The solid electrolyte interphase in lithium batteries is a nanoscale layer where transport efficiency determines cell performance, but its atomic structure is too disordered for clean categorization. Expert classifications — crystalline vs. amorphous, bulk vs. grain boundary — impose labels that may not correspond to the environments that actually govern lithium movement.
Guan and Cai built GET-SEI, a framework that skips expert labeling entirely. Graph contrastive learning encodes each atom's local environment as a node in a graph, then learns representations by contrasting structurally similar and dissimilar neighborhoods — no predefined categories, no human annotation. The emergent clusters are then fed into dynamic mode decomposition and transition path theory to identify dominant transport routes and rate-limiting steps. Tested on sulfide and oxide electrolyte systems, the data-discovered categories identify transport-relevant environments that the conventional ontology misses.
The finding isn't that unsupervised learning is more convenient than expert labeling. It's that the data's own categories outperform expert ontologies for the downstream task. The expert labels carve the interface along axes that matter to crystallographers — symmetry, composition, phase — but lithium doesn't care about crystallographic taxonomy. It cares about local energy barriers and connectivity. The unsupervised clusters carve along axes that matter to the ion, which are different axes.
This is a claim about the relationship between classification and function. The categories that are easiest for humans to define are not necessarily the categories that govern the process. When the ontology is imposed before the data speaks, it can partition the space along irrelevant boundaries and merge regions that should be separated. Letting the structure emerge from the functional signal — transport, in this case — produces categories that are harder to name but easier to use.