By Allison Buenemann, Product Marketing Manager at Imubit
Reinforcement learning (RL) has been proven in the optimization and control of industrial processes, but it is still new in a market where incumbent closed-loop technologies have run for decades. That newness, together with the attention that generative AI tools such as ChatGPT have drawn, has caused confusion about the different types of AI and the roles they can play in the process industries.
This post does not try to map the whole industrial AI landscape. For a wider view, see our article on reinforcement learning in industry. Here we take three of the misconceptions we hear most often about RL-based optimization running in closed loop, and explain how the technology handles each one.
Myth 1: data-first means data-only
Many AI-based approaches start by building the process model from the plant's own historical data, rather than from kinetic or first-principles models. Data is where the model starts, but first principles are not left out.
A traditional model becomes a hybrid when AI is added on top of a first-principles model. An AI-based process model works the other way round, bringing first principles into a model learned from data. Starting from data means the model is not built on limiting assumptions or human bias from the outset. It is still guided by first principles throughout development. Every modeling step involves first-principles choices made by process engineers and subject-matter experts, drawing on their chemical engineering training and plant experience.
The result is enough guardrails to keep the model physically sensible, without so many that they stop it from following what the process data shows. A data-first model does carry first principles, without the limiting assumptions a purely first-principles model has to make.
Myth 2: RL models cannot handle model mismatch or prediction error, so they need frequent, time-consuming retrains
When the optimizer is trained offline with RL, instead of solving online at each step, it works within the space the trained model knows. For a large neural network, that space can be the entire operating history of the process unit.
During training, noise is deliberately injected, so the controller learns from the actual historical scenarios and from millions of slight variations on them. That breadth of training makes the RL controller more robust to prediction error than traditional model predictive control (MPC).
With that much historical context, a change in conditions is often close enough to something the controller saw in training that it can infer the right action without a retrain. In the refining units where these models run, a genuinely unknown operating scenario comes up no more often than the normal maintenance interval for a traditional APC or optimization application. When the unit does move into new operation, the model is retrained on the new data. The retrain is quick and can run overnight, so the controller returns with that new operation inside its range.
Model mismatch therefore does not mean frequent, time-consuming retrains. Most changes in conditions fall within what the controller learned in training, and the ones that do not are handled by a quick retrain.
Myth 3: support is limited to the RL technology vendor
When RL was new to industrial plants, this was true. As interest in AI and RL has grown, so has the number of people who want to understand, implement and support it. Ease of use matters as much as interest. A no-code interface lowers the barrier for third-party and in-house support teams.
Building the algorithms into a modern software platform lets a plant rely on the domain experts it already has, in process engineering, controls, planning and economics, and operations. When process engineers can build and maintain the models themselves, the plant does not need to hire a separate AI team to support the technology.
Several large energy companies already train their own RL closed-loop models this way, without hiring AI specialists. The capability spreads across their organizations, and the chance to work with modern technology helps them attract and retain engineering talent.
How this works in DLPC®
Imubit's Deep Learning Process Control® (DLPC®) is built on these principles. A Deep Learning Predictor learns a nonlinear dynamic model of each unit from its own history, guided by first principles. A Deep Learning Controller, trained by reinforcement learning across millions of simulations, learns the control policy. It runs on-premise and writes to the APC or DCS minute by minute, inside limits set by plant engineers. In DLPC® Builder, process engineers build these models with no Python and no machine-learning background. Imubit applies the approach in refining, chemicals and polymers, cement, and mining and minerals, with over 100 closed-loop controllers deployed worldwide.
