Scalable Reactive Atomistic Dynamics with GAIA

TL;DR

GAIA presents an automated, end-to-end framework for constructing diverse training datasets to train general-purpose reactive MLIPs. By coupling a metadynamics-based Nanoreactor+-driven data generator with a data-improver that targets underrepresented regions, GAIA produces Titan25, a 1.8-million-structure dataset spanning 11 elements, enabling a Titan25-trained MLIP (SNet-T25) that closely matches DFT and experimental results across detonation, CNT coalescence, interfacial adsorption, and catalytic processes. The Titan25-trained model demonstrates broad transferability, outperforming models trained on public datasets in GAIA-Bench and reproducing experimentally observed phenomena with near-ab initio fidelity. These results establish GAIA as a practical, scalable path toward universal, generalizable MLIPs capable of describing diverse materials and chemical processes in realistic conditions.

Abstract

Groundbreaking advances in materials and chemical research have been driven by the development of atomistic simulations. However, the broader applicability of atomistic simulations remains limited, as they inherently depend on energy models that are either approximate or computationally prohibitive for large-scale simulations. Machine learning interatomic potentials (MLIPs) have recently emerged as a promising class of energy models, but their deployment also remains challenging due to the scarcity of systematic protocols for generating training data spanning diverse structural regimes. Here we introduce GAIA, an end-to-end automated framework that streamlines dataset construction for the development of general-purpose reactive MLIPs. GAIA combines a metadynamics-based exploration scheme with closed-loop data expansion for the efficient sampling of a broad spectrum of atomic arrangements, thereby addressing the reliance on heuristics in conventional dataset generation. Using GAIA, we constructed Titan25, a benchmark-scale dataset, and trained an MLIP that closely matches both static and dynamic density functional theory results. The resulting model reproduces key experimental observations across distinct modes of reactivity, including detonation, coalescence, and catalytic processes. GAIA thus helps bridge the gap between simulation and experiment, paving the way toward scalable and general MLIPs capable of describing a wide range of materials and chemical processes.

Figures (19)

