Building Our Playground Reactor

Simulation is powerful, but it has limits. No matter how sophisticated your digital twin, there are phenomena it doesn't capture. Sensors drift. Valves stick. Heat transfer varies in ways that don't match correlations. If you want control systems that work in the real world, you need to test them in the real world.

That's why we built what we affectionately call the "playground"—a small-scale continuous stirred-tank reactor (CSTR) that lets us test our algorithms on real chemistry before deploying at customer sites.

Why Not Just Use Simulations?

We use simulations extensively—they're essential for rapid iteration and exploring edge cases. But simulations have blind spots:

Perfect Sensors Don't Exist

Simulations assume you know exactly what's happening in the reactor. In reality, sensors have noise, drift, delays, and occasional failures. Control algorithms that perform perfectly with ideal measurements can struggle with real instrumentation.

Models Are Always Wrong

As the statistician George Box famously said, "All models are wrong, but some are useful." Simulations are built on models, so they inherit model errors. Our gray-box approach explicitly accounts for model mismatch, but we need real data to validate that it works.

Edge Cases Surprise You

Real systems find failure modes that simulations miss. A valve that sticks briefly. A sensor that goes offline during a critical moment. A combination of disturbances that wasn't in the test plan. You discover these by running real experiments.

Designing the Playground

Our testbed needed to be small enough to be safe and affordable, but realistic enough to stress-test our algorithms. Here's what we built:

Reactor Specifications

• 5-liter jacketed glass vessel
• Variable-speed magnetic stirrer
• Heating/cooling via circulating bath
• Multiple temperature sensors (redundancy testing)
• pH probe and conductivity sensor
• Precision dosing pumps for reagent feeds

Control Infrastructure

• Industrial PLC with OPC UA (typical of small-scale installations)
• OPC UA server for data communication
• Acaysia edge device running our control software
• Local HMI for manual operation and monitoring

Chemistry Selection

We needed reactions that were safe, inexpensive, and representative of industrial challenges. Our go-to test reactions:

• Saponification: NaOH + ethyl acetate → sodium acetate + ethanol. Simple second-order kinetics, mildly exothermic, well-characterized.
• Neutralization: Acid-base reactions for pH control testing. Nonlinear dynamics around the equivalence point.
• Dissolution: Temperature-dependent solubility for crystallization- related scenarios.

What We've Learned

Running thousands of experiments on our playground has fundamentally improved our algorithms. Some highlights:

Noise Handling Is Critical

Our early algorithms worked beautifully in simulation but oscillated on real sensors. The culprit: high-frequency noise that simulation didn't capture. We redesigned our state estimation to filter aggressively without introducing too much lag. The result works on real data.

Valve Dynamics Matter

Simulations often model valves as instantaneous—set a position, get that position. Real valves have travel time, hysteresis, and stiction. Our control needed to account for these dynamics to avoid overcorrection.

Startup Is the Hard Part

Steady-state control is relatively easy. Startup—going from cold and empty to hot and reacting—is where things get tricky. Our models are trained on operating data, but startup looks different. We developed separate startup sequences and smooth handoff to normal control.

Safety Systems Work

We intentionally triggered our safety limits to verify failsafe behavior. Overheat protection? Works. Communication loss fallback? Works. Manual override? Works. Better to find problems on a 5-liter reactor than a 5,000-gallon one.

A Day in the Lab

A typical experimental campaign might look like this:

1. Morning setup: Prepare reagents, calibrate sensors, verify system health
2. Baseline run: Operate with traditional PID control to establish reference performance
3. Shadow mode test: Run Acaysia in observation mode, compare recommendations to actual control
4. Closed-loop test: Enable Acaysia control, run through various scenarios (setpoint changes, disturbances, grade changes)
5. Stress tests: Intentionally create challenging conditions— sensor failures, feed disruptions, rapid changes
6. Analysis: Review data, identify issues, update algorithms

We've run hundreds of these campaigns, each one improving our software.

From Lab to Plant

The playground is essential, but it's not the end of the road. Our 5-liter reactor can't capture everything about 5,000-liter industrial reactors:

• Heat transfer scales differently
• Mixing dynamics change with size
• Industrial instrumentation has different characteristics
• Plant networks and control systems vary widely

That's why our deployment process always starts with shadow mode at the customer site. We use the playground to develop and validate algorithms, then adapt to each specific installation.

Why This Matters

Many AI companies never touch physical systems. They train on data, maybe test in simulation, and deploy hoping for the best. We think that's backwards.

Chemical plants are physical. Chemistry is physical. If you want control systems that work reliably in the real world, you need to test them in the real world—repeatedly, rigorously, and before they touch customer operations.

Our playground reactor is one of our most valuable assets. It's taught us more about practical control than any number of simulations ever could. And it continues to teach us every day.