shiqiao123/umlip-dft-preopt

uMLIP DFT Pre-optimizer Benchmark

Data for "Benchmarking universal machine learning interatomic potentials as DFT pre-optimizers" (Liu, Wang, Yang, Shen; npj Computational Materials, submitted 2026).

Six universal machine-learning interatomic potentials (uMLIPs) pre-relax perturbed crystal structures under one frozen protocol. Every pre-relaxed structure - the handoff - is then given to a real VASP relaxation that is run to its endpoint, and the potential is judged by what that continuation does: how many ionic steps and SCF iterations it spends, whether it converges, and whether it lands in the same energy basin as a from-scratch DFT relaxation of the same input. A DFT-free Pre-optimization Quality Index (PQI), calibrated on the 40-structure bulk subset and then applied frozen, decides which handoffs are safe to accept.

This repository is the data and leaderboard home. The code home is github.com/ShiQiaoL/umlip-dft-preopt: relaxation drivers, the standalone VASP convergence parser, the scoring analyses, the figure builder, and the benchmark harness you need to add your own potential.

• Leaderboard Space: shiqiao123/umlip-dft-preopt-leaderboard
• Machine-readable leaderboard at the repo root: leaderboard.csv, leaderboard.md

---

The census, stated honestly

200 structures are runnable. 159 have a converged from-scratch DFT reference. 148 were scored in the paper.

These three numbers are all correct and they are not the same number. Every input structure can be relaxed by any potential and carries a PQI score, because the index needs no DFT. Only the 159 with a converged reference can be scored against a DFT endpoint. Only 148 of those had converged when the paper's scoring set was frozen: the 40-structure in-domain calibration subset plus 108 transfer cases. The remaining 11 references converged after the freeze and are shipped, flagged, and joined to nothing.

Nothing here implies that 200 structures have references.

Subset / domain	Inputs	Converged reference	Scored in the paper
in-domain calibration (calibration / mp_bulk)	40	40	40
transfer bulk	50	45	45
transfer 2d_monolayer	40	25	25
transfer heterostructure_or_interface	50	32	22
transfer defect_or_doped	20	17	16
total	200	159	148

Counts reproduced from the shipped structures config; the 11-row gap between "converged reference" and "scored" is the scoring_status == 'converged_after_freeze' flag. To reproduce a published number, filter on in_frozen_scoring_set.

---

What is measured

The frozen pre-optimization protocol (identical for all six potentials, and for yours): ASE FIRE (dt=0.1, dtmax=1.0, maxstep=0.04) on a UnitCellFilter, run to fmax = 0.03 eV/A or a 500-step cap.

The judgement is not the handoff geometry itself but the DFT continuation started from it, compared against a from-scratch DFT relaxation of the same perturbed input. Reference and continuation differ only in whether a potential moved the atoms first, which is what makes the comparison honest.

The acceptance decision is hri_score >= 70, computed from the handoff structure alone with no DFT. Applied to the 648 transfer rows it selects exactly the 107 handoffs in the validation config - the ones that were then actually run to a DFT endpoint, so that "accepted" could be checked against reality.

The six potentials: CHGNet, DPA-3.1-3M, M3GNet, MACE, MatterSim-v1.0.0-5M, NequIP-OAM-XL. Checkpoints are not redistributed anywhere in this project; the GitHub repo's environment/model_deployment_manifest.csv records the exact deployed checkpoint, training corpus and environment for each.

The shipped tables label one row MACE-MP-0. The checkpoint that actually ran everywhere in this work is MACE-MPA-0 (medium); the label is left as it ran, and the deployment manifest on GitHub is authoritative.

The four domains

domain (in structures)	n	What it is
mp_bulk	40	Materials Project bulk crystals - the in-domain calibration subset. Every predictor and the PQI were fitted here.
bulk	50	3D periodic crystals, the transfer-set control for the calibration 40
2d_monolayer	40	Single layers in a vacuum slab
heterostructure_or_interface	50	Two dissimilar monolayers stacked into an interface; the genuinely out-of-distribution domain
defect_or_doped	20	Supercells with one atom removed

Every input is deliberately displaced, and undoing that displacement is the work being benchmarked. Recipes differ by domain (positions-only jitter; positions plus cell strain; vacancy-only, unperturbed) and are recorded per structure in perturbation_kind, perturbation_amplitude_ang and cell_strain_amplitude_pct.

