SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials
Jonas Busk, Emil J. P. Frost, Yogeshwaran Krishnan, Henrik H. Kristoffersen, August E. G. Mikkelsen, Xueping Qin, Xin Yang, Heine A. Hansen, et al.
cond-mat.mtrl-sci · cs.LG · physics.chem-ph · physics.comp-ph
Quick verdict
SoLiD26: 15.4M DFT solid-liquid interface structures; MACE training cuts force errors ~3x vs a pretrained baseline. Valuable, but data remains request-only.
Summary
SoLiD26 is a curated dataset of 15,379,823 first-principles atomic structures for training and evaluating machine-learned interatomic potentials (MLIPs) at solid-liquid interfaces. Compiled from density functional theory calculations (VASP, PBE+D3) across prior DTU studies of aqueous metal interfaces and electrode-electrolyte systems, it spans 15 elements, up to 576 atoms per structure (126.3 average), and ~100 GB in .aselmdb format. Most configurations come from ab initio molecular dynamics; bulk liquid and solid references are included. The authors describe a filtering pipeline (energy-cutoff and force thresholds, duplicate removal, rare-element exclusion, and PaiNN cross-validation-based outlier detection) and release preparation code on GitHub. Technical validation trains MACE models on a size-based split (test = structures with >200 atoms; 100,000 validation; 15,006,934 training). On the held-out test set, fine-tuned and from-scratch MACE models reduce force MAE from 0.0501 eV/A (pretrained MACE-MATPES-PBE-0 with D3 added) to about 0.0118 eV/A and per-atom energy MAE from 0.0046 to 0.0027 eV. Main limitations: the dataset is only available on request while a public release is prepared; the >200-atom test split covers only 6 of 15 elements and same-study provenance, so generalization is untested; ML-based outlier filtering risks discarding genuinely reactive configurations; hyperparameters were not optimized and results are labeled preliminary.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SoLiD26 consolidates 15,379,823 VASP PBE+D3 structures (up to 576 atoms, 15 elements, ~100 GB) from ten prior DTU solid-liquid interface studies into a curated MLIP training resource.
- A documented filtering pipeline (cutoff/force thresholds, dedup, rare-element rule, PaiNN cross-validation outlier removal) yields a clean, ML-ready dataset with usable tooling (ASE .aselmdb, fairchem-core, mace-torch).
- MACE models trained or fine-tuned on SoLiD26 reduce test-set force MAE/RMSE to roughly one third of the pretrained MACE-MATPES-PBE-0+D3 baseline, indicating the data adds information not in the foundation model's training set.
- The dataset enables development and benchmarking of MLIPs for structurally and chemically heterogeneous solid-liquid interfaces.
Strengths
Unusual scale for the interface domain
15,379,823 structures with up to 576 atoms (avg 126.3) across 15 elements, ~100 GB; the paper correctly positions this against general-purpose bulk/molecule datasets and the single comparable interface effort, OC25 (Background section).
Transparent, stepwise curation pipeline
Methods lists explicit criteria: encut restricted to 350/400/450, max force <=10 eV/Å, dedup by (formula, energy, max force, center of mass), <1%-element exclusion, and PaiNN 5-fold cross-validation outlier removal; preparation code is public at github.com/team-capex/SoLiD26.
Quantified evidence the data adds information
Honest practical caveats
Weaknesses
Core artifact not yet public
Weak test split design
Provenance correlation and residual near-duplicates
ML-based outlier filtering risks biasing the PES coverage
Validation is thin and explicitly preliminary
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SoLiD26 (arXiv:2609.28013) is a data descriptor in the Scientific Data style. Its central claim is that a curated corpus of 15,379,823 VASP/PBE+D3 structures, assembled from ten prior DTU studies of aqueous metal interfaces and electrode-electrolyte systems (refs [10]-[20]), fills a genuine gap because existing MLIP datasets 'do not systematically sample the coupled response of a liquid and a solid surface' (Background & Summary). The dataset spans 15 elements (H, Li, C, N, O, F, Na, Al, Cl, K, Ca, Cu, Cs, Pt, Au), up to 576 atoms per structure (126.3 average), roughly 100 GB in .aselmdb format. A secondary claim is demonstrated utility: MACE models trained on SoLiD26 outperform a pretrained baseline on a held-out split (Table 1). Both claims are plausible and partially substantiated; the second rests on a deliberately simple split the authors themselves flag as limited.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the SoLiD26 dataset?
SoLiD26 is a curated collection of 15,379,823 first-principles atomic structures (DFT, VASP with PBE functional and D3 dispersion correction) of solid-liquid interfaces, mostly aqueous metal and electrode-electrolyte systems, compiled from DTU studies for training machine-learned interatomic potentials. It spans 15 elements, up to 576 atoms per structure (126.3 on average), and totals about 100 GB in ASE .aselmdb format.
How can I download SoLiD26?
As of this preprint, the dataset is available by request to the authors while a public release is being prepared. The release will contain 16 xz-compressed .aselmdb files (~100 GB uncompressed), MD5 checksums, and a JSON file with the train/validation/test split indices. The preparation code is already public at github.com/team-capex/SoLiD26, and the format is supported by fairchem-core and mace-torch.
How accurate are MACE models trained on SoLiD26?
On the paper's held-out test set (structures with more than 200 atoms), the fine-tuned and from-scratch MACE models reach force MAE of about 0.0118-0.0119 eV/Å and per-atom energy MAE of about 0.0026-0.0027 eV. The pretrained MACE-MATPES-PBE-0 baseline with a D3 correction added gives 0.0501 eV/Å force MAE and 0.0046 eV per-atom energy MAE, so in-domain training roughly cuts force errors to a quarter to a third. Results are labeled preliminary and hyperparameters were not optimized.
Which elements does SoLiD26 cover?
H, Li, C, N, O, F, Na, Al, Cl, K, Ca, Cu, Cs, Pt, and Au; 78.6% of structures contain hydrogen. Note an important caveat: the paper's test split (structures with more than 200 atoms) only exercises H, C, O, K, Pt, and Au, so reported accuracies do not cover the other nine elements.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are described but not analyzable from the extracted text; element distribution beyond the stated 78.6% hydrogen share is unknown.
- Table 1 is explicitly labeled 'Preliminary results' with unoptimized hyperparameters; absolute numbers may change in the public release.
- The dataset itself is request-only and was not inspectable; quality judgments rest on the described pipeline and reported counts.
- arXiv v1 preprint (23 Sep 2026), not yet peer reviewed; the full-text byline includes Bhowmik and Vegge, who are absent from the metadata author list.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.