Catalysis 2026 Issue

• Continuous learning : Models are retrained periodically to incorporate new data, cata- lyst batches, and equipment changes, maintaining perfor- mance over years. • Integration with LP and planning : AIO’s learned yield curves feed back into the refin - ery’s planning models, ensur- ing that LP vectors reflect true non-linear plant behaviour and

Reduce severity (preserve cat life)

Optimised yield prole

Historical yield

Increase severity (spend cat life)

Figure 3 Yield profile of reactor over one-year cycle showing potential optimised moves

severity can produce a positive ROC initially, the long-term consequence is a more negative ROC as deactivation accel- erates. The RL controller then optimised an economic objec- tive function over the full cycle, weighing near-term yield gains against projected catalyst life. In practice, the resulting policy naturally adapts to catalyst condition: • Early in the cycle, it favours higher severity and conversion to capitalise on high activity. • Later in the cycle, it moderates operation to slow deactiva- tion and extend run length. Since the controller continuously reassesses catalyst state, it can shift between these regimes dynamically as plant con- ditions evolve (see Figure 3 ). Results This adaptive policy increased cycle-averaged profit while extending effective catalyst life. At the reference site, AIO achieved a ~1.9 % improvement in total yield across the cat- alyst life period, driven purely by smarter control of degrada- tion and feed composition. These results translated directly to higher total production without requiring new hardware or feed upgrades. Insight and explainability Beyond optimisation, the model provided engineers with new diagnostic clarity. Once the model differentiated short- and long-term effects, it uncovered a previously unconfirmed relationship: the purity of one feed component had a measur- able, cumulative impact on catalyst life and yield decay rate. This finding validated long-standing operational hypothe - ses that had been impossible to quantify using conventional analysis. The result is a unified, explainable framework where engineers can see how each decision affects both today’s yield and next week’s catalyst condition, an essential step toward truly cycle-aware optimisation. Integration, explainability, and organisational adoption Deploying AI in control rooms requires both technical robust- ness and human confidence. This AIO platform incorporates several design principles that enable both: • Transparent models: Variable-importance plots, response surfaces, and prediction-validation dashboards let engineers see how the model makes decisions. • Gradual rollout: Controllers start in advisory mode, gen- erating real-time recommendations before closing the loop, allowing operators to compare AI guidance with their intuition.

closing the loop between economics and execution. Together, these practices ensure safe, explainable, and sustainable adoption across planning, operations, and reli- ability teams. Path to autonomous plant The implications of closed-loop AI optimisation extend beyond individual units: • Multi-unit coordination : AIO can connect dynamic models across FCC, hydrotreaters, and fractionation systems, allow- ing RL agents to balance feed allocation and constraint nav- igation at the site level. • Sustainability integration : By embedding CO₂ and energy terms in the optimisation objective, plants can reduce emis- sions and fuel use while maintaining profitability. • AI governance and safety : Explainability features and human-in-the-loop design ensure AI decisions remain audit- able and aligned with operational standards. As AI-driven optimisation becomes commonplace, engi- neers’ roles evolve from manual tuning to strategic oversight, validating models, optimisation objectives, and guiding digi- tal twins toward autonomous operation. Conclusion The refining and petrochemical industry faces challenges that defy linear assumptions, including non-linear catalyst behaviour, long-term fouling dynamics, and cycle-depend- ent economics. AI optimisation technology addresses these head-on by learning directly from plant data, predicting sys- tem behaviour over time, and continuously optimising oper- ation in a closed loop. The three case studies presented demonstrate how AI can tackle problems once considered too complex or slow-mov - ing for control systems. Each delivers tangible economic and reliability gains while enhancing operator understanding of unit behaviour. In the same way that APC became standard for every major unit decades ago, AI-driven closed-loop opti- misation enables plants to operate at the pace of markets, not planning cycles. By bridging economics and execution through continuous learning, AIO transforms the modern plant into a proactive, self-optimising system that continu- ously balances profitability, reliability, and sustainability. Edison Tan is a business consultant at Imubit, leveraging his decade of experience as an engineer specialising in process optimisation, APC, and data-driven decision-making to improve refinery performance. Email: Edison.tan@imubit.com

Catalysis 2026