acceptodds
Under review as a conference paper at ICLR 2027

Beyond Energies and Forces: OMol-Descriptors-4M, an Open Post-DFT Pipeline and Evaluation Protocol for Machine Learning in Chemistry

Abstract

We present , an end-to-end computational chemistry post-analysis framework that processes repositories of molecular DFT wavefunctions and yields six partial-charge schemes (Hirshfeld, CM5, ADCH, Becke, Mulliken, Loewdin), four bond-order schemes, full QTAIM topology, fuzzy descriptors, and ORCA-derived globals. We built this by wrapping per-tier analysis routines around Multiwfn and ORCA's utility, packaging the result into a scalable workflow for HPC clusters. Users can select compute tiers to trade fidelity for cost, then compact the resulting calculations into portable sharded Lightning Memory-Mapped Databases (LMDBs) keyed by job identity for downstream analysis and machine learning. To demonstrate the pipeline, we processed the recently-released OMol25 electronic structure dataset: a 4M-structure public subset of the 110M+-calculation 2025 release that ships the complete DFT outputs, including density matrices, energy gradients, ORCA files, and wavefunctions. By post-processing the wavefunctions, we extract atom-, bond-, and molecule-level descriptors across 4M structures spanning 34 chemical verticals, supplementing the original dataset's energy and force labels with physically-interpretable quantities. We name this dataset OMol-Descriptors-4M and present it as the first multi-vertical descriptor dataset at this scale beyond energies and forces. Analogous to the source dataset, we establish an evaluation protocol with explicit evaluative claims and scope boundaries: composition-ordered train/val/test splits via blake2b-hashed molecular formula and five domain-specific held-out stress tests (metal-ligand bond pairs, lanthanide-ligand bond pairs, reactivity, large systems, and high-charge regimes). Using this dataset, we establish a baseline set of models for future development and stress test them using these evaluation sets.

Primary area:
datasets and benchmarks
Keywords:
machine learning for chemistry, quantum chemistry, density functional theory, graph neural netowrks, benchmark dataset, molecular descriptors, partial charges

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NjErOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

In this paper, the authors (1) introduce a utility tool to extract atomic descriptor information from ORCA calculations into LMDBs, (2) demonstrate its capabilities by processing specific properties from the wavefunctions of the Omol 4M dataset, which was released as 'omol electronic structures', and (3) create evaluation splits from it.

Critical strengths

This is a workflow paper, and it presents a reasonably difficult engineering effort.

Critical weaknesses

The abstract advertises the wavefunctions, but it seems like the wavefunctions are not actually included in this database. I would increase my score if the authors include full electronic data into this database, eg the Density matrices, Hamiltonian matrices, and electron densities/wavefunction information. That would directly reflect the true utility of the released omol electronic structures data and would save the materials modelling community significant time and effort to do this themselves. As it is, I think it's a useful utility to have, but it seems like the kind of thing individual researchers can set up with agent workflows on a compute cluster. I've done something similar when dealing with this data, and it worked just fine (just needed some minor filtering/cleanup at the end). Eg, I'm not currently convinced this is a level of contribution that can justify an ICLR submission. However, I'm open to being shown wrong, or that there is something I overlooked, and am happy to engage with the authors during the rebuttal stage.

Questions for authors

See 'weaknesses'

1 comment

Thank you for engaging and for stating what would change your assessment. We address the wavefunction point first:

"The wavefunctions are advertised but not included."

Our wording was misleading, and we have corrected the abstract and introduction. The raw files already exist publicly:

  • The 4M raw release. Wavefunction (gbw) files, density matrices and ORCA outputs are public under CC-BY-4.0. They total about 500 TB behind a gated transfer endpoint.
  • Fock matrices. These are public only for a 58k closed-shell subset.
  • Every record joins to its raw job. Each record in our release carries its OMol25 job path, so the raw files for any record are one lookup away.

Re-hosting about 500 TB would duplicate public data and add no information, so we instead:

  1. State the relationship precisely. The revised introduction and storage section do this, and a new appendix measures raw file sizes. A median job is 20 MB of compressed raw files; our descriptor layer is 48 KB per structure (189 GB compressed in total for 3.93M structures).
  2. Point to the raw files directly. The raw files for every structure in our release, including the held-out suites, are already on the public OMol25 endpoint at ALCF (Eagle, via Globus). Each record in our LMDBs stores its OMol25 job path, so the wavefunction, density matrix and ORCA output for any record can be fetched directly, with no re-hosting.
  3. Package the descriptors for ML use. The descriptors are stored as typed LMDBs (charges, bonds, QTAIM, fuzzy-space descriptors, ORCA globals, energies and forces), keyed by structure and aligned across splits and held-out suites (Section 3.2):
    • Size. The whole release is about 2,600x smaller than the raw files.
    • Portability. It downloads to a single workstation.
    • Access. Each record is one random-access read, with no ORCA, orca_2mkl or Multiwfn installation.
    • Integration. The LMDBs convert directly to graph datasets.

