Teaching a force field where water meets metal

A force field can know a metal and know water without knowing what happens when they meet. At their boundary, the solvent arranges itself around a surface that can respond in turn; ions and adsorbed molecules add further local environments. SoLiD26 addresses this meeting place directly. Jonas Busk and colleagues at the Technical University of Denmark have assembled more than 15 million calculated structures and tested whether they improve models of the forces acting on atoms at solid–liquid interfaces.¹

The early answer is encouraging. On the authors' larger-structure test set, training with the collection substantially reduces force errors relative to a pretrained MACE model supplied with a matching dispersion correction. The achievement is useful because it concerns a defined missing environment, rather than a claim that one model now handles every material. Understanding that achievement requires reading the construction of the dataset as carefully as its results table.¹

SoLiD26 is an unreviewed preprint, posted on 23 September. This review examines its full manuscript and supporting preparation code, extending our earlier abstract-based daily mention. The data themselves were available by request while the public release was being prepared, according to the inspected manuscript. We have not downloaded the approximately 100 GB collection or repeated its training runs.¹

A boundary built from earlier calculations

The collection gathers calculations from several DTU Energy projects studying aqueous metal interfaces, electrochemical reactions and electrode–electrolyte systems. Most structures come from ab initio molecular dynamics: trajectories whose energies and forces are calculated using electronic-structure methods. Some bulk reference systems, including water and gold, are included too. The final count is 15,379,823 structures, containing an average of 126.3 atoms and as many as 576.¹

That provenance gives the dataset a clear purpose. It consolidates configurations generated while researchers were investigating actual interfacial questions, preserving energies and forces alongside atomic identities, coordinates, cell dimensions and periodic boundary conditions. It is training material for approximating a potential-energy surface: the relationship between atomic arrangement and energy from which forces can be learned. It is not a set of measured catalytic yields or battery lifetimes.¹

The records share VASP calculations using the PBE density functional and D3 dispersion corrections. The latter is relevant to the comparison because the pretrained baseline also receives a D3 correction. Matching this part of the reference method helps make the improvement interpretable: a missing dispersion contribution should not be mistaken for an advantage conferred solely by new training examples.¹

The collection spans 15 elements, from hydrogen and lithium to platinum and gold. That sounds broad, but elemental membership alone does not specify what a model has encountered. A hydrogen atom in bulk water and one next to an adsorbed intermediate inhabit different environments. The paper's value lies in collecting those coupled solid–liquid configurations, not simply adding names to an elemental checklist.¹

Cleaning data also defines their domain

The preparation pipeline first identifies compatible calculations and then applies several filters. It retains specified energy-cutoff settings, removes structures with maximum forces above 10 eV/Å, screens duplicates and excludes structures containing elements represented in less than one percent of the dataset. A further screen uses five PaiNN models: each predicts a held-out fifth of randomly partitioned data, and structures with high force-prediction errors are removed.¹

There is a sensible motivation for these steps. An unconverged electronic calculation can supply an unreliable training label, while repeated records add cost without proportionate information. Holding each record out of the model that assesses it also avoids treating successful memorisation as evidence of data quality. The authors explicitly describe high prediction error as an indication of chemical outliers or unconverged calculations, however, rather than an infallible detector of bad physics.¹

That distinction matters. A difficult, correctly calculated environment can also be unfamiliar to a model. My interpretation is that this filtering produces a deliberately curated learning domain; it does not demonstrate coverage of every rare configuration discarded along the way. For researchers planning reactive simulations, the useful question is whether their intended environments lie within the retained domain. The paper does not provide a basis for claiming that chemically important rare events were in fact lost, and such a claim would overstate this concern.¹

The public preparation code makes parts of the process inspectable. In the pinned version examined for this review, duplicate screening compares chemical formula, energy, maximum force and the sum of centre-of-mass coordinates after rounding numerical properties to eight decimal places. This is an efficient proxy rather than an exhaustive structural-equivalence calculation. The repository README also describes an additional filtering stage outside that initial script.²

These details are useful without becoming an accusation that the dataset is defective. They tell a future user what kind of duplicate removal was performed and where further provenance would be needed to reconstruct every decision. I inspected the scripts; I did not execute them against the original calculation archive or audit the final retained indices. The relevant scientific claim remains the one the study tests: models trained on the curated collection fit its held-out reference data better.¹²

What the larger-cell test actually asks

The technical validation compares three models. The first is MACE-MATPES-PBE-0 with added D3 dispersion. The second continues training that model on SoLiD26. The third trains a MACE model from scratch on SoLiD26 with similar hyperparameters. Fine-tuning uses a learning rate of 0.0005; both training routes first emphasise forces in the loss and then increase the relative emphasis on energy. The authors describe this as preliminary validation without systematic hyperparameter optimisation.¹