What the benchmark found, in five numbers

All read from the shipped tables, not restated from memory:

• In domain, all 240 continuations converged (convergence_success_binary is 1 for every row), with a median paired ionic-step saving of 0.400 of the from-scratch cost (eta_ionic_steps).
• Out of domain, the safe-handoff prevalence collapses by domain: 96/270 bulk, 107/150 2D, 5/132 heterostructure, 40/96 defect - 248/648 pooled.
• The frozen gate accepts 107 of 648 handoffs.
• Of those 107 run to a DFT endpoint: 100 converged, 2 finished unconverged, 5 failed; 99 of the 100 landed in the reference basin.
• Restarting 22 already-failed from-scratch relaxations from a potential's geometry recovered 15 of 22.

---

Configs

Seven configs, one per measurement layer of the benchmark. The split is not cosmetic: each layer has a different row grain, and joining them by hand is the point.

Config	Rows	Cols	Grain	Source tables in the GitHub repo
structures	200	45	one benchmark structure	data/structures/ (manifest, references, MP parents) + the 359 POSCAR files
handoffs	240	132	one in-domain handoff and its DFT continuation	data/calibration/calibration_predictor_outcome_master.csv + hri_no_dft_readiness_table.csv
transfer_scores	648	26	one transfer handoff, frozen-scored	data/extension/transfer_frozen_hri_scores.csv + transfer_predictor_table.csv + training_overlap_audit.tsv
validation	107	34	one accepted handoff actually run to a DFT endpoint	data/validation_arm/ (verification + basin)
recovery	44	49	one restart of a failed relaxation from a potential's geometry	data/recovery_arm/rescue_arm_*
recovery_control	13	64	one conventional settings-only restart of the same failures	data/recovery_arm/control_arm_*
leaderboard	6	13	one published potential	generated by scripts/benchmark/make_leaderboard.py

Why these seven and not fewer: structures is the only config a user needs to run the benchmark at all, so it stands alone and carries the geometries. handoffs and transfer_scores are the two scoring populations (calibration and frozen transfer) and cannot be stacked - they have different columns, different reference protocols and different energies, and stacking them would invite exactly the cross-subset energy comparison the protocol forbids. validation, recovery and recovery_control are three separate DFT campaigns with disjoint schemas; the two recovery configs are kept apart because the control arm carries INCAR-level restart diagnostics that have no counterpart on the MLIP side, and because merging 44 rows of one kind with 13 of another under a single arm column would leave most cells empty. leaderboard is the six-row summary the Space renders.

Not shipped here as a config: the cross-cutting robustness ledgers (data/ledgers/), including results_ledger.csv - one row per number reported in the paper. Those are paper-index tables rather than benchmark measurements, and they stay on GitHub.

Quickstart

from datasets import load_dataset

structures      = load_dataset("shiqiao123/umlip-dft-preopt", "structures",       split="train")  # 200
handoffs        = load_dataset("shiqiao123/umlip-dft-preopt", "handoffs",         split="train")  # 240
transfer_scores = load_dataset("shiqiao123/umlip-dft-preopt", "transfer_scores",  split="train")  # 648
validation      = load_dataset("shiqiao123/umlip-dft-preopt", "validation",       split="train")  # 107
recovery        = load_dataset("shiqiao123/umlip-dft-preopt", "recovery",         split="train")  #  44
recovery_ctrl   = load_dataset("shiqiao123/umlip-dft-preopt", "recovery_control", split="train")  #  13
leaderboard     = load_dataset("shiqiao123/umlip-dft-preopt", "leaderboard",      split="train")  #   6

Getting a usable structure object out of a row:

from pymatgen.core import Structure

row = structures[0]
s = Structure.from_str(row["input_poscar"], fmt="poscar")   # the perturbed input
if row["reference_poscar"] is not None:
    ref = Structure.from_str(row["reference_poscar"], fmt="poscar")  # the DFT endpoint

import io
from ase.io import read

atoms = read(io.StringIO(row["input_poscar"]), format="vasp")

Or without any structure library, straight from the parsed struct column:

lattice = row["input_structure"]["lattice"]        # 3 x 3, angstrom
species = row["input_structure"]["species"]        # list[str], length n_atoms
frac    = row["input_structure"]["frac_coords"]    # n_atoms x 3, fractional

Three things you will want to do immediately:

import pandas as pd

st = structures.to_pandas()
tr = transfer_scores.to_pandas()