"This could be done with agent workflows on a cluster."

The pipeline itself is not the hard part; producing descriptors that are correct is. Several failure modes exit cleanly and parse without error, so an unsupervised workflow would ship wrong values silently. Examples we found and fixed:

  • Open-shell wavefunction files lose alpha/beta separation. Spin densities, fuzzy bond orders and ELF come out wrong, while charges are unaffected.
  • 557,098 all-alpha UKS singlets got fuzzy bond orders stored at 0.5x or 2x.
  • 38,409 jobs have swapped nuclear critical-point labels on H–H pairs.
  • A sign bug affects h-shell beta orbitals in open-shell relativistic (X2C) actinide files.

The pipeline is also engineered as a reusable tool, not a one-off script:

  • Validated. The cross-package validation (Appendix O: Multiwfn vs HORTON and Critic2 on 87 structures) is how we caught problems like these. It is the kind of check an ad hoc workflow skips.
  • Tested. A test suite covers parsing, LMDB key derivation, sharded conversion and merging end to end, so a fresh checkout reproduces the artifacts we ship (Appendix R).
  • Benchmarked. Throughput and parallel scaling are measured across worker/thread configurations (Appendix T).
  • Portable. The same code ran under PBS Pro, Slurm and Flux at three computing centers, and on a single workstation. Moving to a new center means swapping a scheduler provider and its account fields (Appendix R).

"Is this level of contribution enough for ICLR?"

The new experiments aim to push this. A descriptor layer that most ML groups could not compute themselves (500 TB of input, per-structure Multiwfn runs, the integrity issues above) turns out to carry information that energy/force training does not recover on its own:

  • Feeding model-predicted Hirshfeld charges to a UMA MLIP lowers energy error 21–23% on OMol25 validation.
  • It lowers energy and force error on every held-out stress suite.
  • The largest gain is on high-charge systems absent from training (−35% energy, −28% forces at 100k).

The general response has the table, and Appendix N has the full experiment. The descriptors also come in a form an ML group can use directly: compact, portable LMDBs instead of hundreds of terabytes of raw files. We think this moves the release from infrastructure to a resource with measured downstream value. We would welcome your view on it.

Reviewer_dKyyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Modern computational chemistry datasets often contain limited amount of data: primarily only energy and forces to facilitate training of MLIPs. However there are multiple relevant features that could be extracted from converged density and wavefunctions along with energy and forces and require zero additional DFT calculations. Authors highlight this fact and additionally extract partial charges, bond orders, electron density critical points, and other relevant features from wavefunctions contained in OMol25 dataset. They release extracted data together with processing pipeline.

Critical strengths
  • The work is methodologically sound and concise. I indeed agree with authors' idea that extracting more data available from converged SCF is useful for the field where one could spend multiple cpu hours to obtain a single point calculation results. And utilizing this physically relevant information could improve performance on downstream tasks.
  • The authors carefully approached the dataset splitting strategy and performed held out set tests while honestly estimating the overlaps between held out and train.
Critical weaknesses

Insufficient demonstration of the dataset’s usefulness

  • Neural-network experiments cover only six partial-charge targets. There are no results that demonstrate effective prediction of other properties contained in the dataset. Consequently, it is unclear if current DL models can predict properties with enough accuracy to be employed for subsequent chemical analysis.
  • Moreover, there are no experiments that establish whether predicted descriptors could improve downstream chemical-property prediction. This is particularly interesting, as cited related work on QTAIM-enriched GNNs (Gee et al. (2025), Vargas et al. (2024)) demonstrates that adding QTAIM features only improves the performance of GNNs in some tasks. Since even quantum-chemically computed QTAIM features provide task-dependent benefits, their usefulness cannot automatically be extended to predicted features. The downstream evaluation is therefore necessary to determine if predicted descriptors retain useful information.
  • The described T2 is a valid task for benchmarking but it requires a BCP graph which in term requires a DFT calculation. This makes models trained on this task impractical as they cannot be applied on unseen molecules. One option is to determine BCP graph using model from T3, but this introduces another possible source of error and needs to be evaluated properly.