The most consequential design choice is the split. Every structure containing more than 200 atoms goes into the test set: 272,889 configurations. From the smaller structures, 100,000 are selected randomly for validation and 15,006,934 are used for training. The numbers sum to the stated dataset total. Thus the study tests prediction on larger structures than those used for fitting, rather than simply scattering all configurations randomly across training and test sets.¹

That is a worthwhile test. Size matters when the motivation for an interatomic potential is to simulate systems beyond the practical reach of direct electronic-structure calculations. But larger is not synonymous with chemically unseen. The test set contains only hydrogen, carbon, oxygen, potassium, platinum and gold: six of the collection's 15 elements. The authors state this restriction and leave wider generalisation experiments to future work.¹

The reader should therefore attach the error numbers to a precise claim: transfer to the larger structures selected by this rule, within the chemistries represented there. The aggregate score is not a test of every element in the training collection, every possible electrolyte, or an entirely new surface family. This is a boundary of the experiment, not a reason to discount its result. It also identifies a useful next comparison: hold out the chemistry or interface family that a prospective application actually requires.¹

Forces improve substantially; the metrics deserve their names

Table 1 reports force mean absolute error falling from 0.0501 eV/Å for the pretrained baseline to 0.0118 eV/Å after fine-tuning and 0.0119 eV/Å for training from scratch. Calculating from those displayed values, the fine-tuned model's mean absolute error is about 76.4 percent lower, or roughly one quarter of the baseline. That arithmetic describes the reported benchmark; it is not an independently reproduced accuracy measurement.¹

Root mean squared force error tells a related but distinct story. It changes from 0.0696 eV/Å to 0.0243 eV/Å after fine-tuning and 0.0229 eV/Å from scratch. These are approximately one third of the baseline. Because squaring gives larger errors more influence, reporting both metrics helps avoid compressing the result into a single improvement factor. The paper's broad description of an approximately threefold force improvement is clearest when the individual columns remain visible.¹

Energy also improves. Mean absolute error per atom decreases from 0.0046 eV to 0.0027 eV after fine-tuning and 0.0026 eV from scratch. Expressed in millielectronvolts, those figures are 4.6, 2.7 and 2.6 meV per atom. Whole-structure energy errors have different denominators: the corresponding mean absolute errors are 1.3341, 0.8638 and 0.8953 eV. A small per-atom error should not be silently described as the error in an entire interfacial cell.¹

The two trained models are close in force mean absolute error, while their ordering changes across other columns. Fine-tuning has the lower total-energy mean absolute error; training from scratch has the lower force root mean squared error. These results support the usefulness of the dataset through both routes. They do not establish a universal winner between fine-tuning and training from scratch, especially without an optimised comparison of training budgets or repeated runs.¹

The strongest conclusion is also the most practical: this interfacial training material supplies information that improves prediction beyond the chosen broadly pretrained baseline on the authors' test. The result is about agreement with PBE-D3 reference calculations. Establishing an experimental observable, such as a reaction rate or a potential-dependent adsorption equilibrium, would require that observable to be calculated and compared; the present table does not claim to do so.¹

A neighbouring dataset asks different questions

Earlier OC25 work provides a useful comparison of experimental design. Its version-one dataset paper describes splitting by bulk–solvent combinations and constructing additional tests with unseen solvents, unseen ions and both together. Those tests ask different transfer questions from SoLiD26's atom-count split. Accordingly, their headline errors should not be placed in a common ranking.³

OC25 also recalculated validation configurations with tighter electronic convergence. A model trained on less tightly converged data agreed better with the tighter force labels, illustrating that disagreement with a reference can partly reflect numerical quality in that reference. This is relevant context for SoLiD26's curation, not evidence that either dataset shares the other's specific errors.³

Together, these methodological choices suggest a productive division of labour. One can ask whether larger configurations remain predictable, whether new chemical combinations transfer, and whether reference forces are sufficiently converged. A single attractive error cannot answer all three. SoLiD26 contributes a substantial, focused result to this programme by showing that interface-rich data improve its stated larger-structure benchmark.¹³

Preserve the force the calculation produced

One of the paper's most immediately useful observations is a software detail. Some stored structures retain constraints from the original simulations, including fixed atoms. Reading forces through ASE's ordinary accessor can apply those constraints and set some returned forces to zero. The authors instruct users to request forces with constraints disabled, or to read the underlying raw values.¹

A fixed atom need not experience zero physical force; a constraint can prevent it from moving despite that force. Training against a zero inserted by the accessor would teach a different target from the electronic calculation. This is why data handling belongs in the scientific account. The potential benefit of millions of calculated structures depends on preserving the quantities they actually contain.¹

SoLiD26 earns attention for treating the solid–liquid boundary as a domain worth collecting in its own right, and for demonstrating a sizeable improvement against a recognisable baseline. Its next uses should preserve that specificity: a documented reference method, an explicit retained domain, a test matched to the intended transfer, and unaltered force labels. Those are the conditions under which a large collection becomes useful evidence for a simulation.¹