Autoplex

Machine-learned interatomic potentials (MLIPs) enable large-scale atomistic simulations of materials with near quantum-mechanical accuracy. Training an MLIP requires a dataset of reference calculations, which is typically constructed manually or semi-automatically. Representative structures must be sampled for the material of interest and then evaluated using the ground-truth reference method (typically, using a density-functional theory code), making the process time-consuming, reliant on expert knowledge, and often difficult to reproduce. Autoplex automates this process. Currently, two workflows are implemented, each tailored to a specific simulation task. The first automates the fitting of custom MLIPs, including targeted dataset generation via iterative random-structure search, while the second automates the fitting and benchmarking of MLIPs for phonon calculations. Autoplex interfaces with multiple MLIP fitting frameworks, as well as atomate2 and ASE for high-throughput computations, enabling efficient resource management and remote execution on high-performance computing systems. The primary output is a trained potential, along with the structural configurations used in its development.

Autoplex was developed to address a major bottleneck in atomistic machine learning: the generation and curation of training data for machine-learned interatomic potentials (MLIPs). While fitting frameworks for MLIPs have matured considerably, assembling suitable training datasets often remains time-consuming, difficult to reproduce, and dependent on expert knowledge. Aiming to facilitate the development of new MLIP models and their application in computational materials chemistry, the project was designed from the outset as an open framework for the wider community.

The development team combines expertise in computational chemistry, materials science, machine learning, and scientific workflow automation. The project is a collaboration between the University of Oxford, BAM Berlin, and Friedrich Schiller University Jena, and is supported by a substantial grant from UK Research and Innovation. Rather than creating a standalone software stack, autoplex is built around established community software, particularly atomate2, ASE, pymatgen, and the Materials Project ecosystem. This facilitates interoperability, reuse of existing workflows, and long-term maintainability.

The framework assumes that users are familiar with high-throughput computational materials science and related workflow tools. To support adoption, extensive documentation is provided, including tutorials, workflow examples, installation guides, and teaching material developed for training schools.

Autoplex can connect to high-performance computing resources for both generating reference data and training MLIPs. Development and benchmarking have used national-scale HPC systems, including ARCHER2 in the UK and SuperMUC-NG in Germany. Training data are generated specifically for the target material systems through automated workflows rather than relying on fixed benchmark datasets.

The project is guided by principles of interoperability, reproducibility, and openness. Reproducibility is supported through the publication of workflow definitions, configuration files, datasets, trained models, and analysis notebooks, enabling the reproduction of published results.

Autoplex provides a systematic, reproducible, and scalable approach to developing machine-learned interatomic potentials (MLIPs). By automating dataset generation, model fitting, and evaluation, it reduces reliance on expert-driven trial-and-error workflows and makes it feasible to train and compare MLIPs across a broad range of materials systems. The framework can be reused across increasingly complex scientific problems.

A key benefit is improved efficiency without compromising model qualityThe capability of autoplex to extend MLIP simulations to new material classes was illustrated by the In-Sb-Te material system: autoplex created a machine-learning interatomic potential where, to the best of our knowledge, none had previously existed. More importantly, the training dataset was generated through automated exploration rather than months of expert-guided refinement. The framework also enables systematic studies that would otherwise be prohibitively time-consuming, such as comparing MLIPs trained at different levels of electronic-structure theory. Because the workflow relies on single-point reference calculations, higher levels of theory can be incorporated at moderate computational cost.

With autoplex it is easy to train and evaluate performance of the generated data across different MLIP architectures. In one example, a NequIP model trained on an RSS-generated dataset showed improved performance for previously unseen ice polymorphs, demonstrating that the resulting datasets can support developments beyond the original fitting framework.

Openness and reproducibility are central design goals. The software is openly available and accompanied by documentation, tutorials, and workflow examples. Published studies include workflow definitions, configuration files, datasets, trained models, and analysis notebooks, and are openly available on GitHub, allowing results to be reproduced and reused by others. The framework builds on established community software, including atomate2, ASE, and pymatgen, promoting interoperability across computational environments.

The use of AI methodology (specifically, machine learning) is explicit and transparent: MLIPs are trained from quantum-mechanical reference data, and model quality is assessed through automated validation metrics and comparison plots. Current work focuses on bulk materials, while surfaces and interfaces remain important targets for future development.

Github: https://github.com/autoatml/autoplex
Introduction of the software and early capability demonstrations: https://www.nature.com/articles/s41467-025-62510-6
Application (amorphous arsenic): https://doi.org/10.1021/jacs.5c18688