# 1. the paper's scoring set
paper = st[st.in_frozen_scoring_set]                       # 148 rows

# 2. what the frozen acceptance gate accepts
accepted = tr[tr.hri_score >= 70]                          # 107 rows, exactly the `validation` config

# 3. join a transfer score back to its geometry
merged = tr.merge(st, left_on="case_name", right_on="structure_id")

---

How the configs join

structure_id in structures is the join key for everything, and it is also the POSCAR filename stem in the GitHub repo. It appears under different names by config, because the names are kept exactly as the analysis produced them:

Config	Column holding structure_id	Also keyed by
structures	structure_id	global_index 1-200; sample_index 1-40 (in-domain); array_task_id 1-160 (transfer)
handoffs	folder_name	sample_index + model_short
transfer_scores	case_name	array_task_id + model
validation	case_name	task_id; model
recovery	case_name	task_id; model
recovery_control	case_name	restart_task_id

Five in-domain ids contain parentheses (021_mp-19492_LiMn(PO3)3). They are string values, not paths, and need no escaping here.

Domain vocabulary differs between configs and this is a real trap. The campaign tables use short domain codes. They are not renamed here, because renaming them would break the join back to the GitHub tables.

structures.domain	transfer_scores / validation / recovery	recovery_control
mp_bulk	- (the in-domain set lives in handoffs)	-
bulk	bulk	-
2d_monolayer	2d	2d_monolayer
heterostructure_or_interface	hetero	-
defect_or_doped	defect	-

Reference energies are not comparable across the two subsets, and only loosely comparable across domains within the transfer set. The calibration 40 ran PBE+U at ENCUT 520 eV with tetrahedron smearing and an energy convergence criterion; the transfer 160 ran PBE+D3(BJ) at 500 eV with Gaussian smearing and a force criterion (EDIFFG = -0.02 eV/A). reference_energy_ev is meaningful per structure - against that structure's own continuation, which used the matching protocol - and as a basin check. It is not a formation energy and not a cross-structure ranking. For the same reason, do not filter the calibration 40 on residual force: those runs converged on energy, so 34 of the 40 sit above 0.05 eV/A and are nonetheless correctly converged. Filter on outcar_reached_required_accuracy instead.

---

Column dictionaries

Units are in the column names throughout: _ang angstrom, _ev electronvolt, _ev_ang eV/A, _pct percent, _deg degree, _sec second, mev_atom meV per atom.

structures - 200 rows, 45 columns

Column	Meaning
structure_id	Join key and POSCAR filename stem
subset	calibration (the 40 calibration structures) or extension (the 160 transfer structures)
domain	mp_bulk, bulk, 2d_monolayer, heterostructure_or_interface, defect_or_doped
global_index	1-200, unique across both subsets
sample_index	1-40, calibration subset only; the key most handoffs columns use
array_task_id	1-160, transfer subset only
case_dir	Transfer run-directory name
reference_folder_name	Calibration-subset DFT archive directory name (40 rows; null for the transfer subset). It equals structure_id for 37 of the 40; for three ids the archive name is truncated at the first ( (021_mp-19492_LiMn(PO3)3 -> 021_mp-19492_LiMn), so join on structure_id, not on this
n_atoms, formula	Atom count and realized composition, read from the shipped file
crystal_system	One of cubic / hexagonal / monoclinic / orthorhombic / tetragonal / triclinic / trigonal
space_group	Symbol and number, e.g. Fd-3m (227). Transfer subset only
source_db, source_id, parent_ids	Provenance: which database the parent came from, its id, and (for heterostructures and vacancies) the parent structures it was built from
mp_parent_id, mp_parent_derivation	Materials Project parent entry and how the input was derived from it. Populated for 90 of the 200
perturbation_kind	uniform_cartesian_jitter, uniform_cartesian_jitter_plus_cell_strain, or none
perturbation_amplitude_ang, cell_strain_amplitude_pct	Nominal perturbation amplitudes
input_file, input_sha256	Path within data/structures/ in the GitHub repo, and the SHA-256 of that file
input_poscar	The perturbed starting geometry, as the verbatim POSCAR text
input_structure	The same geometry parsed: {lattice: 3x3 float64, species: list[str], frac_coords: n_atoms x 3 float64}
dft_reference_status	converged (159), failed (34), timeout (4), completed_unconverged (2), not_run (1)
dft_reference_available	Is there a converged from-scratch DFT endpoint. True for 159
in_frozen_scoring_set	Was this structure scored in the paper. True for 148
scoring_status	in_paper_scoring_set (148), converged_after_freeze (11), null (41, no reference)
reference_file, reference_sha256	As above, for the reference. Null where there is none
reference_poscar	The converged from-scratch DFT endpoint, verbatim POSCAR text. Null for 41 structures
reference_structure	The parsed reference geometry, same struct layout. Null for 41 structures
reference_energy_ev, reference_energy_per_atom_ev, reference_energy_sigma0_ev	Final total energy (TOTEN), the same per atom, and final energy(sigma->0)
reference_max_force_ev_ang	max norm(F_i) at the final ionic step - the maximum per-atom force magnitude, not the largest component
reference_volume_ang3	Cell volume of the endpoint
reference_converged_ionic, reference_converged_electronic	Parsed convergence flags
outcar_reached_required_accuracy	VASP's own marker. This is the criterion that decided the census
reference_ionic_steps	Ionic steps the from-scratch run took - the cost baseline the continuations are scored against
ediffg_ev_ang	VASP's EDIFFG, as it ran. Negative is a force criterion in eV/A (-0.02 for the whole transfer subset); positive is VASP's energy criterion in eV, which is what the calibration 40 used (0.0005 to 0.05), so the _ev_ang suffix does not apply to those rows
force_converged_vs_ediffg	Transfer subset only: does the final force meet EDIFFG
training_overlap_max_risk	Transfer subset only. Highest overlap class between this structure and any potential's training corpus: direct_id, derived_from_training_id, probable_structural, possible_structural, low
training_overlap_direct_or_derived_id	Transfer subset only. 1 if some model's corpus contains this exact id or a direct parent of it

