LeMat-Traj: A Scalable and Unified Dataset of Materials Trajectories for Atomistic Modeling

TL;DR

LeMat-Traj tackles data fragmentation in quantum-chemistry trajectories by presenting a unified dataset of over 120 million DFT relaxation configurations from Materials Project, Alexandria, and OQMD, harmonized across functionals PBE, PBESol, SCAN, and r2SCAN. It is complemented by LeMaterial-Fetcher, an open-source pipeline that fetches, transforms, validates, and harmonizes data for reproducible, multi-source dataset construction and distribution via platforms like HuggingFace. The authors demonstrate the dataset’s value through fine-tuning gains on relaxation tasks, improved Matbench Discovery performance, and promising multi-fidelity transfer, highlighting the potential for broader applications including self-supervised pretraining and curriculum learning. While MD data and cross-source biases remain open challenges, LeMat-Traj provides a scalable foundation for training transferable MLIPs and accelerating data-driven materials discovery through community-driven evolution.

Abstract

The development of accurate machine learning interatomic potentials (MLIPs) is limited by the fragmented availability and inconsistent formatting of quantum mechanical trajectory datasets derived from Density Functional Theory (DFT). These datasets are expensive to generate yet difficult to combine due to variations in format, metadata, and accessibility. To address this, we introduce LeMat-Traj, a curated dataset comprising over 120 million atomic configurations aggregated from large-scale repositories, including the Materials Project, Alexandria, and OQMD. LeMat-Traj standardizes data representation, harmonizes results and filters for high-quality configurations across widely used DFT functionals (PBE, PBESol, SCAN, r2SCAN). It significantly lowers the barrier for training transferrable and accurate MLIPs. LeMat-Traj spans both relaxed low-energy states and high-energy, high-force structures, complementing molecular dynamics and active learning datasets. By fine-tuning models pre-trained on high-force data with LeMat-Traj, we achieve a significant reduction in force prediction errors on relaxation tasks. We also present LeMaterial-Fetcher, a modular and extensible open-source library developed for this work, designed to provide a reproducible framework for the community to easily incorporate new data sources and ensure the continued evolution of large-scale materials datasets. LeMat-Traj and LeMaterial-Fetcher are publicly available at https://huggingface.co/datasets/LeMaterial/LeMat-Traj and https://github.com/LeMaterial/lematerial-fetcher.

Figures (11)

• Figure 1: Data curation pipeline of LeMaterial-Fetcher. The library automates the process of fetching, transforming, validating, and harmonizing data from various sources, ensuring a consistent and reproducible dataset. The pipeline currently supports the continuous integration of fully relaxed bulk structures and full relaxation trajectories.
• Figure 2: Chemical distribution in number of trajectories for the PBE split of LeMat-Traj using Pymatvizriebesell_pymatviz_2022.
• Figure 3: Comparison of trajectory length distributions for LeMat-Traj (PBE split), MPtrj, and MatPES, on a log-log scale. For every trajectory, the number of configurations associated is computed. LeMat-Traj exhibits a broader range of trajectory lengths.
• Figure 4: Evolution of mean energy variation ($\Delta E = E^t - E^T$, where $E^t$ is current step energy and $E_T$ is final relaxed energy) per atom (a) and average maximum atomic force (b) as a function of the fraction of relaxation completed. Trajectories from LeMat-Traj, MatPES, and MPtrj are binned by their normalized progress. Solid lines represent the mean values, and shaded areas depict one standard deviation, both on a logarithmic y-axis. LeMat-Traj demonstrates comprehensive sampling from high-energy/high-force initial states to well-converged, low-energy/low-force final states.
• Figure 5: Projected Potential Energy Surfaces (PES) for the metallic Fe-Cu-Al-Ni systems. Atomic configurations are featurized using SOAP descriptors HIMANEN2020106949 and projected onto their first two principal components. The PCA 1 and PCA 2 axes are qualitative representations of structural similarity and do not have a direct physical interpretation. Color indicates formation energy (eV/atom). (a) PES derived from the LeMat-Traj PBE dataset. Green circles and black stars mark initial and final structures of example trajectories (red lines). The visualization highlights LeMat-Traj's dense, high-frequency sampling of the PES, which is crucial for resolving fine details near energy minima. (b) PES derived from the MatPES dataset, showing a broader but sparser sampling of the overall landscape.