From Prediction to Proposal in Catalysis: robust evaluation, LRP explanations, and relevance-guided candidate generation

Parastoo Semnani, Mihail Bogojeski, Florian Bley, Christoph Weisser, Jan Herrmann, Qiong Wu, Zizheng Zhang, Thomas Kneib, Klaus Robert Muller

Keywords: machine learning, explainable AI, catalyst design, material design, oxidative methane coupling, neural networks, support vector machines, decision trees

TL;DR: An ML and XAI pipeline that fairly evaluates skewed OCM data, uses signed LRP to expose high-yield elements, and turns relevance into priors for proposing generating promising catalysts

Abstract:

Machine learning (ML) can accelerate experimentation in chemistry and materials, but models trained on small, and highly class imbalanced datasets often look deceptively strong when judged by accuracy alone and provide limited guidance for follow-up simulations or experiments.
We present a robust ML and explainable artificial intelligence (XAI) framework for catalyst yield classification that emphasizes: (i) robust evaluation under class imbalance, (ii) signed, class-aware explanations via Layer-wise Relevance Propagation (LRP) for neural networks and neuralized support vector machines, and (iii) a simple relevance-guided sampler to propose promising compositions. This framework has been implemented on oxidative coupling of methane (OCM) to evaluate the performance of a range of ML models: tree-based models (such as decision trees, random forest, and gradient boosted trees), logistic regression, support vector machines, and neural networks.
The proposed framework yields reliable generalization estimates under scarcity and mitigates imbalance during training. The attribution layer interrogates model decisions: tree importances are stable but class-agnostic, whereas signed LRP isolates features that contribute positively to the high-yield class. Using these signed signals to bias a validity-preserving sampler enriches model-predicted high-yield candidates.
The resulting workflow forms a practical interface between scalable ML and experimental validation.

Release To Public: Yes, please release this paper to the public

Submission Number: 23