information of individuals in Canada in accordance with the Personal Information Protection and Electronic Documents Act (PIPEDA) and substantially similar provincial laws, including Quebec's Act respecting the protection of personal information in the private sector, as amended by Law 25, and the Personal Information Protection Acts of Alberta and British Columbia. We collect, use, and disclose your personal information with your consent or as those laws otherwise permit, and only for the purposes described in this Notice. You have the right to: - Access the personal information we hold about you and be told how it has been used and to whom it has been disclosed. - Have inaccurate or incomplete information corrected. - Withdraw your consent at any time, subject to legal or contractual restrictions and reasonable notice. - Challenge our compliance with these laws by contacting our Privacy Officer (see "Contact us" below). - Complain to the Office of the Privacy Commissioner of Canada ([priv.gc.ca](https://www.priv.gc.ca)) or, where a provincial law applies, your provincial regulator — in Quebec, the Commission d'accès à l'information ([cai.gouv.qc.ca](https://www.cai.gouv.qc.ca)). ### Appeals If we decline your request and your state's law gives you a right to appeal (Virginia, Colorado, Connecticut, and certain other states do), email us at [founders@acaysia.com](mailto:founders@acaysia.com) with the subject line "Privacy Appeal." We'll respond within the timeframe your state requires. ## California Shine the Light California Civil Code Section 1798.83 lets California residents request information about disclosures of personal information for third parties' direct marketing. We don't disclose personal information to third parties for their direct marketing. ## Do Not Track Browsers can transmit "Do Not Track" signals. We honor them: if your browser sends one, our first-party analytics beacon doesn't run at all. We don't engage in cross-site tracking of users in any event. ## Children The Site isn't directed to children under sixteen (16) and we don't knowingly collect information from them. If you think we have, contact us and we'll delete it. ## Third-party links The Site links to other sites. This Notice doesn't cover them, those sites have their own privacy practices. ## Updates We update this Notice from time to time. When we do, we'll change the "Effective" date at the top. If a change is material, we'll provide additional notice, for example by posting a prominent notice on the Site. Your continued use of the Site after the effective date of a revised Notice indicates you accept it. ## Contact us Acaysia, Inc.\ Attention: Privacy Officer\ Oakland, California, United States\ [founders@acaysia.com](mailto:founders@acaysia.com) Our Privacy Officer is the person responsible for the protection of personal information and is accountable for our compliance with the laws described in this Notice, including PIPEDA and Quebec's Law 25. Reach them at the address above. We have not appointed a representative under Article 27 of the GDPR. EEA, UK, and Swiss residents can reach us directly using the contact details above. --- # AcaysiaRT Canonical: https://acaysia.com/engines/acaysia-rt **Simulation Runtime** A high-throughput GPU simulation runtime built for **ensemble prediction** — the control, optimization, and learning workflows that single-trajectory solvers cannot support. One templated CUDA core serves **127 registered process models**, from a pH-CSTR to a 168-state chromatography column, each one bit-exact parity-tested against a PyTorch reference. Proven first on chemical reactors. It is the runtime underneath the Acaysia MPPI control system, the engine [AcaysiaDRT](https://acaysia.com/engines/acaysia-drt.md) composes for spatial detail, and — through its [Rete](https://acaysia.com/engines/acaysia-rt.md) layer — the engine that runs whole plants with recycle. *Schematic of an MPPI ensemble rollout: thousands of sampled control trajectories fanning out from one process state, cost-weighted toward a setpoint band. Schematic — MPPI samples N rollouts over horizon H, then softmax-weights them by cost.* ## Headline Stats - **127** — Registered process models - **6.1B** — Reactor-steps/sec on one RTX 5080 - **182×** — Faster than the PyTorch reference ## Measured, With Conditions Attached Every number below is a reproducible benchmark from the repository, not a projection. A throughput figure without its batch size, horizon, and GPU is not a claim. | Backend | 10,000 rollouts × H=100 | Throughput | | --- | --- | --- | | PyTorch (CPU) — reference / debug only | 29.85 ms | ~33 M reactor-steps/sec | | **CUDA (RTX 5080, Blackwell sm_120)** | **0.164 ms** | **~6.1 B reactor-steps/sec** | ### Multi-GPU: Data-parallel batch sharding Reactors are independent, so a batch splits across GPUs and results stay **bit-exact** against the single-GPU path. Measured on 2× RTX 5080 at N=200,000 × H=100 — a 646 MB trajectory. | Mode | Time | Throughput | | --- | --- | --- | | 1 GPU (all N) | 5.6 ms | 3.6 B steps/sec | | 2 GPU, shards resident | 2.0 ms | 10 B steps/sec (~2.8×) | | 2 GPU, gathered to one device | 23 ms | 0.9 B steps/sec | The trajectory tensor is large, so gathering it back to one device dominates the kernel. Multi-GPU pays off when the batch exceeds one card's VRAM, or when each GPU keeps its shard resident across many ticks — per-GPU MPPI or training — rather than gathering every call. ## One Core, 127 Process Models The engine and MPPI kernels are `template` and instantiated per model, so the same high-performance core serves every reactor archetype. A new model is a registry drop-in architecturally — it inherits the engine *and* model-generic MPPI without touching dispatch. ### Reactors & separations: The classical core pH-CSTR, reactive CSTR, PFR, batch and semi-batch. Binary and multi-feed distillation, dividing-wall columns, absorbers, strippers, extractors, decanters, flash drums, evaporators, crystallizers. ### Minerals & pyrometallurgy: Heavy industry Leach tanks, autoclave pressure-oxidation, thickeners, grinding