yingheng/mp-predicted-edos-phdos

Predicted eDOS / phDOS for Materials Project materials without DFT labels

Machine-learned electronic and phonon density-of-states predictions for every non-deprecated Materials Project (MP) material missing the corresponding DFT label (MP data release 2026-04-13; 154,377 non-deprecated materials total):

• 91,405 materials without an MP electronic DOS → edos_predictions.npz
• 127,768 materials without MP phonon data → phdos_predictions.npz

Shapes are predicted with the xtal2dos two-stage ensemble (Stage-1 graph-attention encoder–decoder + Stage-2 mass-conserving MIR refinement); the eDOS absolute scale comes from a separate ALIGNN regressor that predicts N_win, the number of states within ±4 eV of the Fermi level (DeepDOSReasoner scheme). Coverage is 100% of each missing-label set.

Files

edos_predictions.npz

array	shape	description
ids	(91405,)	MP material ids
energy_grid_ev	(128,)	linspace(-4, 4, 128), eV; VBM-aligned for gapped materials, E_F for metals (training-label convention)
dos_norm	(91405, 128)	sum-normalized total eDOS shape (rows sum to 1)
dos_states_per_ev	(91405, 128)	absolute total eDOS in states/eV = dos_norm × label_sum_pred
n_win_pred	(91405,)	ALIGNN-predicted # states in [E_F−4, E_F+4] eV
label_sum_pred	(91405,)	16 × n_win_pred (bin-width bridge, corr 0.9996 vs ground truth)
nelect	(91405,)	analytic valence-electron count (Σ MP POTCAR ZVAL)
f_block_flag	(91405,)	True (17,119 materials): f-element compound — absolute scale is extrapolation (shape unaffected)

phdos_predictions.npz

array	shape	description
ids	(127768,)	MP material ids
freq_grid_cm1	(51,)	linspace(0, 1000, 51) cm⁻¹
phdos_norm	(127768, 51)	sum-normalized phDOS shape
phdos_3n	(127768, 51)	physical phDOS in states/cm⁻¹ per unit cell, rescaled so ∫g(ω)dω = 3·n_sites (acoustic sum rule)
n_sites	(127768,)	atoms per unit cell

materials_index.csv.gz

One row per non-deprecated MP material (154,377): formula, spacegroup, energy_above_hull, band_gap, efermi, is_metal, theoretical, MP label flags (has_edos, has_phonon) and prediction flags (pred_edos, pred_phdos).

Quick start

import numpy as np
d = np.load("edos_predictions.npz")
i = list(d["ids"]).index("mp-aaahikhm")
dos = d["dos_states_per_ev"][i]        # states/eV on d["energy_grid_ev"]

p = np.load("phdos_predictions.npz")
j = list(p["ids"]).index("mp-aaahikhm")
g = p["phdos_3n"][j]                   # states/cm^-1 on p["freq_grid_cm1"]

Models & validation

• eDOS shape: xtal2dos Stage-1 (r1/r2, aligned 128-bin σ=0.5) + Stage-2 matched-arch hybrid MIR ensemble (k=5); trained on 37,565 MP materials.
• phDOS shape: xtal2dos Stage-1 ensemble (r1/r2), 51-bin; trained on 1,524 phonon-database materials (test r² 0.729, MAE 0.00671, WD 0.0667).
• eDOS scale: ALIGNN predicting R = N_win/NELECT (test n=4529: N_win MAE 3.58, r² 0.982, MAPE 5.9%).
• The standalone featurizer was verified byte-identical to the training pipeline; inference reproduces frozen reference outputs to ≤5e-5.

Caveats

• ML estimates from structure + composition only; treat chemistries far from the training distributions with care (the phDOS training set is small).
• phDOS: ~6% of materials (light elements) have real modes above 1000 cm⁻¹; their spectra are truncated and the 3N rescale over-weights the window.
• f_block_flag=True eDOS rows: absolute scale is out-of-distribution extrapolation; use dos_norm and treat the magnitude as unreliable.
• eDOS energy zero follows the training convention (VBM for gapped, E_F for metals); small energy-axis misalignments vs. DFT are possible.

Attribution

Input crystal structures and metadata derive from the Materials Project (CC-BY-4.0; release 2026-04-13). Please cite the Materials Project alongside this dataset.