This presentation explores how modern data science methods can be used to support industrial process optimization, with a focus on applying world-model reinforcement learning to a fractionation-style process. Many industrial systems are nonlinear, dynamic, and highly interactive: a change to one controller or operating variable may influence temperatures, pressures, flows, product quality, and throughput over time. Because of this, simple one-step prediction models may be useful for forecasting, but they are often not enough to support practical operating guidance. The main concept presented here is the use of a learned world model (a data-driven simulator) that estimates how a process state may evolve after a proposed operating change.
The project demonstrates a general workflow for moving from prediction to decision support. First, historical process data is used to train a target prediction model that estimates a key production or performance outcome. Next, a separate dynamics model is trained to approximate short-term process response. This world model allows candidate operating moves to be tested through a short simulated rollout before the target outcome is evaluated. Reinforcement learning or policy optimization can then be used to search for bounded action patterns that may improve the predicted future state. This approach is valuable because it makes the recommendation process more realistic than static what-if analysis, so the action is evaluated after the process has had time to respond.
The presentation also discusses broader architecture choices for industrial world models, including feed-forward neural networks, recurrent models, temporal convolutional networks, transformers, physics-informed hybrids, and probabilistic ensembles. Each architecture has tradeoffs in speed, interpretability, data requirements, long-horizon accuracy, and deployment complexity. Finally, the discussion highlights opportunities across refining, chemicals, manufacturing, utilities, supply chain, maintenance, and other industrial domains. The central message is that world-model reinforcement learning should be viewed as a decision-support framework, a way to screen scenarios, improve understanding, and guide expert review. This is not as an immediate replacement for engineering judgment or operational governance but is a tool to help enhance these skills.
