Uncertainty-Aware ML Surrogates for DFT Strain–Bandgap Engineering in CsSnX3 (X = Cl, Br, I)

SummaryThis record contains the complete dataset and analysis code for “DFT-Informed, Uncertainty-Aware ML Surrogates for Strain–Bandgap Engineering in CsSnX₃ (X = Cl, Br, I)”. We provide the raw DFT band-gap vs strain table, cross-validated model benchmarking outputs, dense-grid surrogate predictions with uncertainties, deformation-potential tables (with propagated error bars), and the Python scripts that regenerate all tables and figures reported in the paper and SI. What’s included Data (CSV) data/band_gap_data.csv — raw DFT band gaps Eg (eV) vs strain ε (%) for CsSnCl₃, CsSnBr₃, CsSnI₃ under isotropic and uniaxial-c protocols. Derived outputs written to outputs/ml_compare/:metrics.csv, winners.csv, predictions_best.csv,deformation_potentials.csv, deformation_potentials_with_uncertainty.csv,deformation_potentials_table_wide.csv, deformation_potentials_table_wide_numeric.csv. Code (Python)compare_models.py, summary_winners.py, make_combined_plots.py,make_deformation_potentials.py, make_deformation_potential_uncert.py,make_dp_table_csv.py, make_dp_latex_table.py, make_latex_tables.py,plot_deformation_potentials.py, plus SI figure scripts:Fig_S23_uncertainty_calibration_TNR.py, Fig_S24_residuals_vs_strain_TNR.py,Fig_S25_learning_curves_TNR.py, ML_models_validation_S19.py. Docs & envSI.pdf, environment.yml, requirements.txt, README.md, REPRODUCE.md,LICENSE (MIT, code), LICENSE-DATA (CC BY 4.0, data), CHECKSUMS.txt. Methods (short)DFT (GGA-PBE, non-SOC) gaps were computed on relaxed structures across a strain grid for isotropic and uniaxial-c loading. For each material×loading curve we trained small-N surrogates Eg(ε) using a compact model set (quadratic ridge, SVR-RBF, KRR-RBF, RF, GBDT, kNN, GPR). Model selection used leave-one-out cross-validation (LOO-CV). Predictive uncertainty is analytic for GPR and bootstrap-based for non-GPR. Deformation potentials at ε=0% were obtained by symmetric finite differences with uncertainty propagation from winner-model variances. ReproducibilitySee REPRODUCE.md for step-by-step commands. In brief: create the conda env from environment.yml; run scripts/compare_models.py → metrics.csv, winners.csv, predictions_best.csv; run the make_* and plotting scripts to generate deformation potentials, tables, and figures.All outputs land in outputs/ml_compare/. Seeds and package versions are pinned. Units & conventionsStrain ε: %; band gap Eg: eV; deformation potential aga_gag: eV/%;±1σ = GP posterior σ (for GPR winners) or bootstrap σ (for non-GPR). Licensing Code: MIT (see LICENSE). Data: CC BY 4.0 (see LICENSE-DATA).Please cite both the journal article and this dataset DOI when using these resources. Raw DFT artifacts (size note)Full CASTEP outputs are large and not required to reproduce the ML results. We provide the machine-readable aggregates (band_gap_data.csv) used in all analyses. Representative input/parameter files are available from the corresponding author upon reasonable request. ContactQuestions about data or code: open an issue in the repository or contact the corresponding author.