A Framework for Predictive Catalytic Activity in Heterogeneous Systems

Thesis Abstract

In heterogeneous catalysis, predictive accuracy of activity under diverse reaction conditions remains constrained by limited mechanistic integration of catalyst surface properties, reactant dynamics, and operating environments; this study addresses the gap by proposing a formal framework that unifies physico-chemical descriptors with machine-assisted predictive modeling to forecast catalytic activity across novel heterogeneous systems. The aim is to develop a transferable framework that outputs quantitative activity predictions and qualitative mechanistic insights for catalysts under variable temperatures, pressures, and feed compositions. Specific objectives are (i) to compile a curated dataset of 120 heterogeneous catalysts spanning metal–support combinations, characterized by surface area, acid–base site density, oxidation state distributions, and predictive descriptors from density functional theory (DFT) calculations; (ii) to identify key descriptors correlating with turnover frequency (TOF) and apparent activation energy (Ea) using multivariate regression and partial least squares (PLS) analysis; (iii) to construct a hybrid predictive model integrating mechanistic insights from the Sabatier principle, the Langmuir–Hinshelwood framework, and microkinetic modeling with data-driven calibration; (iv) to validate the framework against an external test set of 30 catalysts and three representative reaction families (oxidation, hydrogenation, and coupling) under bench-scale reactor conditions; and (v) to assess transferability and uncertainty using Bayesian inference and Monte Carlo simulations. Methodologically, the study adopts a mixed-methods design. The population comprises published and in-house catalytic datasets augmented by new experimental measurements. A stratified sampling approach selects 150 catalyst instances to balance metal types, supports, and reaction classes. Data collection employs standardized physicochemical characterization (BET surface area, X-ray photoelectron spectroscopy, CO probe desorption, temperature-programmed reduction/oxidation), in-situ diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) for adsorbate identification, and operando X-ray absorption spectroscopy (XAS) to capture oxidation state dynamics. Experimental data for the validation set are obtained from bench-scale fixed-bed reactors with precise control of temperature (523–773 K), pressure (1–10 atm), and feed composition, recording TOF and Ea via Arrhenius analysis. Analytical strategy proceeds in three tiers. First, descriptive and correlation analyses identify candidate descriptors most strongly associated with activity. Second, a predictive framework is built by coupling a mechanistic base model—rooted in the Langmuir–Hinshelwood or Eley–Rideal regimes with microkinetic elaboration where applicable—with data-driven components implemented through regularized regression (ridge and LASSO) and random forest algorithms to capture nonlinear interactions. Third, the framework is calibrated and validated using cross-validation on the training set and external validation on the 30-catalyst test set; Bayesian hierarchical modeling is used to quantify uncertainties in predictions and descriptor importance. Model performance is evaluated via R-squared, RMSE, mean absolute error, and predictive intervals, while sensitivity analyses probe the robustness of descriptors across reaction families. Ethical considerations include responsible data sharing and reproducibility practices, with all experimental procedures conducted in compliance with institutional safety guidelines. Expected findings include (i) identification of a robust descriptor set—combining geometric, electronic, and thermodynamic features—that consistently predicts TOF with R2 > 0.80 across reaction classes; (ii) demonstration that integrating microkinetic constraints with data-driven corrections improves extrapolative capability to unseen catalysts; (iii) quantification of uncertainty bounds for predictions, enabling risk-informed catalyst screening; and (iv) a transferable framework demonstrated to yield mechanistic interpretations consistent with Sabatier optimality and surface intermediate stabilization trends. The study contributes to knowledge by offering a formal, transferable framework that integrates first-principles insights with empirical data to enable predictive catalysis across heterogeneous systems, advancing rational catalyst design and screening efficiency. It provides a structured approach to reconcile mechanistic theory with empirical variability and establishes a blueprint for incorporating uncertainty quantification into predictive catalysis workflows. Recommendations include expanding the framework to operando data integration, incorporating solvent and electrolyte effects for liquid-phase catalysis, and deploying the model within an active-learning loop to accelerate discovery of high-performance catalysts.