Materials representations rank their own training data: without target labels, a crystal-graph encoder picks the compounds that best supervise a new element
Synopsis
When target compounds contain elements absent from training, it has been unclear which compounds of other elements supply transferable supervision. A pretrained crystal-graph encoder's representation ranks elements by the supervision their compounds provide without using target-property labels; at fixed training-set size, excluding the compounds of its six top-ranked elements raises formation-energy and bandgap errors by 100–110 meV atom-1 and 42–85 meV, while excluding six lower-ranked elements changes little. The rankings hold across properties, regressors and datasets, indicating that the representation itself organizes supervision and that the ranking can be identified before model training.
Interpretation
A materials representation can itself rank the element-indexed compounds of a training pool by how useful they are for a new element, without target-property labels and without a trained regressor. Judging which training data help a new chemistry previously relied on influence functions, data valuation or uncertainty-based acquisition, all of which need target-property labels and typically an already trained model; elemental-property similarity and periodic-table position are independent of the property being predicted. In source-exclusion experiments at fixed training-set size, removing the six source subsets top-ranked by the pretrained crystal-graph encoder raised formation-energy error by 100–110 meV atom-1 and bandgap error by 42–85 meV, whereas removing ranks 11–16 changed formation energy by less than 5 meV atom-1 and bandgap by less than 2 meV.
The ranking comes from geometry in representation space: as an element's atomic fraction in a compound rises, the compound moves along an approximately linear direction, and elements whose directions are closer can better supervise each other's compounds. Engineered descriptors rank largely by elemental-property similarity, while the pretrained crystal-graph encoder ranks by the geometry of its learned representation, and the two orders agree only moderately. Directional alignment is the cosine similarity of direction vectors fitted to representation centroids across five atomic-fraction intervals; computed separately from JARVIS, MP and OQMD, the element-averaged Spearman rank correlations were 0.92 for JARVIS–MP, 0.83 for JARVIS–OQMD and 0.82 for MP–OQMD.
Excluding source subsets by directional alignment raises error more than excluding by elemental-property similarity or same-periodic-group membership, indicating that the encoder captures how elements behave within compounds rather than elemental similarity alone. Elemental-property similarity agrees only moderately with directional alignment, and the same-group control removes a number of source subsets that varies with the held-out element. On OQMD, excluding the top six source subsets by directional alignment raised formation-energy MAE by about 100–110 meV atom-1 and bandgap MAE by about 42–85 meV across the three regressors; excluding the top six by elemental-property similarity also raised MAE but by less, and the same-group control raised it by less as well.
Perspective
The conclusions rest on a leave-one-element-out protocol: all compounds containing the held-out element are excluded from training and form the target set, while the remaining element-free compounds form the training pool. Evaluation covers formation energy and bandgap in the JARVIS, MP and OQMD first-principles databases, paired with random forest, extreme gradient boosting and radial-basis-function support-vector regression. Source-exclusion experiments fix the training-set size at 10,000 compounds, and the main-text results use filtered test sets that exclude compounds containing any of the six top-ranked source elements.
Directional alignment needs no target-property labels but is influenced by the collection of compounds used as probes; the Spearman correlations between dataset rankings range from 0.82 to 0.92 rather than being identical. The error increases from excluding the top six source subsets are given as approximate values and vary with dataset and regressor.