Other weaknesses

  • The manuscript is missing model configurations and training details needed to reproduce the experiments.
  • Averaging across partial charge types hides substantial differences in prediction error. For the equivariant model, the test MAE for Mulliken charges is 0.0801e—approximately eleven times the Hirshfeld MAE—despite a relatively high Mulliken R2R^2 of 0.944. This highlights the need to distinguish absolute error magnitude from relative predictive quality. Averaging also masks differences in generalization: on large systems, the overall MAE improves while the Mulliken MAE increases by approximately 51%.
Questions for authors
  • What explains the differences in error across partial charge types? I am especially interested in larger Mulliken MAE? Could this be due to Mulliken charges using different scale? Additionally, why does the equivariant model achieve substantially lower test R2R^2 for ADCH (0.789) and corrected Becke (0.727) than for Mulliken (0.944) and the other schemes?

Suggestions

  • Does joint training on QTAIM descriptors and energies/forces improve energy/force accuracy or data efficiency compared with energy/force training alone?
  • Subjectively I would propose the authors to change the task and held-out set names from indexes T1-3 and H1-8 to more human-readable names so it would easier to track this through the text without looking for the definition constantly.
2 comments

Thank you. Your central question, whether predicted descriptors help downstream, is the right one, and we ran it as a controlled experiment. The general response has the tables; key points:

W1. Do predicted descriptors improve downstream prediction?

Yes, here. We take your point from Gee et al. (2025) and Vargas et al. (2024) that QTAIM-feature benefits are task-dependent, so we report this one task (energy and forces) as is.

  • Setup. A UMA MLIP given only one predicted descriptor, the per-atom Hirshfeld charge from our Charges model. Everything else is identical to the baseline.

  • Result on OMol25 val (structures outside our dataset).

    • Energy MAE 21% lower at 10k and 23% lower at 100k.
    • Force MAE 14% lower at 10k and 7% lower at 100k.
    • Lower on all five held-out suites, at both sizes, for both seeds.
  • High-Charge is the most informative case.

    • The charge model is least accurate there (R² 0.89), yet the MLIP gains most: −35% energy and −28% forces at 100k.
    • Without charges, the MLIP barely improves on these systems from 10k to 100k.
    • So the predicted descriptor transfers information that energy/force supervision alone does not supply, even when the descriptor is imperfect.
  • Predicted vs DFT charges. Predicted charges recover 66–96% of the energy gain that DFT charges give, but only 31–43% of the force gain. Better charge models would add more, which is what Charges measures.

Suggestion. Does joint training on descriptors and energies/forces help?

We ran this too. UMA trained on energies and forces plus an auxiliary per-atom charge head (weight 10, same subsets and recipe; OMol25 val MAE, mean of 2 seeds):

Arm10k forces (eV/Å)10k energy (eV)100k forces (eV/Å)100k energy (eV)
E + F only0.05351.5550.02560.681
+ our Hirshfeld (aux target)0.0499 (−7%)1.409 (−9%)0.0248 (−3%)0.666 (−2%)
+ OMol25 Löwdin (aux target)0.0511 (−4%)1.462 (−6%)0.0253 (−1%)0.683 (+0%)
+ predicted Hirshfeld (input)0.0460 (−14%)1.222 (−21%)0.0238 (−7%)0.525 (−23%)

(Appendix N, Table 16. Values are from training logs, so arm P differs from the general response in the third decimal.)

  • Auxiliary supervision helps a little, and our Hirshfeld beats Löwdin. OMol25 ships Löwdin charges. Most of the auxiliary gain is gone by 100k.
  • Predicted charges as an input help several times more, and the energy gain persists at 100k.
  • Caveat. The auxiliary weight was chosen on 10k validation forces using one seed of the Hirshfeld arm.

W2. BCP-Exist / BCP-Props need a DFT BCP graph, which is impractical for new molecules.

Agreed. BCP-Props is an oracle-topology benchmark by design. In practice it would chain with BCP-Exist (predicting the BCP graph from geometry), and that compound error needs its own evaluation. We have not done that and say so in the paper. The charge experiment above uses only geometry-derived inputs, so it applies to unseen molecules as is.

W3. Missing configurations.