mills, flotation cells, hydrocyclones, screens, filter presses, EAF and BOF steelmaking on JANAF thermodynamics. ### Energy transition: Electrochemical & capture SOEC electrolyzers, vanadium redox-flow batteries, solid-sorbent direct-air capture on a temperature/vacuum swing, fired heaters, boilers, compressors, expanders, FCC regenerators. ### Bio, water & food: Living and regulated processes Immobilized-enzyme packed beds, penicillin fed-batch fermentation, photobioreactors, ASM2d biological nutrient removal, UV advanced-oxidation, UHT pasteurization. | Representative model | What it is | Controlled output | | --- | --- | --- | | **pH-CSTR** | Neutralization with carbonate and weak acid/base equilibria — the reference model | pH | | Distillation column | Binary constant-molar-overflow, tray-by-tray (Skogestad "Column A") | Distillate purity | | Reaction-network CSTR | Data-defined mass-action reactor, fixed-dim, any network — hosts 45 curated processes | Product concentration | | Crystallizer | Cooling MSMPR via the method of moments | Temperature → crystal size | | Autoclave POx leach | Pressure-oxidation hydrometallurgical leach | Metal extraction | | Dividing-wall column | 82-state thermally-coupled distillation | Side-draw purity | | Penicillin fermenter | Bajpai–Reuss fed-batch antibiotic fermentation | Product titer | | SOEC electrolyzer | Solid-oxide electrolysis cell | Hydrogen rate | A representative selection. Alongside the 127 registered models sit **45 curated reaction-network processes** running on the shared reaction-network engine. **126 of 127** are MPPI-controllable — the exception is `pipe`, a transfer-line connector. The largest state dimension in the registry is 168; 43 models carry the stream contract that makes them wireable into a [Rete](https://acaysia.com/engines/acaysia-rt.md) plant. ## Technical Specifications ### Backend policy: CUDA is production The CUDA engine is the only production backend. PyTorch exists for distribution/ABI, the Python API surface, and reference correctness checking — it is not a second execution engine, and selecting it emits an explicit warning. ### Kernels: One core, instantiated per model Adding a process model does not mean writing a new engine. Each model plugs into the same templated CUDA core and the same controller, so performance work done once benefits every model in the registry — and a new one arrives without a bespoke code path to maintain. ### API: Tensors in, tensors out `step` / `rollout` / `observe` dispatch through the registry; `step_soa` and `rollout_soa` are the structure-of-arrays fast path for MPPI hot loops. TORCH_LIBRARY bindings keep tensors zero-copy on device. ### State design: Conserved totals, projected equilibrium Speciation is never carried in state. The step marches conserved totals and clamps to physical bounds; `observe()` solves equilibrium on demand — a log-domain Newton iteration for pH — and applies sensor bias at the output. ## Model-Generic MPPI One sampling-based CUDA controller works with every registered model. Its cost is data-driven over a model's observable vector — per-output target, weight, and bounds — so a model picks what it tracks by setting weights. There is no per-model controller code to write. ### Sampling: Thousands of futures per tick MPPI needs thousands of parallel rollouts over a 100–200 step horizon with consistent, deterministic stepping. That requirement is the reason this runtime exists, and it is what makes path-integral control practical for process plants. ### Closed loop: Convergence-tested per model Every MPPI-controllable model ships a closed-loop convergence test, not just a smoke test — the controller has to actually drive that model's tracked output to target in CI. ### Biasing: Surrogate corrections, cleanly bounded The biasing layer is the entry-side transform that shapes a learned surrogate's predicted corrections into a model's declared correction slots — 117 of the 127 models have them. It is deliberately standalone: torch and the ontology only, enforced by test. ### Plants: Whole-flowsheet control Single units compose into plants through the Rete layer below, where one MPPI sample becomes one whole-plant rollout — recycle loops included. ## Rete: Whole Plants, Not Just Units A plant is not a pile of unit operations — it is a **loop**. Product streams get fed back into the units that made them, and every recycle changes the duty of everything upstream. Rete is the layer that wires AcaysiaRT models into a full flowsheet and marches the whole thing forward at once, recycle included. *Flowsheet schematic: fresh feed enters a reactor, flows to a decanter, which splits into a product draw (30%) and a recycle stream (70%) returning to the reactor inlet.* ### Recycle changes the answer Run that loop at 70% recycle and the two conversion numbers move in *opposite* directions. Per-pass conversion falls from 0.4623 to 0.3068 — each trip through the reactor does less work. Overall conversion on fresh feed rises from 0.4623 to **0.5961**. That divergence is the textbook recycle result, and it is exactly what a unit-by-unit model cannot tell you. It is also why plant decisions made from single-unit simulations go wrong: the number that matters is the one a loop produces. The balances are cross-checked against **DWSIM**. ### Control: Tune the plant, not the loop Plant-wide MPPI treats one control sample as one whole-plant rollout, so the controller optimizes across units instead of fighting itself — the classic failure when each loop is tuned in isolation and the recycle carries the disagreement around. ### Scale: Dozens of units, thousands of futures Plants of 50-plus mixed unit types march on the GPU, batched over thousands of samples at once. Fusing a plant into a single kernel is worth roughly **22×** over the same work run unit-by-unit on the same card. ### Portable: A plant is a document Topology, feeds, connections, and targets live in a declarative definition you can save, diff, and hand to someone else. Configuration is an artifact, not a state buried in a