bshepp/rl-surrogate-dft-gap

Crystal-RL: surrogate-vs-DFT effective-mass gap

Two related tables from the crystal-rl materials-discovery project:

1. bootstrap_signed_mstar — 794 Quantum ESPRESSO DFT calculations on perturbed semiconductor crystal structures, with signed effective mass values (681 negative, 113 positive). Signed-m* DFT data is rare: JARVIS, Materials Project, and AFLOW all store magnitudes only, which destroys the sign that distinguishes band-inverted from normal-curvature bands. This subset is the primary reusable artifact for anyone training signed-m* surrogate models.

2. surrogate_dft_pairs — 24 surrogate-prediction-vs-DFT-truth pairs across three independent RL+surrogate validation rounds. This subset exists to support a methodological claim: across three independent pipeline variants (unsigned-m* baseline, signed-m* with bootstrap data absent, signed-m* with bootstrap data loaded), zero of 24 surrogate- selected candidates DFT-validated as low-positive-effective-mass semiconductors. 19 of 24 instead showed negative DFT band curvature (band-inversion signature). The dataset is published so other groups can study the gap directly.

Dataset summary

Subset	Records	What's in it
bootstrap_signed_mstar	794	Per-structure DFT outputs + 156-dim structural fingerprint
surrogate_dft_pairs	24	Per-candidate surrogate prediction and DFT measurement

Both subsets load with the standard 🤗 Datasets library:

from datasets import load_dataset

# Default config (the bootstrap table)
boot = load_dataset("bshepp/rl-surrogate-dft-gap", split="train")

# Explicit subset
pairs = load_dataset("bshepp/rl-surrogate-dft-gap",
                     "surrogate_dft_pairs", split="train")

Subsets

bootstrap_signed_mstar (794 rows)

Effective-mass + band-gap DFT outputs for 794 perturbed semiconductor structures, plus the 156-dimensional structural fingerprint used by the crystal-rl surrogate. Generated on AWS EC2 (c5.4xlarge) in February 2026 over 4 rounds of bootstrap exploration starting from 10 seed families (Si, Ge, C-diamond, GaAs, AlAs, InAs, GaP, SiC-3C, InP, AlN).

Column	Type	Description
row_id	int64	Stable record index (0..793). Matches the row index in the source bootstrap_expanded_all.json.
seed	string	Seed family the structure was derived from (e.g. "Si", "GaAs").
label	string	Free-form provenance label from the bootstrap pipeline.
formula	string	Chemical formula of the final structure (e.g. "As2Ga2").
dft_m_electron	float64	Signed electron effective mass at the conduction band edge (mₑ units). Negative ⇒ negative band curvature (band inversion).
dft_m_hole	float64	Signed hole effective mass at the valence band edge (mₑ units).
dft_m_min_signed	float64	Smallest-magnitude effective mass found in the band structure, signed by curvature. This is the canonical signed-m* training target.
dft_band_gap_ev	float64	PBE band gap in eV. 0 means metallic.
dft_vbm_ev / dft_cbm_ev	float64	Valence/conduction band edge positions in eV.
is_direct_gap	bool?	True if Γ-point gap, False if indirect, null if undetermined.
qe_runtime_seconds	float64	Wall-clock seconds the QE pw.x job took on a c5.4xlarge core.
fingerprint_156	list<float32>[156]	12 composition + 4 elemental property + 8 lattice + 64 RDF + 64 partial-RDF features. Composition palette: H, C, N, O, Si, P, Ge, Ga, As, In, Sn, Al, Sb, Bi, Se, Te. See qe_interface/structures.py for the exact fingerprint code.

Signed-m* distribution: 681 records with dft_m_min_signed < 0, 113 with dft_m_min_signed > 0. The skew toward negative is a property of which structures the bootstrap perturbation chose to retain — not of materials in general.

surrogate_dft_pairs (24 rows)

Side-by-side surrogate prediction and DFT measurement for the top candidates produced by three independent RL training runs.

Column	Type	Description
run_id	string	One of phase7, pathb_run1, pathb_run3. See Methodology below for what changed between runs.
formula	string	Chemical formula of the candidate produced by the PPO agent.
seed	string	Seed structure the agent perturbed from.
surrogate_m_star	float64	Surrogate's predicted effective mass for this candidate.
surrogate_reward	float64	PPO reward the agent received from the surrogate.
dft_m_min_signed	float64	DFT signed m* of the minimum-magnitude band.
dft_m_electron	float64	Signed conduction-band m*.
dft_m_hole	float64	Signed valence-band m*.
dft_band_gap_ev	float64	PBE band gap in eV.
dft_converged	bool	True if QE SCF + bands converged. (All 24 are True.)
dft_runtime_seconds	float64	Wall-clock QE time.
unusual_topology	bool	True if DFT showed negative band curvature (band-inversion signature).
is_low_positive_semiconductor	bool	Convenience: dft_m_min_signed > 0 AND < 0.5 AND dft_band_gap_ev > 0.1. True for 0 of 24 records.

