Auditing Domain Alignment in Latent Representations for Cross-Domain Materials Machine Learning

Machine learning models are increasingly applied to heterogeneous materials datasets spanning different synthesis routes, measurement protocols, and structural classes. Although multi-task and representation-learning approaches are commonly used to improve predictive performance, the latent representations learned by such models are rarely examined directly, allowing domain misalignment to remain hidden until model transfer fails. Here, we present a practical, representation-level framework for auditing domain alignment in latent spaces learned by multi-task neural networks for materials property prediction. Using porous carbon materials and metal–organic frameworks as a representative cross-domain case study, we analyze a shared latent space trained across multiple target properties and regularization strategies. Global and local dimensionality-reduction visualizations are combined with quantitative measures of domain separability and latent–target correlation structure to assess alignment quality. We find that domain-alignment objectives induce measurable geometric reorganization and partial large-scale mixing in latent space, while domain identity remains locally recoverable and physically interpretable monotonic associations between latent dimensions and target properties are preserved. These diagnostics reveal structural differences in learned representations that are not apparent from conventional performance metrics alone and enable model comparison based on latent-space geometry rather than predictive accuracy. The proposed auditing framework is model-agnostic and can be readily integrated into existing materials machine learning workflows, providing a practical tool for improving robustness and interpretability in cross-domain materials modeling.