Reaction Anomaly Mining: Systematic Detection of Unexpected Computational Outcomes to Guide Catalyst Discovery

Current computational workflows often discard “failed” or unexpected prediction results as noise, but as Worakul et al. (2024) and Staub et al. (2023) highlight, these anomalies can be rich sources of insight. I propose building a platform that systematically tracks, clusters, and explains deviations from expected catalytic trends (e.g., outliers in volcano plots, unexpected selectivity, or kinetic barriers). Using unsupervised learning and explainable AI, the system would not just flag anomalies but also generate mechanistic hypotheses—such as unusual intermediate stabilization or electronic effects—that can be tested experimentally or computationally. This approach transforms noise into knowledge, potentially uncovering new catalyst archetypes and reaction mechanisms that conventional screening would miss.

References:

1. Microkinetic Molecular Volcano Plots for Enhanced Catalyst Selectivity and Activity Predictions. Thanapat Worakul, Rubén Laplaza, Shubhajit Das, M. D. Wodrich, Cl´emence Corminboeuf (2024). ACS Catalysis.
2. Challenges for Kinetics Predictions via Neural Network Potentials: A Wilkinson’s Catalyst Case. Ruben Staub, Philippe Gantzer, Yu Harabuchi, Satoshi Maeda, A. Varnek (2023). Molecules.