Source-linked AI summary
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, Arghya Bhowmik, Tejs Vegge
TL;DR
Existing atomistic datasets do not systematically sample the coupled response of liquid and solid surfaces needed for solid–liquid interface MLIPs. SoLiD26 curates first-principles structures from interface studies, mostly generated by AIMD, and provides a large dataset for training and evaluation. In held-out testing, SoLiD26-trained MACE models improve on the pretrained baseline, particularly for force errors.
Problem
Common atomistic datasets do not systematically sample coupled liquid–solid interface response, which involves complex environments and is demanding to sample with first-principles calculations.
Source: Background & Summary · p. 2Method
SoLiD26 curates first-principles interface-study calculations, mostly from AIMD, and filters and prepares the structures for MLIP use.
Source: Background & Summary · pp. 2–3, Methods · p. 3Results
Both SoLiD26-trained MACE models improve over the pretrained baseline, with force MAE and RMSE reduced to approximately one third.
Source: Technical Validation · p. 5Takeaways & Limitations
The 15,379,823-structure dataset is intended for training, fine-tuning and benchmarking MLIPs on solid–liquid interfaces and related reference structures.
Source: Background & Summary · p. 3, Data Overview · p. 4Constrained atoms may have forces zeroed by ASE’s standard get_forces() method, so raw forces require apply_constraint=False or direct access to the underlying data.
Source: Usage Notes · p. 6
Abstract
from arXiv · showhide
Machine-learned interatomic potentials (MLIPs) for solid-liquid interfaces in advanced materials applications, e.g., electrochemistry, catalysis and corrosion, require training data that samples both liquid environments, the solid and the interface itself. We present SoLiD26, a curated solid-liquid interface dataset, containing 15.4 million first-principles atomic structures with up to 576 atoms and 15 chemical elements for training and evaluating MLIPs. The structures were compiled from density functional theory (DFT) calculations performed in studies of solid-liquid interfaces, with most configurations originating from ab initio molecular dynamics (AIMD) simulations. Each record contains atomic species, positions, simulation cell, periodic boundary conditions, potential energy and atomic forces. SoLiD26 includes aqueous coinage metal interfaces, electrode-electrolyte systems, and selected bulk reference structures, calculated with VASP using the PBE functional and D3 dispersion corrections. We describe the data ingestion and preparation pipeline used to construct the dataset. The application of SoLiD26 for training and evaluating MLIPs is demonstrated with a suite of MACE models on a simple training, validation and test split. The dataset enables development and benchmarking of MLIPs for structurally and chemically heterogeneous solid-liquid interfaces.
Background & Summary
Existing atomistic datasets do not systematically sample the coupled response of liquid and solid surfaces, despite the distinctive and demanding environments of solid–liquid interfaces. SoLiD26 consolidates first-principles data to support MLIP development for these settings.
Background & Summary
- MLIPs can accelerate materials discovery by simulating more atoms and longer timescales than current ab initio methods, while foundation models require large, chemically diverse training datasets.Few interface-focused datasets exist, and they complement general-purpose datasets with targeted data for fine-tuning and evaluation in electrochemical and catalytic environments.Source: Background & Summary · p. 2
- Existing datasets largely cover molecules, bulk crystals, relaxed structures or small high-temperature trajectories, but do not systematically sample coupled liquid–solid response.Interface configurations include solvent layering, surface reconstruction, ion adsorption, solvent-mediated bond changes and fluctuations across length and time scales; first-principles sampling is computationally demanding.Source: Background & Summary · p. 2
- Solid–liquid interface evolution remains an unsolved challenge in energy materials because kinetic barriers and long-range electrostatics govern rate-limiting processes.These regions experience electric fields and chemical reactions, motivating bespoke training data for MLIPs.Source: Background & Summary · p. 2
- SoLiD26 consolidates first-principles calculations from studies of aqueous metal interfaces, electrochemical reactions and electrode–electrolyte systems into a curated dataset.The release contains 15,379,823 structures, up to 576 atoms, spanning 15 elements, with selected liquid and solid reference structures as well as interfaces.Source: Background & Summary · pp. 2–3
Methods
SoLiD26 was curated from first-principles calculations across research projects, with most data generated by AIMD. The pipeline collected, filtered and prepared structures for publication.