Renamed from the GitHub source for clarity within this config: references.csv:converged_ionic -> reference_converged_ionic, converged_electronic -> reference_converged_electronic, ionic_steps -> reference_ionic_steps; materials_project_parents.csv:mp_id -> mp_parent_id, derivation -> mp_parent_derivation; training_overlap_audit.tsv:max_risk_any_corpus -> training_overlap_max_risk, direct_or_derived_id_any_model -> training_overlap_direct_or_derived_id. Nothing else is renamed, in any config.

handoffs - 240 rows (40 calibration structures x 6 potentials), 132 columns

One row is one in-domain handoff and the DFT continuation started from it. This is the config where cost and correctness can both be measured, because these 40 structures all have a from-scratch reference. It is also where every predictor and the PQI were calibrated.

Column groups (click to expand)

Identity - sample_index, folder_name (= structure_id), reference_folder_name, mp_id, formula_label, n_atoms, crystal_system, chemistry_arity, dft_reference_volume_ang3, dft_ionic_steps (the from-scratch cost this continuation is compared against), model_short, display_name, model_name.

Handoff geometry vs. the DFT reference - coord_mae_handoff_ang, coord_rmse_handoff_ang (the primary geometry metric), mean_site_disp_ang, max_site_disp_ang, lattice_mae_pct, angle_mae_deg, volume_error_pct, structure_match_scaled, structure_match_strict, final_volume_ang3, a, b, c, alpha, beta, gamma, and the Matbench-Discovery-style threshold scores adwt, afbt, ewt_avg, efwt_avg, trs. trs is present in the data but was excluded from every manuscript figure and table.

Handoff vs. the unrelaxed input (the do-nothing baseline) - initial_mean_site_disp_ang, initial_max_site_disp_ang, initial_lattice_mae_pct, initial_volume_error_pct, initial_angle_mae_deg, initial_structure_match_scaled, initial_structure_match_strict, and the three better_than_initial_* flags that make "the model degraded the input" measurable.

What the ML relaxation cost - relax_success, relax_steps, error_message, wall_time_sec, model_init_time_sec (kept separate because it amortises over a batch), optimizer_runtime_sec, postprocess_runtime_sec, time_per_atom_sec, atoms_per_sec, optimizer_time_per_step_sec, steps_per_sec, final_max_force_ev_ang (as the potential itself reports it), final_ml_energy_ev, final_ml_energy_per_atom_ev, run_started_at_utc, run_finished_at_utc.

Potential vs. DFT at the same geometry - both evaluated at the handoff positions, so the difference is purely the potential's error: dft_handoff_energy_ev, dft_handoff_energy_per_atom_ev, dft_handoff_forces_ev_ang (the full force array, JSON text), fmax_dft_handoff_ev_ang, ml_energy_handoff_ev, ml_energy_handoff_per_atom_ev, ml_forces_handoff_ev_ang (JSON text), ml_fmax_handoff_ev_ang, ml_stress_available, fmae_ev_ang, frmse_ev_ang, fcos, delta_e_ml_dft_mev_atom, delta_e_handoff_mev_atom. The force-discrepancy metrics are the best predictors of continuation cost.