session. ### Settling: Answers about dynamics It marches in time rather than solving for a steady state, so it reports *when* a recycle loop settles and how it got there — startup, disturbance rejection, and the transients a steady-state flowsheet package never shows you. ## Validation Provenance Validation runs on a T0–T5 ladder where **tier, numerical outcome, and health are orthogonal**. A strong oracle does not lose its tier when it exposes a model gap — which is exactly what keeps the incentives honest. | Tier | What it means | | --- | --- | | **T0** | Running-deployment validation, tied to a real deployment and its committed operating envelope | | **T1** | Full-model oracle — an independent reference covering primary behavior, regimes, and outputs | | **T2** | Partial independent oracle — measurements, standards, or an authoritative external tool on a named subset | | **T3** | Local scientific reference — a good validator-authored analytical model, not externally grounded | | **T4** | Cross-implementation parity — catches numerical and indexing defects; *cannot* establish that the physics is right | | **T5** | Supporting assurance — calibration replay, invariants, regression | ### Census: 92 of 127 models at T0–T2 72% of the roster carries independent evidence at the top three tiers, across **264 logical validators** over 200 cached runs. Current health: 261 fresh, 0 stale, 0 failed. ### Oracles: Best-in-class, per domain A deliberate collage rather than one tool: Cantera for gas and combustion, PHREEQC for aqueous and hydrometallurgy, JANAF for pyrometallurgy, CoolProp for thermophysical properties, and published benchmarks — ASM2d/BSM, Bajpai–Reuss, HCH-1 — for biological units. ### Calibration: Blind, fit, or held-out Every validator declares how it was calibrated. That is the structural defense against the classic failure mode of fitting a model to its own oracle and then citing the agreement as evidence. ### Test suite: 2,650 tests Per-model CUDA↔PyTorch parity, physics invariants, and MPPI closed-loop convergence for all 127 models; plus MPPI cost-layout contracts, biasing slot writes, and multi-GPU sharding. CUDA tests are hardware-gated. Distances are reported as multiples of the declared tolerance: 0.58× means 42% headroom remains, 1.83× means 83% over. A gap never demotes a tier — it is the result the validator was built to find, and several validators have surfaced and documented real fidelity gaps. ## Simple, Powerful API Tensors in, tensors out. Pick a model by name; the registry handles dispatch. **Installation (source):** ```bash # AcaysiaRT ships from source with the CUDA extension pip install -e . python setup.py build_ext --inplace # Requires torch, numpy, and the shared # acaysia-ontology-schemas contracts package ``` **Quick Start (Python):** ```python import torch import acaysia_rt as art # pH-CSTR state [N, 8]: T, V, DIC, ALK, # WAC_TOT, WBASE_TOT, BIAS_PH, UA_SCALE states = torch.tensor([[ 298.0, 100.0, 0.002, 0.002, 0.0, 0.0, 0.0, 1.0 ]], device="cuda") # Roll out a horizon of 100 steps traj = art.rollout(states, controls_seq, dt=0.1) # Derived quantities: pH, speciation, ... pH = art.observe(states)[:, art.OutputIndex.PH] ``` ## Ready to Accelerate Your Control Systems Contact us to discuss how AcaysiaRT can power your MPPI control and simulation workloads. [Request Early Access](https://acaysia.com/contact.md) --- # AcaysiaDRT Canonical: https://acaysia.com/engines/acaysia-drt **Spatial Simulation Engine** 3D spatial simulation engine for physical process unit operations. Extends AcaysiaRT with Lattice Boltzmann transport for full 3D spatial modeling of any unit-op class in the Acaysia typed process ontology. Proven first on chemical reactors. Enables digital twin capabilities for industrial process design, optimization, and predictive maintenance. ## Capabilities ### Scale: nL - 100s L Microfluidics to industrial scale reactor simulation in a single framework. ### LBM Stencil: D3Q19 Full 3D velocity space with BGK collision operator for accurate transport physics. ### Accuracy: 2nd Order Spatial Second-order spatial accuracy with BGK collision for reliable simulation results. ### Coupling: Strang Splitting Operator splitting for accurate transport-reaction coupling in complex systems. ## Technical Specifications ### Transport Physics D3Q19 Lattice Boltzmann with BGK collision. Bounce-back and interpolated boundary conditions. Variable viscosity and diffusivity fields. Thermal LBM for temperature-dependent reactions. ### Industrial Reactor Types Continuous stirred-tank reactors (CSTR). Plug flow reactors (PFR). Batch and semi-batch reactors. Microfluidic and lab-on-chip devices. ### GPU Implementation Structure-of-Arrays memory layout. Multi-GPU domain decomposition. Overlapped communication and compute. Adaptive mesh refinement (AMR) support. ### Visualization & Export ParaView/VTK compatible output. napari plugin for live visualization. USD export for 3D platforms. Time-series animation export. ## Digital Twin Use Cases ### Virtual Commissioning Test reactor configurations and control strategies before physical deployment. Reduce commissioning time and identify design issues early. ### Process Optimization Explore operating conditions across the full parameter space. Identify optimal setpoints for yield, selectivity, and energy efficiency. ### Predictive Maintenance Model fouling, catalyst deactivation, and equipment degradation. Predict maintenance windows and optimize turnaround scheduling. ### Control System Integration Designed to integrate with the Acaysia control system, enabling predictive control strategies built on full spatial awareness of process dynamics. ## Ecosystem Integration ### Cloud Deployment Kubernetes-native orchestration across AWS, Azure, and GCP. Auto-scaling GPU clusters with S3/Blob storage integration and REST API for job submission. ### Visualization Ecosystem Real-time streaming to ParaView and napari visualization clients. OpenUSD scene description export. VR/AR-ready spatial data formats. ## 3D Simulation Made Simple Full spatial reactor simulation