Methods
- Most SoLiD26 data came from AIMD simulations across projects studying solid–liquid interfaces for energy and catalysis applications.The simulations provide variation in energies and forces for learning the potential energy surface.Source: Methods · p. 3
- The pipeline collected VASP structures calculated with PBE and D3, filtered the initial dataset for quality and consistency, then prepared a clean publication release.Python scripts using ASE supported the ingestion and preparation steps.Source: Methods · p. 3
- Filtering removed calculations outside selected energy cutoffs, structures with forces above 10 eV/Å, duplicates, rare-element structures and high-force-error outliers.For outlier detection, five PaiNN models used five-fold held-out evaluation; structures with high force prediction error were excluded.Source: Methods · p. 3
Data Records
SoLiD26 is distributed as sequentially numbered ASE database files, with each structure represented as an ase.Atoms object.
Data Records
- The approximately 100 GB release uses sequentially numbered .aselmdb files, storing each structure as an ase.Atoms object with atomic numbers, positions, cell, periodic boundary conditions, energy and forces.The files are provided in ASE format.Source: Data Records · p. 4
Data Overview
SoLiD26 contains more than 15 million atomic structures, primarily solid–liquid interfaces, with selected bulk references. Its structures cover 15 elements and were calculated using VASP with PBE and D3.
Data Overview
- The dataset contains 15,379,823 atomic structures, up to 576 atoms each (average 126.3), primarily representing solid–liquid interfaces with periodic boundary conditions, alongside some bulk structures such as water and gold.Potential energies and atomic forces were calculated with VASP using the PBE functional and D3 dispersion correction.Source: Data Overview · p. 4
- The 15 represented elements are H, Li, C, N, O, F, Na, Al, Cl, K, Ca, Cu, Cs, Pt and Au.Figure 2 reports the proportion of structures containing each element.Source: Data Overview · p. 4, Figure 2 · p. 4
- 78.6% of structures contain H, and Figure 2 shows the proportion of all structures containing each represented element.The figure provides a per-element view of dataset composition.Source: Figure 2 · p. 4
Technical Validation
On the held-out test set, both SoLiD26-trained MACE models improve on the pretrained baseline, especially for force errors. Evaluation uses a test set of structures larger than 200 atoms, while the model hyperparameters were not systematically optimized.
Technical Validation
- The evaluation compares a D3-corrected MACE-MATPES-PBE-0 baseline, its SoLiD26-fine-tuned version, and a model trained from scratch on SoLiD26.Fine-tuning continued training of the pretrained model at a learning rate of 0.0005 until convergence; both SoLiD26 training approaches used two stages with reversed energy and force loss weights.Source: Technical Validation · pp. 4–5
- The test set comprises all 272,889 structures with more than 200 atoms, testing generalization to larger structures but only a subset of chemistries and elements.Only H, C, O, K, Pt, and Au occur in this test set; broader generalization experiments are left for future work.Source: Technical Validation · p. 5
- Both SoLiD26-trained models significantly improve on the pretrained baseline, with force MAE and RMSE reduced to approximately one third.The authors expect further performance improvements with systematic hyperparameter optimization.Source: Technical Validation · p. 5
- The authors did not systematically optimize model hyperparameters for this technical validation and expect further performance improvements in future work.Source: Technical Validation · p. 5
Usage Notes
The usage notes explain how to load the dataset, summarize the metrics and models in Table 1, and warn that ASE constraints can alter forces unless disabled when retrieving them.
Usage Notes
- Install ASE and ase-db-backends to read .aselmdb files with ASE’s standard connection method; fairchem-core and mace-torch also support the format.Source: Usage Notes · p. 5
- Table 1 reports MAE and RMSE for energy, energy per atom, and forces across pretrained, fine-tuned, and from-scratch MACE models.Source: Table 1 · p. 6
- Structures may retain fixed-atom constraints from their original calculations, so use get_forces(apply_constraint=False) or retrieve forces directly to avoid zeroed values.Standard ASE atoms.get_forces() may set some forces to zero, potentially causing erroneous results in machine-learning applications.Source: Usage Notes · p. 6
Funding
The work acknowledges support from the Danish National Research Foundation and the Novo Nordisk Foundation.
Funding
- Support came from the Danish National Research Foundation’s CAPeX center, the Novo Nordisk Foundation’s data-science infrastructure grant, and Gefion AI Supercomputer access grants.Source: Funding · p. 11