Catalysis 2026 Issue

AI optimisation in closed-loop control

Three case studies show how non-linear problems of the modern refining and chemicals industry are solved using AI to deliver breakthrough value

Edison Tan Imubit

R efineries and petrochemical complexes operate in an environment defined by volatility and complexity. Feedstock variability, shifting product spreads, and increasingly stringent environmental regulations converge to test the limits of traditional control systems. Units like fluid catalytic crackers (FCC), cracking furnaces, and cata - lytic reactors are the economic engines of the plant, yet they are deeply non-linear, constraint-limited, and interdepend - ent. Small inefficiencies in these units can translate to mil - lions of dollars of lost margin annually. Advanced process control (APC) and model predictive control (MPC) once revolutionised plant operation by pro - viding structured, multivariable coordination of process variables. However, these systems were built upon lin - ear approximations and static yield assumptions that are increasingly detached from modern operating realities. Their linearised models cannot capture catalyst ageing, additive chemistry, or long-term fouling. Their update cycles, weeks- to-months, leave a persistent gap between planning intent and plant execution. AI optimisation (AIO) closes that gap. By combining dynamic process models (DPMs) trained directly from plant data with reinforcement learning (RL) control agents, AIO continuously learns, forecasts, and optimises plant performance in a closed loop. It transforms operations from static, rule-based control into an adaptive system that optimises profitability, reliability, and sustaina - bility in real time. Three unique challenges where AIO deliv - ered breakthrough value include: • Learning and optimising the non-linear impact of FCC additive chemistry (Z-cat) on olefin selectivity. • Predicting and mitigating long-term fouling in cracking furnaces to improve reliability. • Managing catalyst life cycles by dynamically trading off between short-term yield and long-term activity preservation. Each example highlights a class of problems that tradi - tional APC cannot handle, where AI optimisation delivers new, data-driven intelligence to the plant. APC and MPC limitations Traditional APC and MPC assume that small changes in inputs produce proportional responses in outputs. These assumptions simplify optimisation but fail when non-lineari - ties dominate, as they do in modern units: • FCC systems : Selectivity to olefins, liquefied petroleum gas (LPG), and gasoline depends non-linearly on feed quality, riser temperature, and catalyst activity. Additive chemistry alters product distribution in complex, non-proportional ways.

• Cracking furnaces : Fouling and coking follow time- dependent kinetics that change with crude composition and run length, far beyond the reach of linear disturbance models. • Reactor systems : Catalyst deactivation and regeneration cycles introduce strong hysteresis; yield responses vary over time even under constant conditions. Linear MPC frameworks also suffer from slow model update cycles. Yield vectors used in refinery linear program - ming (LPs) may be refreshed monthly, but the plant’s actual operating regime can shift hourly. As a result, planning and operations diverge, the so-called ‘planning vs actual’ gap, leaving margin uncaptured. AIO eliminates this gap by replacing static models with continuously learning dynamic process models that adapt to changing plant conditions. These models are embedded in closed-loop RL controllers capable of optimising every few minutes. AI-driven closed-loop optimisation Imubit’s closed-loop AIO directly addresses these limitations. AIO combines two core elements: u Dynamic process models : These are high-fidelity, data- driven digital twins of process units. Trained on years of plant history, they capture non-linear cause-and-effect rela - tionships among manipulated variables (MVs), controlled variables (CVs), and key economic indicators. Unlike simpli - fied yield vectors, DPMs can represent the real curvature of product yields versus severity, additive dosage, or feed com - position. The model also considers the context of the plant by using a control window consisting of continuous history data of one to two hours for context before forward prediction. v Reinforcement learning optimisation : Once trained in a simulated environment using the DPM, RL agents learn through iterative trial-and-error how to maximise an eco - nomic objective (such as product margin or energy effi - ciency) while respecting safety and equipment constraints. The agents then operate in a closed loop, generating new optimised setpoints every few minutes, continuously adapt - ing to changing conditions. Together, they deliver real-time, economics-driven optimi - sation that connects planning intent directly to plant execution. Case 1: Learning the impact of Z-additive on FCC olefins selectivity The challenge In FCC operations, the relationship between Z-additive dos - age, riser outlet temperature (ROT), and olefin selectivity is inherently non-linear.

Catalysis 2026