What the DFT continuation did - run_status, parse_status, parse_error, vasp_version (6.5.0 throughout), run_date, n_atoms_cont, the settings encut / isif / ismear / nsw / nelm, converged_electronic, converged_ionic, ionic_steps and total_scf_iterations (the two cost units; they do not always move together), final_energy_ev, final_energy_per_atom_ev, final_volume_ang3_cont, and the five has_* output file flags.

Where the continuation ended up vs. the reference - structure_match_scaled_final, structure_match_strict_final, coord_mae_final_ang, coord_rmse_final_ang, lattice_mae_final_pct, angle_mae_final_deg, volume_error_final_pct, final_energy_error_per_atom_ev, reference_final_energy_ev, reference_final_energy_per_atom_ev, reference_ionic_steps, handoff_endpoint_max_site_shift_ang, handoff_endpoint_vol_shift_pct.

Outcome labels - delta_n_ionic_steps (from-scratch minus continuation, so a positive value means the handoff saved work and beneficial_steps_binary == (delta_n_ionic_steps > 0)), eta_ionic_steps (the same as a fraction of the from-scratch cost, delta_n_ionic_steps / reference_ionic_steps; median 0.400), beneficial_steps_binary, convergence_success_binary (constant 1 across all 240 rows - nothing failed in domain, which is precisely why the transfer set was needed), final_match_scaled_binary.

Readiness index, DFT-free - initial_handoff_mean_disp_ang, initial_handoff_max_disp_ang, initial_handoff_lattice_change_pct, initial_handoff_volume_change_pct, initial_handoff_angle_change_deg (how far the potential moved the structure from its own input, which needs no DFT), hri_local_readiness, hri_model_prior, hri_score, plus the consensus variant intermodel_mean_disp_ang, intermodel_lattice_change_pct, intermodel_volume_change_pct, hri_consensus_score, hri_consensus_gate, and the outcome label safe_to_skip_dft_label.

Bookkeeping - patch_pending, patch_pending_reason, outcome_geometry_stale.

transfer_scores - 648 rows (108 transfer cases x 6 potentials), 26 columns

The frozen recipe applied out of domain with zero refitting - same features, same weights, same thresholds as the calibration fit.

Column	Meaning
case_name	Case identifier; equals structures.structure_id
domain	bulk (270 rows), 2d (150), hetero (132), defect (96)
array_task_id	Joins to structures.array_task_id
model	chgnet, dpa, m3gnet, mace, mattersim, nequip
n_atoms	Atoms per cell
coord_mae_ang	Mean atomic-coordinate error of the handoff vs. the from-scratch DFT reference
lattice_mae_pct, volume_error_pct, angle_mae_deg	Cell-parameter, volume and angle error vs. the reference
scaled_match, strict_match	StructureMatcher verdicts vs. the reference
relax_steps, relax_success, wall_time_sec, final_max_force_ev_ang	What the ML relaxation cost and returned
initial_handoff_mean_disp_ang, initial_handoff_max_disp_ang, initial_handoff_lattice_change_pct, initial_handoff_volume_change_pct, initial_handoff_angle_change_deg	How far the potential moved the structure away from its own input. Displacements, not errors - they need no DFT, which is what makes the readiness score usable before any DFT is run
hri_local_readiness	The part of the score computed from this structure's own handoff features (seven weighted percentile features)
hri_model_prior	The part contributed by which potential produced it, calibrated in domain and then held fixed
hri_score	Pre-optimization Quality Index (PQI), 0-100, 0.5 * hri_local_readiness + 0.5 * hri_model_prior. The acceptance gate is hri_score >= 70. The index was named the Handoff Readiness Index while the analysis was frozen, which is why the shipped columns, file names and script names keep the hri_ prefix; renaming them would break the no-refitting provenance. PQI is the same quantity.
safe_geo	The outcome label the index is scored against, not the acceptance decision. 1 = the handoff geometry was in fact inside all four tolerances against the reference (248 rows); 0 = it was not (400 rows). Prevalence by domain: bulk 96/270, 2D 107/150, hetero 5/132, defect 40/96
training_overlap_risk_this_model	Overlap class between this case and this row's own potential's training corpus
training_overlap_max_risk_any_corpus	Highest overlap class over all audited corpora

safe_geo is not "accepted". 24 of the 107 gate-accepted rows have safe_geo == 0, and only 83 of the 248 safe_geo == 1 rows were ever run. Read the training-overlap columns before drawing any conclusion from the transfer set; the full audit and its caveats are data/extension/training_overlap_audit_summary.md on GitHub.