in just a few lines. **Installation (pip):** ```bash # Install AcaysiaDRT pip install acaysia-drt # Requires AcaysiaRT as dependency # CUDA support included automatically ``` **Quick Start (Python):** ```python import torch from acaysia_drt import SpatialCSTR3D, DRTParams # Create 32x32x16 reactor grid params = DRTParams( Nx=32, Ny=32, Nz=16, V_cell_L=1.0, dx_m=0.1 ) # Initialize and simulate reactor = SpatialCSTR3D(params, device="cuda") states, stats = reactor.step_soa( states, controls, velocity, dt=0.001 ) ``` ## Build Your Digital Twin Contact us to discuss your digital twin requirements and how AcaysiaDRT fits your process simulation needs. [Contact Us](https://acaysia.com/contact.md) --- # AcaysiaGEM Canonical: https://acaysia.com/engines/acaysia-gem **Geometry Engine** A spatial simulation is only ever as good as the shape you run it in. Real vessels are not boxes — they have dished bottoms, baffles, downcomers, sparger rings and nozzles set at awkward angles, and every one of those changes the answer. GEM is the layer that builds that shape. It authors the vessel's interior as a parametric model, proves the model is fit to simulate *before* a GPU hour is spent on it, and hands it to [AcaysiaDRT](https://acaysia.com/engines/acaysia-drt.md) ready to run. *Photoreal render of a process plant built by AcaysiaGEM: vessels, columns, routed piping and structural steel on a hardstanding.* **Not a picture of a plant — the plant itself.** Twelve units, routed lines and procedural steel, authored parametrically and rendered straight from the same model the solver runs on. 678,374 triangles, rendered in about 6 seconds. ## Model the Space, Not the Steel Ordinary CAD describes the vessel: the shell, the flanges, the weld prep. A solver does not care about any of that. It cares about the shape of the *liquid* — the negative space the fluid actually occupies. GEM models that space directly, and labels its edges as it builds them: this face is where fluid enters, that one is a wall, this volume is swept by an impeller and is not fluid at all. That is why its output drops into a solver without the usual translation losses. *Cross-section of a stirred vessel showing the fluid domain as tagged negative space: an inlet nozzle, an outlet, wall boundaries, and an impeller exclusion zone.* ## Stats - **18** — Parametric geometries, lab duct to full vessel - **100%** — Boundary faces tagged, or the build fails - **0** — Wasted GPU hours on a leaking domain ## How a Domain Gets Built Four properties do most of the work here, and each one exists because the alternative fails quietly rather than loudly. ### Parametric: Dimensions, not drawings A vessel is described by the numbers an engineer already has — diameter, liquid height, impeller ratio, clearance, nozzle positions — and rebuilt from them on demand. Changing a design means changing a number, not redrawing a model. ### Tagged at build: The builder knows what it built Boundary faces get their meaning — inlet, outlet, wall, symmetry, exclusion — at the moment the builder creates them. No fragile post-hoc guessing about which surface was the inlet, and no selector language to get wrong. ### Grounded: Checked against real CAD Where a trusted reference part exists, GEM's model is regression-checked against it — volume, surface area, centroid, inertia, bounding box — so a simplified flow domain stays honest about the equipment it represents. ### Internals: What the solver sees, and what it doesn't Agitator shafts, turbines, drive motors, baffles and nozzle flanges can be carried for visual fidelity without polluting the physics. A rotating impeller is handed to the solver as the volume it sweeps, which is the honest representation of it. ## Eighteen Shapes, Three Families | Family | Geometries | What it is for | | --- | --- | --- | | **Flow domains** | Channel, cylinder in channel, backward-facing step, sudden expansion, serpentine mixer, lid-driven cavity | Lab-scale ducts — the shapes the solver is benchmarked on, where an analytic or published answer exists | | **Inline components** | Tee mixer, orifice plate | Round pipe fittings — blending at a junction, and pressure loss across a restriction | | **Vessels** | Stirred tank, distillation column, bubble column, packed-bed reactor, fermenter, flash drum, decanter, cyclone, drum, horizontal separator | The real equipment — internals and all, at the scale a plant actually runs | **Stirred tank.** Shaft, turbine, drive motor, baffles and nozzle flanges are all here — carried for fidelity, and none of them confusing the solver about where the fluid is. **Distillation column, walls made transparent.** The tray stack and downcomers are genuinely modelled, not implied by a texture — which is why a spatial solver can be asked what happens between them. Each geometry sweeps across its whole parameter range in the test suite, so a mis-tagged large impeller, a baffle poking through a dished bottom, or an unsealed vessel is caught automatically on any change — not discovered halfway through a run. ## Fail on the Desk, Not on the Cluster The expensive failure in spatial simulation is not a wrong answer — it is a run that consumes a night of GPU time and produces nonsense because the domain leaked. Every GEM domain has to pass a health check before it can be exported. ### Watertight: A real solid Valid, watertight, a single connected body with positive volume. The check that catches a geometry which merely *looks* closed on screen. ### No untagged faces: Every boundary has a rule Each boundary face must carry exactly one tag. An untagged face is a hole in the physics — a boundary the solver has no instruction for — and it leaks silently rather than erroring. ### Nothing stranded: One connected fluid The fluid must be one region with no isolated pockets, every declared opening must actually reach it, and the domain must be sealed everywhere an opening does not exist. ### Nothing protruding: Internals stay inside Anything declared as submerged has to genuinely sit within the fluid. When it doesn't, the check names the offending point