Auditing domain alignment in latent representations for cross-domain materials machine learning

Abstract Machine learning (ML) models for materials discovery increasingly combine heterogeneous experimental and computational datasets, yet it remains unclear whether learned representations capture transferable structure–property relationships or primarily encode domain-specific biases. Here, we audit domain alignment in latent representations for cross-domain materials ML using experimentally measured porous carbons, experimentally measured metal–organic frameworks (MOFs), and simulated MOFs. A target-free neural encoder was trained on shared physicochemical descriptors with explicit missingness masks, while domain-adversarial and noise-consistency objectives were used to examine how alignment reshapes latent-space geometry. Across the alignment grid, the learned representations exhibited partial rather than complete alignment. Cross-domain neighbourhood mixing increased and global domain separability decreased, but local domain structure persisted. The selected aligned model achieved the strongest cross-domain neighbour mixing; however, improved geometry did not translate into improved cross-domain prediction. Ridge transfer probes showed degraded transfer performance for capacitance targets, demonstrating that reducing domain separability is insufficient to ensure transferable target semantics. Target-semantic probing further showed that electrochemical properties were encoded unevenly across latent dimensions. Spearman and domain-controlled partial Spearman analyses revealed persistent target associations, especially for volumetric capacitance, but these associations were strongly domain-dependent. Linear Ridge-probe attributions confirmed that different domains often relied on different latent coordinates for the same target property, indicating that alignment redistributed target relevance without producing a uniformly shared semantic representation. These results show that geometric alignment, target-semantic preservation, and predictive transfer are distinct properties of learned materials representations. We, therefore, introduce a latent-space auditing framework that combines geometry, neighbourhood mixing, target probing, transfer diagnostics, and attribution analysis to diagnose when domain alignment is only partial. This framework provides a practical route for evaluating the limits of cross-domain generalization in heterogeneous materials ML.