Smart Catalytic Process Optimization via AI-Driven Reaction Modelling

Thesis Abstract

The rapid evolution of catalytic processes in chemical industries demands intelligent optimization strategies that can adaptively model complex reaction networks, minimize energy consumption, and maximize product yield under variable feedstock and operating conditions. Traditional process optimization often relies on static kinetic models and rule-based heuristics, which fail to capture nonlinearities, multi-scale transport phenomena, and catalyst deactivation patterns that emerge in real-time operations. This study addresses the problem of achieving robust, data-driven optimization of catalytic processes by integrating AI-driven reaction modelling with physics-informed constraints to deliver real-time control and decision-support capabilities. The aim is to develop a unified framework, Smart Catalytic Process Optimization (SCPO), that combines machine learning, reaction engineering, and catalytic science to predict optimal operating trajectories and catalyst performance metrics under uncertainty. Specific objectives are (i) to construct a multi-fidelity modelling architecture that fuses first-principles kinetic models with neural network surrogates trained on reactor data; (ii) to embed physics-informed neural networks (PINNs) within dynamic optimal control to enforce mass, energy, and species balance as well as catalyst deactivation kinetics; (iii) to implement an active learning protocol that prioritizes data collection from experiments and pilot-scale reactors to reduce model error in critical operating regions; (iv) to validate the framework on a representative selective oxidation/coupling reaction system using real-time analytics; and (v) to quantify energy savings, carbon footprint reduction, and yield improvements relative to conventional optimization methods. The methodology adopts an iterative, data-centric research design. The population consists of pilot and bench-scale fixed-bed and slurry reactors representative of industrial hydrocarbon processing, with a target dataset of approximately 15–20 datasets per reactor type, each containing 6–12 weeks of operation. A mixed-methods data collection approach is employed high-frequency instrument data (temperature, pressure, flow rates), spectroscopic in situ analytics (Diffuse Reflectance Infrared Fourier Transform Spectroscopy, DRIFTS; Raman spectroscopy) for surface species monitoring, and off-line product analysis via GC–MS and HPLC. The analytical instruments include online gas chromatography (GC) and online mass spectrometry (MS) for real-time composition monitoring. The data will be integrated within a hierarchical modelling framework a physics-informed reactor model governing reaction kinetics and transport, augmented by neural networks capturing non-idealities and catalyst ageing. The controller employs model predictive control (MPC) enhanced with Bayesian optimization to handle uncertainty, and the learning loop uses active learning to select informative experiments. The data analysis plan includes regression analyses (Gaussian process regression for surrogate modelling and uncertainty quantification), comparative assessments via ANOVA for control strategies, and time-series analyses for dynamic performance. Theoretical grounding rests on reaction engineering, catalysis science, and AI theory, anchored by the Theory of Dynamic Optimisation and the Theory of Physically Informed Neural Networks. Key expected findings include (i) a validated SCPO framework capable of delivering near-optimal operating policies with reduced energy consumption and improved selectivity; (ii) demonstrable accuracy gains from PINN-enhanced kinetic models over conventional mechanistic models in predicting conversion and selectivity under perturbations; (iii) a quantified reduction in catalyst deactivation impact through predictive maintenance scheduling; and (iv) an actionable active learning protocol that accelerates model refinement with a limited experimental burden. The study anticipates demonstrating statistically significant improvements in yield (5–12% relative) and energy intensity (7–15% relative) across tested scenarios, with uncertainties explicitly characterized by Bayesian methods. The contribution to knowledge lies in (a) bridging mechanistic reaction engineering with advanced AI through physics-informed learning for catalytic systems; (b) delivering a scalable, real-time optimisation framework applicable to diverse catalytic processes; and (c) providing empirical evidence on the benefits and limitations of integrating active learning with MPC in industrial reactors. The conclusion emphasizes the practical viability of SCPO for industry-wide deployment, recommending phased implementation, maintenance of data quality standards, and ongoing refinement of physics-informed constraints to accommodate novel reaction chemistries as bi-metallic and nano-structured catalysts evolve.