A Machine Learning\nand Explainable AI Framework Tailored\nfor Unbalanced Experimental Catalyst Discovery

The successful application of machine learning (ML) in\ncatalyst\ndesign has been made difficult by the challenges associated with collecting\nhigh-quality and diverse data. Due to the complex interactions between\ncatalyst components, the design of novel catalysts has long relied\non trial-and-error, a costly and labor-intensive process that results\nin scarce data that is heavily biased toward undesired, low-yield\ncatalysts. Such data presents a challenge for training ML models that\ngeneralize well to novel compositions, which is necessary for the\nsuccess of ML-guided catalyst discovery. Despite the growing popularity\nof ML applications in this field, most efforts so far have not focused\non dealing with the challenges presented by such experimental data.\nIn this work, we introduce a robust ML and explainable artificial\nintelligence (XAI) framework that incorporates a series of well-established\nML methods designed to improve model performance and provide reliable\nevaluations for catalytic yield classification in the context of scarce\nand class-imbalanced data. We apply this framework to classify the\nyields of different catalyst combinations in the oxidative coupling\nof methane reaction and use it to evaluate the performance of a range\nof ML models: tree-based models (such as decision trees, random forest,\nand gradient boosted trees), logistic regression, support vector machines,\nand neural networks. Our experiments demonstrate that the methods\nused in our framework lead to more robust performance estimates and\nreduce the effect of class imbalance on model training, resulting\nin significant improvements in the predictive capability of all but\none of the evaluated models. Additionally, the XAI component of the\nframework analyzes the decision-making process of each ML model by\nidentifying the most important features for predicting catalyst performance.\nOur analysis found that XAI methods that provide class-aware explanations,\nsuch as Layer-wise Relevance Propagation, managed to identify key\ncomponents that contribute specifically to high-yield catalysts. These\nfindings align with chemical intuition and existing literature, reinforcing\ntheir validity. We believe this framework can serve as a blueprint\nand a set of best practices for ML applications in catalysis, driving\nfuture research while delivering robust models and actionable insights\nthat can assist chemists in designing and discovering novel catalysts\nwith superior performance.