DeepSciReasoner Spectrum Benchmarks

Prepared crystal graphs and spectrum targets for the six structured-spectrum benchmarks of Deep Scientific Reasoning under Physical Constraints: Structure-Aware Spectrum Prediction (ICML 2026). Code: github.com/yingheng-wang/DeepSciReasoner.

Benchmarks

Directory	Spectrum	Samples	Train / Val / Test	Bins	Spectral grid
edos_caltech	Electronic DOS, CalTech	37,565	30,052 / 3,757 / 3,756	128	[−4, 4] eV, binned
edos_tri	Electronic DOS, TRI	14,140	11,312 / 1,414 / 1,414	256	[−5, 5] eV, Gaussian-smoothed
phdos	Phonon DOS	1,524	1,220 / 152 / 152	51	[0, 1000] cm⁻¹
xanes_full	K-edge XANES, all absorbers	49,895	39,916 / 4,990 / 4,989	128	[−5, 40] eV relative to E₀, σ = 0.5 eV
xanes_o	K-edge XANES, oxygen	5,094	4,075 / 510 / 509	128	[−5, 40] eV relative to E₀, σ = 0.5 eV
raman	Raman	2,283	1,826 / 229 / 228	128	[0, 1500] cm⁻¹, σ = 10 cm⁻¹

Sample IDs are Materials Project IDs. For xanes_full the ID is <mpid>_<absorber> (one record per absorbing element). For phdos the ID is the numeric part of the Materials Project ID. Curation, smoothing and gating are described in Appendix D.1 of the paper.

Files

<benchmark>/
  manifest.csv     sample IDs, one per row (column "mpid")
  split.pkl        fixed (train, val, test) row indices into manifest.csv
  graphs.tar.gz    graphs/<id>.chkpt, one prepared crystal graph per sample
atom_init.json     92-dim atom feature table used for the node features
SHA256SUMS         checksums of all benchmark files

With the code repository, bash scripts/download_data.sh [benchmark ...] downloads, verifies and unpacks everything into DATA/<benchmark>/. Alternatively:

hf download yingheng/DeepSciReasoner --repo-type dataset --local-dir DATA

tar -xzf DATA/edos_caltech/graphs.tar.gz -C DATA/edos_caltech

Graph format

Each graphs/<id>.chkpt is a torch.saved dictionary. Load it with torch.load(path, weights_only=False); pymatgen must be installed because atom_id holds pymatgen Element objects.

Key	Shape / type	Content
atom_fea	[n_atoms, 92] float32	Atom features from atom_init.json
nbr_fea	[n_atoms, 12, 41] float32	Distances to the 12 nearest neighbours within 8 Å, Gaussian-expanded on 0–8 Å in 0.2 Å steps
nbr_fea_idx	[2, 12·n_atoms] int64	Edge index (atom, neighbour)
groups	[n_atoms] int64	Cluster label per atom (clustering of Cartesian coordinates)
enc_compo	[1, 103] float32	Fractional elemental composition (H to Lr)
coordinates	[n_atoms, 3] float64	Cartesian coordinates (Å)
target	[1, bins] float32	Spectrum. For eDOS and phDOS this is a tuple whose first element is the target
cif_id	str or int	Sample ID
atom_id	list of Element	Species of each site
absorber_z	int	Absorber atomic number (XANES); 0 for Raman; absent otherwise

Targets are stored on their native scale; the DeepSciReasoner code normalizes each spectrum by its sum.

Sources and license

The data are released under CC BY 4.0. They are derived from the following public sources; please cite them alongside the paper:

• Crystal structures and electronic structure: Materials Project (Jain et al., APL Materials 1, 011002, 2013).
• CalTech eDOS benchmark: Kong et al., Nature Communications 13, 6098 (2022).
• TRI eDOS benchmark: Mueller et al., PNAS 121, e2320134121 (2024).
• phDOS: Chen et al., Advanced Science 8, 2004214 (2021).
• XANES: Materials Project XAS database (Mathew et al., Scientific Data 5, 180151, 2018).
• Raman: Bagheri & Komsa, Scientific Data 10, 80 (2023); Materials Cloud Archive, doi:10.24435/materialscloud:ze-58.

Citation

@inproceedings{wang2026deepscireasoner,
  title     = {Deep Scientific Reasoning under Physical Constraints: Structure-Aware Spectrum Prediction},
  author    = {Wang, Yingheng and Yu, Tao and Kong, Shufeng and Ricci, Francesco and Gregoire, John M. and Gomes, Carla P.},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  volume    = {306},
  year      = {2026}
}