Code and data for "Dynamic evaluation of uncertainty quantification under distribution shift in materials property prediction"
Wenbin Wan, Kexin Liu, Shanlin Tong, Wu Lu, Liu Y, Xingwen Jiang, Jianghai Qian
This record contains the supplementary code and data supporting the manuscript “Dynamic evaluation of uncertainty quantification under distribution shift in materials property prediction.” The archive contains raw and processed data tables, crystallographic structures, Materials Project data-acquisition and processing workflows, controlled-shift construction and validation notebooks, split indices, seed-specific model configurations, raw and calibrated sample-level predictions, full-database five-fold cross-validation outputs, aggregated result tables, supplementary material, and figure- and table-generation inputs. The study covers three materials-property regression tasks: formation energy per atom, band gap, and Young’s modulus. Five uncertainty-aware regression approaches are included: deep ensembles, heteroscedastic neural networks, Monte Carlo dropout multilayer perceptrons, natural-gradient boosting, and random forests. The controlled distribution shift is constructed using unit-cell site count as a reproducible structural-size variable. The archive supports two reproduction routes: rerunning model training, calibration, evaluation, and visualization from the archived processed datasets; or reacquiring the underlying Materials Project records using a user-provided API key. README.md, REPRODUCIBILITY.md, DATA_DICTIONARY.md, and FIGURE_PROVENANCE.md describe the directory structure, execution order, variable definitions, and output provenance.