• Figure 1: Large-scale atomistic simulations with the Titan25-trained MLIP.a, MLIPs have so far shown limited transferability across diverse reactions, making it difficult to apply them broadly as general models. Here, we show representative simulations enabled by a single Titan25-trained model, spanning four scales from left to right: molecular adsorption, nanoscale coalescence, non-equilibrium assembly, and catalytic reactions. Together, these examples demonstrate the broad applicability of the MLIP. b, Schematic overview of the GAIA framework for generating training datasets and deploying MLIPs. From left to right, user-provided input structures are augmented by GAIA into diverse datasets (for example, Titan25) comprising a wide range of atomic arrangements. These datasets can then be used to train MLIPs, which are subsequently applied in the simulations. c, Comparison of dataset scales, with axes showing the release year, the number of elements, and the number of data points. GAIA can generate datasets with scalable sizes and elemental diversity. As a representative case, Titan25 comprises eleven elements and 1.8 million data points. An inset table qualitatively summarizes whether each compared dataset includes de novo configurations and is capable of describing diverse reactions and interfacial processes. Further details of this summary are provided in Supplementary Note \ref{['supp:related_work']}. d, Element-wise atom counts in Titan25. e, Probability density distributions of per-atom energies (left panel) and norms of atomic forces (right panel). P95 denotes the 95$^\text{th}$ percentile. See Supplementary Fig. \ref{['sfig:titan25']} for additional details on Titan25.
• Figure 1: a, Details of GAIA data-generator. Various combinations of atomic arrangements are generated using the user-provided inputs (top) and the builders described below. $N^{\text{CB}}_\text{U}$, $N^{\text{NANO}}_\text{U}$, $N^{\text{BK}}_\text{U}$, $N^{\text{SB}}_\text{U}$, $N^{\text{AD}}_\text{U}$, and $N^{\text{AM}}_\text{U}$ indicate the maximum number of unlabeled structures that can be generated by each builder (Checkerboard, Nanoreactor$^+$, Bulk, Slab, Adatom and Admol) and $N^{\text{DI}}_\text{U}$ for diatomic curve add-on. The same symbols with the subscript "L" denote the maximum number of labeled structures. For detailed explanation, see Methods \ref{['method:dg']}. b, Illustration of bondmap examples. Averaged bond matrices of trajectories generated from a mixed structure comprising a Rh single atom, a single CO molecule, and two H$_2$O molecules are shown. The left and right panels depict the first and second unique averaged bond matrices derived from this mixed structure, obtained with an SSIM threshold of 0.98, respectively. c, Representative configurations generated by the Checkerboard (bi-elemental complexes) and the Nanoreactor$^+$ builder (metal–organic and organic complexes).
• Figure 2: GAIA data-generator and -improver produce diverse datasets.a, Illustration of the DG module. Simple periodic and non-periodic input components are expanded into diverse unlabeled configurations by builders such as Checkerboard, Bulk, Slab, Adatom, Admol, and Nanoreactor$^+$. The first five builders generate structures by retaining key input features or combining structures, while Nanoreactor$^+$ explores reactive configurations by inducing bond formation and cleavage via enhanced dynamics. The generated structures are periodically replicated to form supercells with random rotations, which are then labeled at the DFT level. b, Feature-space embedding of structures generated by the DG module (dataset G), obtained from a Titan25-trained MLIP. Each point corresponds to a generated structure, and the color denotes the builder from which the structure originated. c, Illustration of the DI module. For the exploratory data points, basis snapshots are extracted through categorical analyses that respectively rely on data distribution statistics from the training dataset and error-based measures derived from model predictions. New seed structures selected from these snapshots are fed into the Nanoreactor$^+$ builder, where they are expanded into diverse configurations to form dataset I. d, Comparison of structures generated by DG and DI using atom--atom distance diversity ($d_\Lambda$) and total energy. Heatmap elements show the relative dominance of each dataset by counts, with gray indicating dataset G and red indicating dataset I. The distribution of dataset G is further illustrated as a histogram on the side. Red regions correspond to dataset I populating sparse areas of dataset G, demonstrating its contribution of additional structures beyond those obtained from DG alone.
• Figure 2: a, Four representative transformations from input (QCG results) to output (Nanoreactor$^+$ results), indicated by arrows, during the data-generator stage. Spheres are color-coded by element; larger spheres represent metals, whereas smaller spheres represent organic elements. b, As in a, but during the data-improver stage. The inputs are expanded seeds, and the outputs are again the Nanoreactor$^+$ results.
• Figure 3: The Titan25-trained MLIP achieves reliable accuracy in both static and dynamic benchmarks.a, Plot of normalized mean errors on GAIA-Bench, the static benchmark suite. The normalization scales all bars to a common maximum for relative comparison. ANI-1xnr is shown only for the mol2mol task, as it does not include metallic elements. b, Mean absolute errors (MAEs) of relative energies (top) and forces (bottom) for the four GAIA-Bench tasks. Each plot is individually sorted in ascending order; each point on the horizontal axis corresponds to a group of snapshots derived from the same reference structure. In the lower panel, shaded regions represent the range between the smallest and largest MAE values across different values of $\epsilon$, where $\epsilon$ denotes the structure perturbation strength at which the force norm matches the threshold. For mol2surf, no additional perturbations were applied; instead, only the structures used for energetic calculations were considered. An inset provides an enlarged view of the plot near its end. Root-mean-square error (RMSE) results are provided in Supplementary Fig. \ref{['sfig:brmse']}. c, Angle and magnitude errors in atomic forces. Magnitude errors are shown with a reference level of 100 meV/ Å. d, Evolution of the Pt–C vertical distance during MD simulations of CO and H$_2$O adsorbed on the Pt(111) surface using the SNet-T25 model. Ivory, gray, white, and red spheres represent Pt, C, H, and O atoms, respectively. The moving-average curve is presented for SNet-T25; for DFT, the mean value is shown together with a horizontal dashed line; for UMA-OMat, -OC20, and -ODAC, only the mean values are provided. We denote these models as UMA-X, since the forward path of the UMA model is adapted depending on the task input X. Snapshots from the simulations with UMA-OMol and -OMC are also displayed, illustrating that the Pt surface structure was not preserved. e, As in d, except that CO is replaced by CO$_2$. The distance curves are displayed for the four MLIPs, whereas only the mean value is shown with a horizontal dashed line for DFT. The apparent truncation of the UMA-X curves arises from cases where the Pt--C distance exceeds 4.0 Å.