rather than leaving you to find it in a render. ## One Vessel Is the Primitive. A Plant Is the Point. The same machinery that places one vessel places thirty of them. GEM grows from independent geometries into a laid-out, connected, named plant — the physical counterpart of a [Rete](https://acaysia.com/engines/acaysia-rt.md) flowsheet. **Thirty units, four parallel trains, 1.6 million triangles** — spanning 105 × 172 metres of plot. Built from eight distinct designs, because identical units share one underlying shape. ### Plot plan: Place and orient Units are positioned as a plot plan, with each placed unit carrying its boundary tags along with it — the detail that separates a real assembly from a pile of shapes that lost their meaning when they moved. ### Piping: Routed, not drawn Connections between units are routed through the plant as real line work rather than sketched, with nozzle-to-nozzle port matching and collision awareness. ### Structure: Racks, steel and access Pipe racks, supports and access structure come with it, so what you are looking at reads as a plant rather than as vessels floating in space. ### Cheap by design: Thirty units, eight builds Identical units share one underlying shape, so a thirty-unit plan built from eight distinct designs costs eight builds and not thirty. That is what makes assembling a whole plant routine instead of an overnight job. ## The Model Is the Picture The geometry that runs the simulation is the same geometry that renders it. There is no separate marketing model to fall out of sync — when the design changes, the picture changes with it. **Metal.** The plant as equipment — what it looks like standing on the hardstanding. **X-ray.** Same scene, one switch. The walls go transmissive and the internals the geometries already built — trays, downcomers, impellers, packing — are simply there. **Or skip the picture entirely.** This is the exported model itself, running in the browser — the same geometry the solver meshes, at 147,710 triangles. Nothing was rebuilt for the web; it is one export flag on the plant above. ### Two looks: Equipment, or x-ray The pair above is one scene rendered twice. Nothing was modelled specially for the transparent version — the internals were always in the geometry, because the solver needs them there. ### Fast: Seconds, not hours The twelve-unit plant above renders in about 6 seconds on a GPU; the thirty-unit one, at 1.6 million triangles, takes under 8. Fast enough that a render is something you do while iterating, not a deliverable you schedule a day for. ### Web-native: Ship the model, not a screenshot Output is a still image, an editable scene, or a 3D file a browser can open — usually the better answer, since it is the model itself rather than a picture of it. ### Composable: Drops onto a page Renders come out on transparent backgrounds with the plant still standing on its own shadow, so it composites cleanly onto whatever sits behind it. ## Straight Into the Solver ### Export: Lattice-ready, three ways A validated domain converts into exactly what the solver consumes. Where a geometry carries an exact mathematical description of its surface, the export uses it, giving sub-cell accuracy at curved walls instead of a staircase. ### Contract: No translation losses The export encodes the solver's own conventions directly, pulled from its source rather than reimplemented — the class of mismatch that produces a simulation which runs happily and means nothing. ### Downstream: Then it's a DRT run From there it is an [AcaysiaDRT](https://acaysia.com/engines/acaysia-drt.md) job: solved flow, transport, and the chemistry that [AcaysiaRT](https://acaysia.com/engines/acaysia-rt.md) supplies per cell. ### In the app: Without the toolchain [AcaysiaCORE](https://acaysia.com/core.md) brings this into the browser — pick a geometry, set the parameters, see it in 3D, check the regions, and export it for a run without installing anything. ## Build, Check, Export Three steps, and the third one only runs if the second one passed. **Build a Domain (Python):** ```python from gem.geometries import Channel, ChannelParams from gem.core.regions import collect_regions geo = Channel() params = ChannelParams( length=0.10, height=0.02, width=0.02 ) part = geo.build(params) # tagged solid meta = geo.metadata(params) # serializable for region in collect_regions(part): print(region.name, region.kind.value) ``` **Validate, Then Export (Python):** ```python from gem.validate import validate from gem.export import get_exporter result = get_exporter("voxel_flags").export( part, meta ) # Health + leak checks, per-check verdict report = validate(part, result) print(report.summary()) assert report.ok # no ERROR-severity check failed # Solid mask in the solver's own cell order solid = result.to_dart_linear("solid") ``` ## Start From Your Geometry Contact us about modelling your vessels — and what changes when the simulation runs in the real shape instead of an idealized one. [Contact Us](https://acaysia.com/contact.md) --- # What Is Advanced Process Control (APC)? Canonical: https://acaysia.com/resources/advanced-process-control Guide Advanced process control (APC) is a category of model-based control methods that use a predictive model of a physical process to look ahead, coordinate many variables at once, and keep the process at its most profitable operating point while respecting safety and quality constraints. ## Overview Where a basic PID loop simply reacts to the current error, APC anticipates how a reactor, column, or crystallizer will behave over the next several seconds or minutes and acts before a disturbance forces the operator to intervene. The result is higher yield, lower energy consumption, and steadier throughput from the same equipment. APC is not a single algorithm. It is an umbrella term that has evolved through three broad generations: classical PID loops, model predictive control (MPC), and a newer class of AI-driven, sampling-based controllers such as **MPPI (Model Predictive Path Integral)**. Understanding how these differ is the fastest way to decide