Alloy-XRD-Bench v1.1 Dataset Card

Summary

Alloy-XRD-Bench v1.1 contains two independent simulated powder-XRD tasks. HEA-SRO predicts masked first- and second-shell Warren-Cowley short-range order. Results are reported for ID, composition OOD, chemistry OOD, and SRO OOD splits with parent-group macro averaging. Gamma-Decomp predicts unordered gamma/gamma-prime phase/HKL-group reflection components and their center, FWHM, integrated intensity, and profile parameters from full patterns and overlap windows.

No cross-task total score is defined. HEA family and occupancy identifiability remain diagnostics rather than leaderboard tasks.

Family labels are automated diagnostics and are not expert-validated ground truth. Occupancy pairs are used only to diagnose identifiability and abstention under frozen simulated perturbations; they do not produce a v1.1 ranking score.

HEA Source And License

• Source: DFT dataset for high entropy materials, Kangming Li and Jason Hattrick-Simpers.
• DOI: 10.5281/zenodo.10854500.
• License: CC BY 4.0.
• Source CSV SHA-256: 42722654902f03efde4e1d53284741033f4c43b77d7ba5ced29e4cf873c455fe.
• Accepted canonical structures: 84,019 of 84,024 source rows.

Each released XRD sample retains a structure_state_id, xrd_sample_id, and parent_group_id. Clean, counterfactual, and perturbed views inherit one parent group and cannot cross a split boundary.

Gamma-Decomp Data

Gamma-Decomp v1.1 contains 10,000 synthetic Ni-Al gamma/gamma-prime mixtures, 175,150 raw reflection provenance rows, and 165,150 grouped full-pattern scoring components. Exactly coincident gamma-prime (221)/(300) reflections remain separate in provenance and are merged into one hkl_group_id for scoring. Train patterns expose a clean spectrum only for the frozen no-perturbation ablation; validation and test expose only observed intensity.

Evaluation

The HEA primary metric is masked parent-macro MAE. RMSE and per-shell Spearman correlations are secondary. Composition Ridge is the deterministic baseline. Peak XGBoost, ResNet1D, Spectrum Transformer, and the physics-constrained Transformer use seeds 2027-2031. Confidence intervals use 2,000 parent bootstrap replicates with seed 2027 and are reported separately from random-seed standard deviations. Gamma-Decomp reports component F1, phase/HKL-group and count accuracy, peak-parameter errors, calibration, and forward reconstruction separately for full and window views.

External Sanity Sets

The benchmark release includes four experimental patterns used only for external-validity checks. Two AlCoCrFeNi patterns come from Zenodo DOI 10.5281/zenodo.4623002 (CC BY 4.0). Two precipitation-heat-treated Ni-based superalloy patterns for IN718 and AM718R come from Apollo DOI 10.17863/CAM.87406 (CC BY-ND 4.0); the original source ZIP and .xy files are redistributed without modification. Their manifests, source hashes, and allowed claims are under data/external_sanity.

None of these experimental patterns has exact SRO or Gamma-Decomp component truth. The associated Ni study explicitly reports that gamma-prime and gamma-double-prime superlattice reflections are not distinguishable.

Limitations

All primary-task XRD patterns and v1.1 observation perturbations are simulated. The external experimental sanity patterns have no exact latent labels and are not leaderboard samples. No repeated-measurement experimental series is available, so v1.1 does not claim real-noise calibration. Atomic occupancy can remain non-identifiable from powder XRD even when the underlying computational structures differ. Repeated-measurement import and noise calibration are v1.2 release requirements.

Gamma-Decomp is not a benchmark for elemental site occupancy, phase-fraction thermodynamics, or experimental lattice-misfit inference. Its exact component truth exists because the mixtures are generated by a recorded forward model.

The immutable source is /home/GJM/data/hea/benchmark_v1; v1.1 artifacts are written only below /home/GJM/data/hea/benchmark_v1_1.