A Framework for Predicting Catalyst Performance Using Machine Learning Models

Thesis Abstract

Catalyst performance is a critical determinant of efficiency and sustainability in numerous industrial chemical processes, yet predicting catalyst efficacy remains a complex challenge due to the multifaceted nature of catalyst-material interactions and the influence of diverse synthesis parameters. Traditional empirical methods often require extensive trial-and-error experimentation, leading to significant time and resource expenditure. This study aims to develop a comprehensive predictive framework utilizing machine learning models to accurately forecast catalyst performance based on compositional and process parameters. The primary objectives are to (1) compile and curate a robust dataset of catalyst properties and performance metrics from existing literature and experimental sources; (2) identify key features influencing catalyst efficacy through feature selection techniques; (3) evaluate the predictive capabilities of various machine learning algorithms—including random forest, support vector machines, and artificial neural networks—using cross-validation approaches; and (4) formulate an integrated framework that combines the most effective models to facilitate rapid and reliable catalyst performance prediction. The research adopts a quantitative, experimental design employing a combination of retrospective data analysis and machine learning model training. The population comprises catalyst datasets extracted from over 150 peer-reviewed publications, encompassing both characterization data (such as surface area, pore size distribution, and elemental composition) and performance outcomes (such as conversion rates, selectivity, and turnover frequency). A sample size of approximately 2,500 catalyst instances was assembled, with stratified random sampling ensuring representation across various catalytic reactions, substrate types, and synthesis methods. Data collection involved systematic extraction and standardization of parameters into a structured database, with supplementary experimental validation performed via laboratory synthesis and testing of 50 novel catalyst formulations, to enhance dataset robustness. Data analysis integrated advanced preprocessing techniques including normalization, missing data imputation, and dimensionality reduction through principal component analysis (PCA). Subsequently, feature importance was assessed through recursive feature elimination and correlation analysis. Machine learning models were trained and optimized using grid search hyperparameter tuning, with performance evaluations based on metrics such as coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE). Model validation employed k-fold cross-validation (k=10) to ensure robustness and prevent overfitting. The study also applied the theory of supervised learning and the concepts of ensemble modeling to enhance predictive accuracy, aligning with established frameworks in computational materials science. Expected findings include identification of the most significant physical and chemical parameters influencing catalyst performance, alongside the development of highly accurate predictive models with R² values exceeding 0.85 and RMSE within acceptable experimental margins. The findings will demonstrate the superiority of ensemble machine learning algorithms over individual models in capturing complex nonlinear relationships within catalyst data. The resulting predictive framework aims to significantly reduce experimental trial periods, guide targeted synthesis strategies, and facilitate the rational design of high-performance catalysts. This research contributes to the expanding field of data-driven materials discovery by providing a reproducible, scalable machine learning-based framework adaptable across various catalytic systems. It bridges computational modeling with experimental catalysis, offering a valuable tool for researchers and industrial practitioners aiming to accelerate catalyst development cycles. The main conclusion emphasizes that machine learning models, particularly ensemble approaches, can reliably predict catalyst performance with minimal experimental input, thereby optimizing resource allocation and fostering innovation in catalyst design. The study recommends expanding the dataset to include real-time process monitoring data and integrating the framework within digital twin systems for dynamic performance prediction, proposing avenues for future work in AI-driven catalysis research.