what your process actually needs. ## The evolution: PID → classical MPC → AI/MPPI control Process control has progressed from reacting to a single error signal, to optimizing a model over a horizon, to sampling thousands of possible futures in parallel. Each step spends more computation to handle harder problems: stronger nonlinearity, more variables, tighter constraints. | Capability | PID | Classical MPC | AI (MPPI) | | --- | --- | --- | --- | | Handles nonlinearity | Poorly, only near its tuning point | Partially, usually through a linearized model | Natively, by sampling the full nonlinear model | | Constraint handling | None (limits enforced externally) | Explicit, via constrained optimization | Explicit, via penalized trajectory costs | | Retuning effort | High: manual, per loop, and it drifts over time | Moderate: re-identify the model, re-tune weights | Low: adjust the cost function and the model adapts | | Compute approach | Closed-form, near-zero cost | Single optimization solve per step (CPU) | Massively parallel trajectory sampling (GPU) | | Safety fallback | Is the fallback | Typically falls back to PID | Supervised fallback to PID in under 100 milliseconds | The pattern is consistent. PID is robust and cheap but blind beyond its tuning point. Classical MPC adds foresight and constraint awareness but leans on a well-behaved model. MPPI keeps the foresight and the constraints, works with the full nonlinear physics, and pays for it with parallel computation instead of mathematical shortcuts. ## Where APC applies APC delivers the most value on processes that are nonlinear, interacting, constrained, or economically sensitive. Those happen to be exactly the unit operations that are hardest to run well with hand-tuned loops. - **CSTRs and other reactors.** Continuous stirred-tank reactors couple temperature, concentration, and reaction kinetics in ways that make them strongly nonlinear and prone to runaway. This is where Acaysia is proven on real hardware. - **Distillation columns.** Columns are highly interacting multivariable systems where reflux, reboiler duty, and product purity trade off against energy. Coordinated APC captures value that single loops leave on the table. - **Crystallizers.** Crystal size distribution depends on tightly coupled supersaturation and cooling trajectories, making crystallization a natural fit for predictive, constraint-aware control. - **Batch processes.** Batches follow recipe trajectories with no steady state, so the controller has to track a moving target and adapt from batch to batch. Predictive control clearly outperforms reactive tuning here. ## Benefits: yield, energy, throughput Because APC continuously drives a process toward its true optimum instead of a conservative, hand-tuned setpoint, the benefits show up directly on the plant's bottom line. - **Yield.** Holding a reactor closer to its ideal conditions converts more feedstock into product and less into waste or off-spec material. On real CSTR hardware, the MPPI controller measured a yield improvement of more than 2% over a tuned PID baseline. - **Energy.** Predictive control avoids the over-heating, over-cooling, and excess reflux that conservative loops use as safety margin. The same tests measured an energy reduction of around 10%. - **Throughput.** Steadier, constraint-aware operation lets a plant run nearer to its real limits with fewer trips and less variability, increasing sustained output from existing equipment rather than requiring new capital. Quantified figures above are measured on real continuous stirred-tank reactor hardware, not simulation. ## How Acaysia's MPPI approach differs Acaysia is an APC platform built around [MPPI control](https://acaysia.com/resources/mppi-vs-pid-mpc.md) rather than classical MPC. The practical differences come down to how it computes, how it stays safe, and how it fits into a plant you already have. - **GPU-parallel trajectory sampling.** Instead of solving one optimization per step, the controller samples thousands of candidate control trajectories in parallel on a GPU and weights them by cost. This handles strongly nonlinear dynamics without linearizing the model. The sampling runs on Acaysia's simulation engines: [AcaysiaRT](https://acaysia.com/engines/acaysia-rt.md), a high-throughput GPU simulation runtime; [AcaysiaDRT](https://acaysia.com/engines/acaysia-drt.md), a 3D spatial simulation engine; and [Rete](https://acaysia.com/engines/acaysia-rt.md), which couples units into whole plants with recycle. - **Trust Arbiter on every move.** A supervisory component called the Trust Arbiter checks every control action the optimizer proposes before it reaches the process, rejecting moves that fall outside validated bounds. - **PID fallback in under 100 milliseconds.** If the Trust Arbiter loses confidence, control reverts to proven PID in under 100 milliseconds. PID is the safety net, not a competitor. - **ASIL-D inspired, SIL compatible.** The safety architecture is ASIL-D inspired and SIL compatible, and it never interferes with the plant's independent Safety Instrumented System (SIS). - **Brownfield integration.** Acaysia is a brownfield drop-in that speaks OPC UA and EtherNet/IP, so it works with existing control systems from Rockwell, Siemens, Beckhoff, Schneider, and ABB. No rip-and-replace. ## How APC is deployed without disrupting operations A common objection to advanced process control is risk: handing a profitable, safety-critical process to an algorithm sounds like a large step. In practice, modern APC is introduced in stages so the plant earns trust before the controller ever moves a valve. Acaysia follows a three-phase rollout that mirrors how operators themselves build confidence. - **Shadow mode.** The controller runs alongside the existing loops, reading live data and computing what it *would* do, but taking no action. This validates the process model against real plant behavior and surfaces any modeling gaps with zero operational risk. - **Advisory mode.