validation - 107 rows, 34 columns

The 107 handoffs the gate accepted, each actually run to its DFT endpoint. Every verdict is re-derived from the VASP output files by the standalone parser (scripts/verification/parse_vasp.py), deliberately without trusting the scheduler's own status file.

Column	Meaning
task_id, arm, model, domain, case_name	Identity; case_name joins to structures.structure_id
status_file	The verdict the run's own status file claimed
tier1_files_ok, tier1_missing	Tier 1: are all expected output files present, and which are missing
tier2_normal_term, tier2_fatal	Tier 2: did VASP terminate normally, and any fatal error signature
tier3_ionic_conv, tier3_elec_conv	Tier 3: ionic and electronic convergence, read from the output rather than assumed. The literal string NA means the check could not be evaluated - it is preserved, not turned into a null
tier4_energy_finite, tier4_forces_parsed	Tier 4: is the final energy finite, are forces parseable
parser_exit, parser_verdict	0 / converged (100 runs), 1 / completed_unconverged (2), 2 / failed (5)
verdict_agrees_status	Does the independent verdict agree with status_file. It agrees for all 107
status_ionic_steps, parser_ionic_steps	Ionic-step count as claimed vs. as re-parsed
status_E0, parser_E0, E0_agrees	Final energy as claimed vs. as re-parsed, and whether they agree (tolerance 1e-2 eV)
note	Free-text annotation (empty for all 107)
continuation_status, ref_status, n_atoms	Basin-check context
coord_mae_final_ang, coord_max_final_ang, lattice_mae_final_pct, volume_error_final_pct	Endpoint geometry error vs. the from-scratch reference
scaled_match, strict_match	StructureMatcher(ltol=0.2, stol=0.3, angle_tol=5) fit of the continuation endpoint against the reference endpoint
basin_verdict	correct (99), wrong (1), null for the 7 runs that did not produce a comparable endpoint
false_acceptance	True for the 8 accepted handoffs that did not deliver a converged run in the correct basin

recovery - 44 rows (22 failed cases x 2 potentials), 49 columns

22 structures whose from-scratch DFT relaxation failed were restarted from a potential's pre-relaxed geometry, with MatterSim and NequIP. Columns 1-36 are the same four-tier independent verification as validation, plus tier_verdict, reconciliation, ionic_agrees, the split between parser_ionic_completed and parser_ionic_started, the two energy definitions (parser_E0_last_ionic_step vs parser_E0_last_electronic_step, reconciled by E0_defn_split), the k-point and band counts as claimed vs. re-parsed, and a free-text note that spells out every timeout.

Columns 37-49 are case-level context attached from the case table, identical for the two rows sharing a case: n_atoms, isif, first_pass_error_hint_statusfile, first_pass_bucket_ondisk, first_pass_signature_ondisk, first_pass_hint_disagrees, first_pass_ionic_steps, first_pass_wall_sec, rescued (yes for 15 of the 22 cases), rescued_by, both_converged, any_timeout, itt_bestcase_rescued.

The files on GitHub are prefixed rescue_arm_; the manuscript calls this the recovery arm. The two names denote the same 44 runs, and the arm column here says rescue.

recovery_control - 13 rows, 64 columns

The same failures restarted the conventional way - by changing DFT settings only, no potential involved. This is the comparison baseline for recovery. Columns 1-59 mirror the control-arm verification ledger: identity (restart_task_id, parent_task_id, case_name, domain, n_atoms), the failure it was restarted from (error_class, profile, poscar_source), the four-tier verification, the INCAR knobs that were changed (incar_ediff, incar_ediffg, incar_potim, incar_isif, incar_isym, incar_ldipol, incar_kspacing, nsw, nelm), the force trace (Fmax_first_eV_per_A, Fmax_last_eV_per_A, Fmax_min_eV_per_A, Fmax_last_over_target), and the timeout adjudication block.

Columns 60-64 are the head-to-head verdict against the MLIP restart of the same case: mlip_recovered, mlip_recovered_by, control_recovered, pair_category (both 4, MLIP-only 4, control-only 2, neither 3) and union_recovered.