Added as Appendix M: shared inputs and edges, a per-model table (architecture, encoder, cutoff, batch, learning rate, weight decay, parameters), and the training setup (optimizer, schedule, early stopping). Configuration files are released with the code. Writing this up, we found and fixed an input problem (DFT dipole magnitudes used as inputs; see R1 Q1). Appendix N gives the full UMA recipe for the downstream experiment.

W4 / Q1. Per-scheme errors (Mulliken MAE about 11x Hirshfeld at R² 0.944; ADCH and corrected Becke R²).

Thank you; this was a good catch. We computed label statistics on the full test split (20.9M atoms) and added them as a new appendix table.

  • Mulliken is on a different scale.
    • Its labels spread 5x more than Hirshfeld's (std 0.61 vs 0.12 e), and 8.6% of Mulliken charges exceed 1 e in magnitude, vs 0.03% for Hirshfeld.
    • Normalized by label spread, Mulliken's MAE is only about 2x Hirshfeld's.
    • So a large MAE and a high R² are consistent.
  • ADCH and corrected Becke are hard on every measure.
    • They have the heaviest tails (excess kurtosis 1,066 and 132), the highest MAE/std, and tail-dominated errors (RMSE/MAE 2.8).
    • Both carry the atomic-dipole correction. We note that as a shared property without claiming it causes the difficulty.
  • Per-scheme results are reported separately from now on. The six-scheme mean appears only alongside them.
  • The Mulliken increase on Large-Systems is plausibly basis-set delocalization in large systems. We flag it as a hypothesis.

Rename suggestion. Done throughout:

  • T1–T3 are now Charges, BCP-Props, BCP-Exist.
  • H1–H8 are now Metal-Ligand, Reactivity, Lanthanide-Ligand, Large-Systems, High-Charge.
  • The release directories keep their short IDs, with the mapping in the appendix and data card.

Thank you to the authors for the detailed response and additional experiments. My main concern regarding downstream applicability has been at least partially addressed. I have decided to raise my score to 3.

However, I believe that further evaluation of other descriptors, particularly bond orders and QTAIM features, would be more informative and would strengthen the paper. Several existing machine-learning force fields already incorporate atomic partial charges into energy prediction, so the downstream utility of partial charges was arguably the least uncertain among the descriptor types introduced in this work.

I would therefore encourage the authors to investigate the downstream utility of these other descriptors in future work.

Reviewer_8ctROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces OMol-Descriptors-4M, a post-DFT resource derived from approximately four million OMol25 structures. It combines multiple charge and bond-order definitions, QTAIM descriptors, an extraction pipeline, composition-based splits, and five stress-test suites. Learned results are reported only for charge prediction, using five baselines.

Critical strengths
  1. Broad, aligned labels. Joint descriptor, energy, and force labels across 34 chemical domains could support interpretable and multitask learning.
  2. Useful infrastructure. The extraction pipeline, integrity checks, and targeted stress suites increase the resource’s practical value.
Critical weaknesses
  1. Incomplete benchmark validation. Baselines use a split differing from the release by 48,791 records (1.2%), while input, graph-construction, and training details remain incomplete. T2/T3 have no learned results. This limits reproducibility and leaves the practical value of the distinctive descriptors insufficiently demonstrated.
  2. Inconsistent quality claims. Section 3.3 claims BCP correlations of at least 0.998, but Table 13 reports 0.4029 for ellipticity and 0.9471 for the Hessian determinant. Table 5’s 30.2 million charge records also exceed six schemes times 3.985 million structures. Both discrepancies require clarification to support the stated data quality.
  3. Insufficient validation of the dataset’s distinctive value. QTAIM topology and bond-critical-point labels are key contributions, yet learned results are limited to charge prediction. Their practical value, alongside the benefits of scale and cross-domain coverage, remains unclear. Representative T2/T3 results and a controlled scaling or downstream experiment would substantially strengthen the paper.
Questions for authors
  1. Can the authors provide baselines on the exact released split, with training data, complete configurations, and explicit input and edge definitions?
  2. How do the authors reconcile §3.3 with Table 13 and explain Table 5’s charge-record total?
  3. Are completed T2/T3 or controlled downstream results available to establish the utility of the distinctive descriptors?
1 comment

Thank you for the careful reading. Two of your points exposed real errors, which we have fixed.

Q1. Baselines on the exact released split, with configurations and input/edge definitions.

We re-built the Charges graphs on the released split and retrained. The released split has train 3,122,674, val 398,777 and test 403,680 graphs, plus the five held-out sets.