** The controller presents its recommended setpoint moves to operators, who approve or reject them. The plant captures the benefit of the recommendations while keeping a human in the loop and building a track record. - **Closed-loop control.** Once the model and the Trust Arbiter have demonstrated reliability, the controller is allowed to act directly. It is still supervised on every move, can still fall back to PID in under 100 milliseconds, and still leaves the Safety Instrumented System untouched. This staged path is what makes APC adoptable on brownfield plants: the technology proves itself against the incumbent controls before it is trusted with them, and the operator retains a clear, fast path back to known-good PID behavior at every stage. ## Frequently asked questions ### What is the difference between APC and PID control? PID is a single-loop feedback controller that reacts to the current error between a measured value and its setpoint. Advanced process control (APC) is a broader category of model-based methods, such as MPC and MPPI, that use a model of the process to look ahead over a prediction horizon, coordinate many variables at once, and respect operating constraints. PID corrects after a disturbance arrives; APC anticipates it. ### Does advanced process control replace my existing PID loops or safety system? No. APC typically sits on top of existing regulatory PID loops and adjusts their setpoints rather than replacing them. In Acaysia's case, PID also remains the failsafe: the Trust Arbiter can hand control back to PID in under 100 milliseconds. APC never interferes with the Safety Instrumented System (SIS), which remains independent. ### What kind of results does advanced process control deliver? Because APC pushes a process closer to its true optimum while honoring constraints, it commonly improves yield, reduces energy per unit of product, and increases throughput. On real CSTR hardware, not simulation, the MPPI controller measured a yield improvement of more than 2% and an energy reduction of around 10% versus tuned PID baselines. ### How is AI (MPPI) control different from classical MPC? Classical MPC solves a constrained optimization problem at each step, which assumes a relatively well-behaved (often linearized) model. MPPI (Model Predictive Path Integral) instead samples thousands of candidate control trajectories in parallel on a GPU and weights them by cost, so it handles strongly nonlinear dynamics and non-convex objectives without linearizing. It trades a single deterministic solve for massively parallel sampling. ## Related - [MPPI vs PID vs MPC: how path-integral control differs](https://acaysia.com/resources/mppi-vs-pid-mpc.md) - [The Acaysia control system](https://acaysia.com/product.md) - [AcaysiaRT: high-throughput GPU simulation runtime](https://acaysia.com/engines/acaysia-rt.md) - [Rete: whole-plant flowsheets with recycle](https://acaysia.com/engines/acaysia-rt.md) ## See APC on your process If you run reactors, columns, crystallizers, or batch processes and want to know what advanced process control could recover in yield, energy, and throughput, talk to the team. [Book a discovery call](https://acaysia.com/contact.md) --- # MPPI vs PID vs MPC: How Path-Integral Control Differs Canonical: https://acaysia.com/resources/mppi-vs-pid-mpc Comparison PID, classical MPC, and MPPI all answer the same question: what should the controller do next? They just answer it with very different levels of capability. PID reacts to error one loop at a time. Classical MPC plans ahead using a linearized model. MPPI plans ahead by sampling thousands of nonlinear control trajectories in parallel, which makes it the natural fit for highly nonlinear, tightly constrained processes where the other two run out of headroom. ## What is MPPI? MPPI stands for **Model Predictive Path Integral** control. It is a sampling-based, GPU-parallel form of model predictive control. Instead of solving a single optimization equation each cycle, MPPI draws a large population of candidate control trajectories, rolls each one forward through a process model, scores every trajectory with a cost function, and then blends them into one control move using an exponential, cost-weighted average. Lower-cost trajectories carry more weight; high-cost or constraint-violating trajectories are effectively pushed aside. Because MPPI only needs to *evaluate* a model rather than invert or linearize it, it handles nonlinear dynamics, hard constraints, and non-convex objectives directly. The work of sampling and rolling out trajectories is naturally parallel, so a GPU can evaluate thousands of candidates per control cycle in real time. That combination of nonlinear handling, constraint awareness, and massive parallelism is what sets path-integral control apart from the controllers most plants run today. ## MPPI vs PID vs classical MPC at a glance The table below summarizes the practical differences. No single method is best everywhere. PID is simple and battle-tested, classical MPC adds prediction, and MPPI adds nonlinear, sampled optimization on top. | Capability | PID | Classical MPC | MPPI | | --- | --- | --- | --- | | Handles nonlinearity | Limited, only near one operating point | Partial, often via local linearization | Native, evaluates the nonlinear model directly | | Constraint handling | Indirect (clamping, anti-windup) | Explicit in the optimization | Explicit via cost penalties on every trajectory | | Requires linear / quadratic model | No (single-loop heuristic) | Typically yes for tractable solves | No, any forward model works | | GPU-parallel sampling | No | No (sequential numerical solve) | Yes, thousands of rollouts in parallel | | Retuning effort on process change | High: manual gain retuning | Moderate: rebuild the model and solver setup | Lower: adjust cost terms and reuse the model | | Safety fallback | Is itself the common fallback | Usually drops back to PID | Supervised; PID fallback in <100ms | | Brownfield fit | Ubiquitous; already installed | Common in refining / large continuous units | Drop-in supervisory layer over existing loops | ## PID: simple, robust, and everywhere Proportional-Integral-Derivative