Note that recovery_control.domain spells the 2D domain 2d_monolayer while recovery spells it 2d. Both are preserved as written.

leaderboard - 6 rows, 13 columns

Column	Meaning
potential	Model label as the shipped tables carry it
source	published (data/) for all six rows - these are read off the measurement tables, not recomputed
notes	Footnote letters, expanded in leaderboard.md
in_domain_n, extension_n	40 and 108: the row sets each cell is measured over
in_domain_hri_accept_pct_70, extension_hri_accept_pct_70	Percent of handoffs at or above the deployed operating point, PQI >= 70. A decision rate, not an accuracy
in_domain_hri_mean, extension_hri_mean	Mean composite score, 0-100
in_domain_safe_pct, extension_safe_pct	Percent of handoffs whose geometry is inside all four tolerances against the DFT reference
in_domain_median_relax_steps, extension_median_relax_steps	Median optimizer steps to reach fmax 0.03 eV/A or the 500-step cap

leaderboard.csv and leaderboard.md at the repository root are byte-identical copies of what scripts/benchmark/make_leaderboard.py emits in the GitHub repo; the paths they mention (data/ledgers/..., scripts/benchmark/...) are GitHub paths.

---

Benchmark your own model

The 200 inputs and 159 references in the structures config are everything you need to measure any ASE calculator on the same suite, the same frozen protocol and the same acceptance score as the six published potentials. The whole contract is one function:

# my_potential.py
def get_calculator():
    return SomeASECalculator(...)

Then, in a clone of the GitHub repo:

git clone https://github.com/ShiQiaoL/umlip-dft-preopt
cd umlip-dft-preopt

python scripts/benchmark/run_relaxation.py --calculator my_potential.py --name MyPotential
python scripts/benchmark/compute_metrics.py --results results/MyPotential
python scripts/benchmark/score.py           --results results/MyPotential
python scripts/benchmark/make_leaderboard.py

Your row appears beside the six published ones. scripts/benchmark/examples/README.md has a ready-to-run mock calculator, a real MACE template and expected wall times; scripts/benchmark/test_regression.py replays a published potential's 40 in-domain endpoints through the harness and asserts every column against the shipped rows.

Two things to keep in mind. The readiness index needs no DFT reference, so all 200 structures produce a score, but every reference-based number is computed only where a reference exists - and the leaderboard states its row set per row, because a comparison across different row sets is not a comparison. And the protocol is deliberately not configurable: change fmax or the step cap and you are no longer on this benchmark.

---

Provenance, licensing and attribution

License: CC BY 4.0 for this dataset repository as a whole.

The composition is not uniform, and the reason for the choice matters:

• 90 of the 200 structures descend from Materials Project entries (all 40 calibration structures, 40 transfer bulk, and 10 of the 20 defect_or_doped supercells). Materials Project data is CC BY 4.0, which requires attribution. mp_parent_id gives the parent entry for every one of them, and mp_parent_derivation says how the input was made from it. The geometries here are perturbed derivatives, not verbatim MP entries; the unperturbed parent can be re-downloaded from materialsproject.org with the mp_id.

  If you use these structures, cite the Materials Project:

  A. Jain, S.P. Ong, G. Hautier, W. Chen, W.D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, K.A. Persson, Commentary: The Materials Project: A materials genome approach to accelerating materials innovation, APL Materials 1, 011002 (2013). doi:10.1063/1.4812323

• 60 structures descend from NIST JARVIS-DFT (JVASP-* ids: 10 bulk from JARVIS-DFT-3D, 40 2d_monolayer from JARVIS-DFT-2D, and 10 vacancy supercells cut from JARVIS-2D parents). source_id gives the parent entry for every one of them, and the unperturbed parent can be re-downloaded from jarvis.nist.gov. If you use those rows, cite:

  K. Choudhary, K.F. Garrity, A.C.E. Reid, B. DeCost, A.J. Biacchi, A.R. Hight Walker, Z. Trautt, J. Hattrick-Simpers, A.G. Kusne, A. Centrone, A. Davydov, J. Jiang, R. Pachter, G. Cheon, E. Reed, A. Agrawal, X. Qian, V. Sharma, H. Zhuang, S.V. Kalinin, B.G. Sumpter, G. Pilania, P. Acar, S. Mandal, K. Haule, D. Vanderbilt, K. Rabe, F. Tavazza, The joint automated repository for various integrated simulations (JARVIS) for data-driven materials design, npj Computational Materials 6, 173 (2020). doi:10.1038/s41524-020-00440-1

• The 50 heterostructures were generated for this work and carry the authors' licence.