While doing this we found a flaw in the original inputs, and we want to state it plainly. The original graphs carried four DFT dipole magnitudes (from the Hirshfeld, ADCH, CM5 and Becke charges) as molecule-level inputs. Our converter appended them automatically. Net charge and spin were not inputs, although the paper said they were.

  • Size of the effect. Replacing the dipoles with their training mean at test time raised Hirshfeld test MAE by 1.5–1.9x, and the six-scheme mean by 1.08–1.15x, across four models. The effect is material.
  • What we changed. The released graphs now drop the dipoles and add net charge (one-hot, −10…10) and spin multiplicity (one-hot, 1…11).
  • Retrained results. We retrained on them. The retrained SchNet model matches or beats the original:
    • Hirshfeld test MAE 0.0070 e, R² 0.978.
    • Six-scheme mean MAE 0.046 e and mean R² 0.856 on the released test split, vs 0.047 and 0.854 in the submitted table (earlier split, with dipole inputs).
    • So the original headline was not an artifact of the leak, though it did rely on it.

SchNet and GAT are retrained on the released graphs; the remaining models will be updated in the camera-ready version. The correction and the retrained numbers are in the new Appendix M.

The new Appendix M (Table 13) gives, for every model:

  • Inputs.
    • Atoms: element one-hot, degree, H count, ring flags.
    • Bonds: Gaussian distance RBF (50), bond-orientation harmonics, ring flags.
    • Molecule: atom and bond counts, molecular weight, charge and spin one-hot.
  • Edge definition. Structural (coordinate-based) bonds. Nothing is derived from DFT.
  • Architecture, optimizer, schedule, early stopping, and parameter counts.

The training configuration file for every model is released with the qtaim_embed code (configs/omol_descriptors_4m/charges/).

Q2a. Section 3.3 ("r ≥ 0.998") vs Table 13 (ellipticity 0.4029, Hessian determinant 0.9471).

You are right; the sentence overstated the result. Table 13 explains the gap:

  • All-electron structures (62). Every BCP field agrees with r = 1.0000.
  • ECP structures (25). The low correlations all come from these. Critic2 does not read the EDF core densities, so Hessian-derived fields diverge.
  • Ellipticity. It is λ1/λ2 − 1 and ill-conditioned for near-cylindrical bonds; its median absolute difference is still 2.8e-7.

Section 3.3 now says exactly this. Correcting it, we also found that 3 all-electron structures have mismatched BCP sets, so the text now says "31 of the 39 unmatched BCPs are in ECP structures".

Q2b. Table 5's 30.2M charge records exceed 6 schemes x 3.985M structures.

Correct; the census over-counted. It counted:

  • 7 records per job: the 6 schemes plus a duplicate of Mulliken, which the release no longer ships.
  • Three extra schemes (CHELPG, MBIS, VDD) present on a subset of jobs in seven verticals.

Restricted to the six schemes described in the paper, the total is 23,910,630, which is 6.00 per job. Table 5 is corrected, and the data card now lists the extra schemes.

Q3 / W3. Learned results for BCP-Props / BCP-Exist, and a controlled downstream or scaling experiment.

  • Downstream and scaling. See the general response: descriptor inputs to an MLIP at 10k and 100k training structures, scored on OMol25 val and all five held-out suites.
  • Auxiliary training target. We also tested using our Hirshfeld charges as an auxiliary target rather than an input (see our response to R2).
  • BCP-Props / BCP-Exist. We do not have learned results within the discussion window and will not claim otherwise. Both tasks are fully specified in the paper (targets, splits, metrics, baselines), and learned baselines for both will be in the camera-ready version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers. All three asked for the same thing: evidence that the descriptors are useful beyond the charge-prediction benchmark. R2 put it most sharply: do predicted descriptors, not DFT ones, help a downstream model? We ran that experiment. Its summary is below; per-reviewer details follow.

Setup.

  • Model and training. UMA-S (eSCN-MD, 8 experts, direct forces), trained from scratch on nested 10k and 100k subsets of our training split. Every arm uses the same recipe: 40 epochs, identical subsets and record order, 2 seeds.
  • Arm A is UMA with energies and forces only.
  • Arm P adds one per-atom input: the Hirshfeld charge predicted by our charge model (Charges task, SchNet encoder). The charges are shifted per atom so they sum to the system charge. The extra input is the only difference between the arms. The new embedding is created with an isolated RNG, so every other parameter initializes identically to arm A for a given seed.
  • Evaluation sets.
    • OMol25 validation: a fixed 5 x 20k sample. These records are outside our dataset, so the charge model never saw them.
    • A 20k sample of our test split.
    • All five held-out stress suites.
  • No leakage. No evaluation record is in either model's training data. The charge model's Hirshfeld error is the same on UMA's training records (0.0069 e) as on test (0.0067 e), so UMA is not trained on unrealistically accurate charges.