Methodology

The surrogate_dft_pairs subset spans three runs of the crystal-rl pipeline:

run_id	Surrogate variant	What changed	Surrogate m* corr	DFT-validated low-pos. m*
phase7	Unsigned-m* (pre Phase-9 fix)	Original 152-dim, 12-element palette	0.928	0 / 8
pathb_run1	Signed-m*, palette expanded to 16, bootstrap data accidentally not loaded (fingerprint-dim mismatch with cached surrogate_data.npz)	156-dim fingerprint, 16-element palette, Sb/Bi/Se/Te pseudopotentials registered	0.495	0 / 8
pathb_run3	Signed-m*, palette expanded to 16, bootstrap data loaded (after zero-padding the cached 152-dim fingerprints to 156-dim via scripts/regen_bootstrap_for_palette.py)	as above	0.768	0 / 8

The cumulative finding is 0 / 24 — the same answer regardless of the surrogate's accuracy on its own validation set. The 24 candidates were picked by three different surrogate models trained on three different data mixes, but the DFT outcome is consistent: surrogate-selected low-m* predictions don't validate.

Source data & collection

• bootstrap_signed_mstar was collected by perturbing the 10 seed crystal families (formula edits, lattice scaling, position perturbation) and running PBE DFT (Quantum ESPRESSO 7.3.1, SSSP Efficiency pseudopotentials, 30 Ry wavefunction cutoff, 4×4×4 k-points for SCF, 20-point band path for the m* extraction). Effective mass was extracted from a parabolic fit to the band curvature at extrema — sign-preserving. All 794 calculations converged.
• surrogate_dft_pairs was collected by running each trained PPO agent for 30 deterministic episodes, deduplicating by formula, taking the top 8 by surrogate-predicted reward, and DFT-validating each. All 3 × 8 = 24 calculations converged.

Intended uses

• Train signed-m* surrogates without losing the negative-curvature examples. The 681 negative records are particularly load-bearing — most public datasets remove the sign.
• Benchmark RL-discovery pipelines against a known failure mode. If your method produces surrogate predictions that disagree systematically with DFT in this same way, you're hitting the same gaming dynamic.
• Study the surrogate-DFT calibration gap. The surrogate_dft_pairs subset is small (24 rows) but each pair is a controlled comparison: same fingerprint, same structure, same DFT settings, different surrogate.
• Cite the negative result. This dataset is intentionally published to make the finding citable rather than letting it die as a side comment in a repo README.

Limitations & what this dataset is not

• This is not a benchmark of "best low-m* materials" — the candidate formulas listed in surrogate_dft_pairs were chosen by an RL agent that turned out to game its surrogate. Treating them as recommended materials would invert the intended use.
• PBE-level DFT. All values are PBE outputs from Quantum ESPRESSO with SSSP Efficiency pseudopotentials. PBE is known to underestimate band gaps and, for some compositions, give qualitatively wrong band ordering. Hybrid-functional (HSE06) or GW calculations would shift some values significantly. The signed-m* sign is generally robust to functional choice; the magnitudes are not.
• Small validation cohort. 24 surrogate-DFT pairs is a strong signal for a single project but not large enough to claim a universal law about RL-surrogate gaming. We report what we observed across our 3 runs; we make no claim about every possible surrogate architecture.
• Bootstrap structures cluster around 10 seed families. They are not a uniform sample of the periodic table.

Reproducibility

• Code: github.com/bshepp/crystal-rl
  • Use tag v0.1.1 or later. The earlier v0.1.0 Zenodo archive (10.5281/zenodo.20362336) is known to be incomplete — its release notes describe code changes (Sb/Bi/Se/Te PSEUDO_MAP additions, fingerprint palette expansion to 16 elements, input_dim 152→156) that were not actually staged in the underlying commit. The AWS runs that produced the data in this dataset used a working-tree bundle that had those changes, so the published numbers are unaffected, but anyone running git checkout v0.1.0 will get code that doesn't match the methodology. v0.1.1 is what v0.1.0 was supposed to be. The concept DOI 10.5281/zenodo.20362335 always resolves to the latest version.
• Pipeline: the assembly script for this dataset is scripts/build_hf_dataset.py.
• Run artifacts: the path-B runs (pathb_run1, pathb_run3) wrote full per-stage logs, surrogate weights, PPO checkpoints, and manifests to a private S3 bucket; they are not redistributed here because they're bulky and replicable from the code. The surrogate_dft_pairs table is the distilled scientific output.
• Quantum ESPRESSO version: 7.3.1, built from source. SSSP Efficiency pseudopotentials.
• Surrogate architecture: MultiTaskMLP(156 → [192 × SiLU + LN + Drop] × 4 → m\* head + gap head), ~142k parameters, trained with two-phase protocol (JARVIS+bootstrap then frozen-trunk MP gap fine-tune).

Citation

Please cite both the dataset and the underlying software:

@misc{crystal_rl_surrogate_dft_gap_2026,
  title  = {Crystal-RL: surrogate-vs-DFT effective-mass gap (negative-result dataset)},
  author = {Sheppard, Brian},
  year   = {2026},
  url    = {https://huggingface.co/datasets/bshepp/rl-surrogate-dft-gap},
  note   = {Companion dataset to github.com/bshepp/crystal-rl}
}

@software{crystal_rl_2026,
  title     = {bshepp/crystal-rl: Signed effective-mass pipeline for RL-driven semiconductor discovery},
  author    = {Sheppard, Brian},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20362335},
  url       = {https://doi.org/10.5281/zenodo.20362335},
  note      = {Concept DOI; resolves to the latest version. v0.1.0 specifically: 10.5281/zenodo.20362336}
}

License

CC-BY-4.0. Use freely, please cite.