• All measurement tables, scores, labels and the leaderboard are the authors' own work. The companion GitHub repository releases its code under the MIT licence; the authors additionally offer these tables under CC BY 4.0 as part of this collection. CC BY 4.0 is the label on the repository because it is the most restrictive term that any component carries, and it is satisfiable by attribution alone.

None of the six model checkpoints is redistributed here or on GitHub; each remains under its own upstream licence.

Not in this repository: the raw VASP run trees (INCAR / KPOINTS / POTCAR, OUTCAR, vasprun.xml, CONTCAR) and the per-structure handoff geometries. Those are tens of gigabytes and are archived on Zenodo, with the DOI assigned at acceptance.

---

Citation

The article:

@article{liu2026umlip,
  title   = {Benchmarking universal machine learning interatomic potentials as
             {DFT} pre-optimizers},
  author  = {Liu, Xiaoqing and Wang, Yangshuai and Yang, Ziduo and Shen, Lei},
  journal = {npj Computational Materials},
  year    = {2026},
  note    = {Submitted 2026. Volume, pages and DOI are placeholders until
             acceptance; please check for the published version before citing.}
}

The code and curated data:

@software{liu2026umlip_software,
  title    = {umlip-dft-preopt: code and curated data for benchmarking universal
              machine learning interatomic potentials as {DFT} pre-optimizers},
  author   = {Liu, Xiaoqing and Wang, Yangshuai and Yang, Ziduo and Shen, Lei},
  year     = {2026},
  version  = {1.0.0},
  license  = {MIT},
  url      = {https://github.com/ShiQiaoL/umlip-dft-preopt}
}

The Zenodo DOI for the raw-run archive is registered at acceptance; until then there is no identifier to cite, so the entry carries the repository URL only. Check the GitHub repository for the live DOI before adding one.

Author affiliations: Xiaoqing Liu and Lei Shen, Department of Mechanical Engineering, National University of Singapore; Yangshuai Wang, Department of Mathematics, National University of Singapore; Ziduo Yang, College of Information Science and Technology, Jinan University, Guangzhou.

---

How this repository was built

Every value here is read from a file in the companion GitHub repository. Nothing was recomputed, re-fitted or re-derived, and no number was rounded, reformatted or re-typed on the way in.

• Serialization choice. Structures ship as the verbatim POSCAR text of the source .vasp file plus a parsed {lattice, species, frac_coords} struct. CIF was rejected: the sources are POSCAR files, so a CIF column would insert a lossy conversion between the reader and the benchmark. The text column is byte-identical to the source file, and its SHA-256 is carried alongside so a reader can prove that independently.
• Round-trip proof. All 359 shipped geometries (200 inputs + 159 references) were read back from parquet, parsed with pymatgen, and compared against pymatgen's parse of the source .vasp: 0 byte differences, 0 SHA-256 mismatches, 0 species or atom-count mismatches, maximum lattice deviation 0.0, maximum fractional coordinate deviation 0.0 (Cartesian coordinates agree to 7.1e-15 A, which is the floating-point cost of the frac @ lattice product, not a data difference).
• Cell-exact fidelity. Each config was built twice - once with numeric typing and once with every cell kept as raw CSV text - and the parquet round-trip was diffed against the raw text cell by cell: 63,432 cells checked, 0 mismatches across all seven configs. Two source-faithfulness traps were found and fixed by that check, and both are worth knowing about if you re-read the source CSVs yourself: pandas' default CSV float parser is not correctly rounded and perturbs the last bits of 587 cells across 35 columns and 7 of the source tables (float_precision="round_trip" fixes it), and pandas' default NA list silently converts the 20 literal NA verdict cells in the validation-arm ledger - which mean "this check could not be evaluated", not "missing" - into nulls (keep_default_na=False, na_values=[""] fixes it).
• Numbers untouched, names mostly untouched. Renamed columns are listed in full under structures above and nowhere else.
• Dropped, deliberately. Seven columns of the in-domain master table held filesystem paths into the authors' run tree and the compute system; they point at an archive that is not distributed here and they are not part of the measurement, so they are not shipped. One duplicated join column (reference_folder_name.1, byte-identical to reference_folder_name) was also dropped.

Total repository size: 1.3 MB.