Result (MAE, mean of 2 seeds, A → P). On every set, at both sizes, both arm-P seeds beat both arm-A seeds, for both energy and forces.

Set10k energy (eV)10k forces (eV/Å)100k energy (eV)100k forces (eV/Å)
OMol25 val1.542 → 1.216 (−21%)0.0534 → 0.0459 (−14%)0.682 → 0.526 (−23%)0.0256 → 0.0238 (−7%)
Test (20k)0.729 → 0.573 (−21%)0.0643 → 0.0567 (−12%)0.229 → 0.192 (−16%)0.0321 → 0.0308 (−4%)
Metal-Ligand1.601 → 1.266 (−21%)0.0950 → 0.0847 (−11%)0.537 → 0.447 (−17%)0.0547 → 0.0528 (−3%)
Reactivity1.560 → 1.249 (−20%)0.1746 → 0.1526 (−13%)0.641 → 0.515 (−20%)0.1082 → 0.1031 (−5%)
Lanthanide-Ligand2.426 → 1.507 (−38%)0.1119 → 0.0967 (−14%)0.566 → 0.526 (−7%)0.0555 → 0.0544 (−2%)
Large-Systems5.136 → 4.381 (−15%)0.0365 → 0.0319 (−13%)3.207 → 2.689 (−16%)0.0182 → 0.0163 (−10%)
High-Charge30.68 → 24.06 (−22%)0.1099 → 0.1011 (−8%)26.16 → 16.97 (−35%)0.1031 → 0.0748 (−28%)

Findings.

  1. The energy gain does not wash out with more data: −21% at 10k and −23% at 100k on OMol25 val. The force gain narrows from 10k to 100k on most sets.
  2. The largest gain is on High-Charge (|Q| ≥ 5). Our training split contains no such structures. There, arm A improves little from 10k to 100k (forces 0.110 → 0.103 eV/Å), while arm P improves substantially (0.101 → 0.075). This holds even though High-Charge is where the charge model itself is least accurate (Hirshfeld R² 0.89, vs about 0.98 elsewhere).
  3. The predicted charges are cheap. Predicting 220k structures took about 15 minutes on one A100 but this model was not highly tuned for inference speed and could certainly be much faster here.

Ceiling: DFT charges as input (arm O, 10k, 2 seeds). Arm O is identical to arm P but is fed the DFT Hirshfeld charges instead of predicted ones. It can be scored only on our test split and the held-out suites, because OMol25 val has no DFT charges. We excluded 6 records with corrupted DFT Hirshfeld labels (charges summing to the wrong total).

SetEnergy A → P → O (eV)Share of O's energy gain P recoversForces A → P → O (eV/Å)Share of O's force gain P recovers
Test (20k)0.729 → 0.573 → 0.50971%0.0643 → 0.0567 → 0.045842%
Metal-Ligand1.601 → 1.266 → 1.09066%0.0950 → 0.0847 → 0.068238%
Reactivity1.560 → 1.249 → 1.15477%0.1746 → 0.1526 → 0.111835%
Lanthanide-Ligand2.426 → 1.507 → 1.38588%0.1119 → 0.0967 → 0.072539%
Large-Systems5.136 → 4.381 → 4.02868%0.0365 → 0.0319 → 0.025843%
High-Charge30.68 → 24.06 → 23.8096%0.1099 → 0.1011 → 0.081631%
  • Predicted charges recover most of the energy benefit of DFT charges (66–96%).
  • They recover only about a third of the force benefit (31–43%). Exact DFT charges cut force error by 26–36%.
  • This points to more value in better charge models, which is the purpose of the Charges benchmark and the training data it provides. We, again, only tested with one charge scheme as of now.

Two caveats:

  • These are two seeds per cell. The seed-to-seed spread is 2–3%, well below the A/P gaps.
  • A 1M-structure arm is still training and will not finish within the discussion window.

The revised PDF adds this experiment as Appendix N, with Figure 12. All revisions are in red.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.