Measuring the impact of materials composition on MXene property prediction using explainable AI

MXenes offer a vast compositional space for energy applications, but model accuracy depends on which materials are included in the training and testing data sets. This work measures how this fundamental choice affects predictions of voltage, capacity, charge, and electronic levels in MXenes with a variety of terminations. Two MXene data sets are encoded using structure-free elemental descriptors, and gradient boosting and models are trained and analyzed with Shapley-based explainability, addressing the role of descriptive feature-space and the material instance-spaces. We use Shapley value analysis and residual decomposition to quantify instance-to-instance influence within the residuals. RSHAP is used to rank which materials improve generalization and which degrade it, identify cross-influences between samples, and exposes weaknesses that can potentially be remediated. The results show that a small subset of compositions disproportionately controls the model error, regardless of how sophisticated the features used to describe them. The study yields practical guidance for curating MXene training sets to maximize model performance. • M, X, T and Z influence the prediction of MXene electronic and energy properties in different ways. • RSHAP quantifies pairwise interactions between M, X, T and Z that determine model performance. • Forensic analysis using RSHAP explains the impact of data remediation strategies to improve predictions.