control is the workhorse of industrial automation, and for good reason. It is easy to understand, cheap to implement, and runs on essentially every PLC and DCS in service. A PID loop reacts to the current error between a setpoint and a measurement, with three terms that respond to the size of the error, its accumulated history, and its rate of change. The limitation is that PID is fundamentally a single-loop, reactive controller tuned around one operating point. It does not look ahead, does not reason about constraints, and does not coordinate interacting variables. When a process is strongly nonlinear, like an exothermic reactor whose gain changes with conversion, a single set of PID gains rarely performs well across the full operating envelope, and retuning becomes a recurring manual cost. PID remains an excellent regulatory and fallback layer; it is simply not designed to optimize a process. ## Classical MPC: planning ahead with a model Model Predictive Control was a major step forward: instead of reacting to the current error, it uses a process model to predict future behavior over a horizon and solves an optimization problem each cycle to choose the best sequence of moves, then applies the first move and repeats. MPC handles multivariable interactions and explicit constraints far better than PID, which is why it is well established in refining and large continuous units. The trade-off is in the math. To keep the per-cycle optimization fast and reliable, classical MPC typically relies on a linear model with a quadratic cost, which makes the problem convex and quick to solve. That works well near the operating point it was built around, but real nonlinearity, hard non-convex constraints, and large operating-range changes strain the linear assumption. Building and maintaining accurate linearized models, and re-solving when the process drifts, is real engineering effort. MPC is powerful and proven. It just inherits the cost of its linear-quadratic structure. ## MPPI: nonlinear optimization by sampling MPPI keeps the predictive, horizon-based idea of MPC but changes how the optimization is done. Rather than solving an equation, it samples a large set of candidate control sequences, simulates each through a forward model, and weights them by cost. Because it only evaluates the model, the model can be fully nonlinear, whether first-principles, learned, or a hybrid, and the cost function can encode non-convex objectives and safety penalties without breaking the solver. MPPI has limits of its own. Its quality depends on sampling enough trajectories, and on a forward model good enough to rank them sensibly. That sampling workload is also why GPU parallelism is so important: it is what makes thousands of real-time rollouts feasible. Like any optimizer, MPPI must be supervised so that a single bad sample can never reach the plant. Inside the right safety architecture, though, that ability to optimize nonlinear, constrained processes directly is exactly what suits MPPI to the problems where PID and linear MPC plateau. For the bigger picture, see our overview of [advanced process control](https://acaysia.com/resources/advanced-process-control.md). ## How Acaysia uses MPPI Acaysia uses MPPI as the optimization core of a control system for physical processes, focused on improving yield, energy, and throughput. The optimizer never acts alone. Every proposed control move is screened by a supervisor we call the **Trust Arbiter**, which validates each move before it can reach the plant and rejects anything outside safe bounds. If the optimizer is unavailable or a move is rejected, a deterministic **PID fallback engages in under 100 milliseconds**, so the process always has a safe, known controller in command. The safety design is **ASIL-D inspired** and **SIL compatible**, and the system is built to sit alongside an existing safety instrumented system, never to interfere with it. The approach is **proven on real CSTR hardware**, not simulation. On the integration side, Acaysia speaks the standard industrial protocols, **OPC UA and EtherNet/IP**, so it drops into brownfield plants alongside existing PLC and DCS infrastructure. The heavy GPU rollout work behind MPPI is carried by our simulation runtime; you can read more about it on the [AcaysiaRT engine page](https://acaysia.com/engines/acaysia-rt.md), and see the full system on the [Acaysia product page](https://acaysia.com/product.md). ## Frequently asked questions ### Is MPPI a replacement for PID? No. MPPI is a supervisory optimization layer that sits above existing regulatory control. In Acaysia deployments, a PID loop remains the safety fallback and engages in under 100 milliseconds if the optimizer is unavailable or a control move is rejected. ### Does MPPI require a perfect model? No. Because MPPI evaluates many sampled control trajectories against a cost function, it tolerates approximate models and uses the cost weighting to favor lower-cost outcomes. Model quality still matters, but MPPI does not require the linear or quadratic structure that classical MPC depends on. ### How is MPPI different from classical MPC? Classical MPC solves a numerical optimization problem each cycle and typically assumes a linear model with a quadratic cost. MPPI instead samples thousands of candidate control trajectories, rolls each one through a forward model, and computes a cost-weighted average. This sampling approach handles nonlinear dynamics and non-convex costs directly and parallelizes well on a GPU. ## Related - [What is advanced process control?](https://acaysia.com/resources/advanced-process-control.md) - [The Acaysia control system](https://acaysia.com/product.md) - [AcaysiaRT: high-throughput GPU simulation runtime](https://acaysia.com/engines/acaysia-rt.md) - [Rete: whole-plant flowsheets with recycle](https://acaysia.com/engines/acaysia-rt.md) ## See MPPI on your process Want to know whether path-integral control can recover yield, energy, and throughput on your reactors or columns? Talk to the team. [Book a discovery call](https://acaysia.com/contact.md)