A Framework for Predictive Catalytic Performance in Industry-Grade Processes

Thesis Abstract

Industrial catalytic processes face variability in feedstock quality, operating conditions, and reactor aging, which collectively erode long-term predictive accuracy for performance, selectivity, and catalyst lifetime. This study addresses the gap between empirical performance benchmarks and robust, deployable predictive frameworks capable of guiding industrial operators in process optimization and catalyst management. The aim is to develop a framework that integrates mechanistic catalytic understanding with data-driven predictions to forecast catalytic activity, selectivity, and deactivation trajectories under real-world industrial conditions. Specific objectives include (i) to characterize key factors governing catalytic performance across representative refinery and chemical-processing platforms; (ii) to develop a hybrid predictive model combining microkinetic/steady-state reactor theory with machine learning to capture both mechanistic dependencies and complex nonlinearities; (iii) to quantify uncertainty and establish reliability bounds for predictions across process scales; (iv) to validate the framework against pilot-plant and production-scale data; and (v) to provide decision-support guidelines for catalyst selection, regeneration scheduling, and process control. The methodology adopts a mixed-methods approach under a pragmatist paradigm. The population comprises industrially relevant catalysts used in naphtha reforming, hydrocracking, and petrochemical hydro-processing. A stratified sample of 120 datasets is assembled from collaborating refinery partners, including 40 datasets for reforming, 40 for hydrocracking, and 40 for hydro-processing, each containing feed composition, temperature, pressure, contact time, catalyst age, and measured performance metrics over multiple cycles. Data collection employs instrumented process logs, operando spectroscopy (Raman, FTIR) for surface intermediates, inductively coupled plasma–mass spectrometry (ICP-MS) for metal leaching, and chromatographic analysis (GC-FID/MS) for product yields. Instrument validation ensures measurement uncertainties are below 4% for major species and 6% for trace components. A complementary laboratory program with 12 carefully selected catalysts is conducted to generate high-resolution microkinetic data and deactivation pathways, including temperature-programmed oxidation and chemisorption studies. Analytical methods comprise a layered modeling strategy (i) microkinetic modeling to establish mechanistic correlations between active-site coverage, reaction steps, and observed rates; (ii) a Bayesian hierarchical framework to fuse mechanistic models with empirical data, enabling uncertainty quantification across scales; (iii) machine learning modules (Gaussian process regression and gradient boosting) to capture nonlinear interactions and latent variables such as support effects and mesostructural features; (iv) time-to-deactivation models to predict catalyst lifetime under varying regimes; and (v) sensitivity analysis and global optimization to identify dominant factors and optimal operating windows. Model validation uses k-fold cross-validation, back-testing on hold-out refinery data, and comparison against conventional first-principles kinetics. Hypothesis testing includes ANOVA to assess factor significance and likelihood ratio tests to evaluate model improvements from hybridization. The study also incorporates theoretical framing from transition-state theory and attrition theory for catalyst aging, complemented by reliability-centered maintenance concepts. Expected findings indicate that the hybrid framework yields statistically significant improvements in predictive accuracy over purely mechanistic or purely data-driven models, with reductions in error metrics such as RMSE by 25–40% and improved calibration of predictive intervals (coverage within 95% credible bounds in over 92% of cases). The model is anticipated to reveal critical interactions between feedstock sulfur species, metal dispersion, and support acidity that govern selectivity shifts and deactivation rates, enabling actionable insights for regeneration scheduling. The framework is expected to generalize across processing platforms, with transferability demonstrated through successful cross-domain validation between reforming and hydro-processing datasets. Contributions to knowledge include (i) a validated hybrid predictive framework that unites microkinetic insight with data-driven adaptation for industrial catalysis, (ii) a quantified uncertainty-aware decision-support tool for catalyst management and process control, and (iii) a methodological template for integrating operando measurements with large-scale production data to advance predictive catalysis. The study concludes that integrative modeling enhances resilience of industrial operations to feedstock and aging variability and recommends adopting the framework in pilot-plant trials, expanding data-sharing collaborations with industry, and extending the approach to emerging catalytic systems such as biobased feedstocks and heterogeneous electrocatalysis.