Accelerating Multi-Elemental Catalyst Discovery with Interpretable Machine Learning and Automated Experimentation

Multielemental catalysts (MECs) offer broad compositional freedom for tuning catalytic performance, yet practical optimization is often limited by trial-and-error synthesis and testing. Here we implement a transferable, reproducible closed-loop discovery workflow on a fully commercial robotic automation stack and couple it with machine-learning–guided optimization to accelerate MEC development for tetracycline degradation in a peroxymonosulfate (PMS)–based Fenton-like system. An adaptive-learning genetic algorithm (GA) was used to design an initial campaign of 144 MECs to train a multilayer perceptron (MLP) surrogate model. The GA–MLP closed loop then proposed and experimentally validated 25 additional candidates, identifying four high-performing MECs and increasing tetracycline degradation efficiency from ∼78% in the initial data set to 93% for the best catalyst. To substantiate the executability and reproducibility of the digital recipe, inductively coupled plasma optical emission spectrometry confirmed that recommended precursor ratios were translated into measured catalyst compositions with minor deviations, and catalysts prepared robotically and manually under identical protocols exhibited consistent degradation performance. Finally, SHapley Additive exPlanations (SHAP) analysis enabled model interpretation and revealed nonlinear composition–performance contributions, providing quantitative guidance for metal fraction and promoter loading windows. This work demonstrates a reproducible, automation-ready strategy for sample-efficient optimization of multielement catalysts for advanced oxidation processes in environmental applications.