betaFeedback

WorksPhysics and Astronomy

Orchestr.AI: A Unified Workflow for Training and Benchmarking of Machine-Learned Interatomic Potentials

Abstract

Machine-learning force fields enable atomistic simulations at near-reference accuracy and substantially reduced computational cost, but their development remains fragmented across software frameworks with incompatible data formats, configuration schemes, dependencies, and execution procedures. Orchestr.AI addresses this limitation through a unified workflow that coordinates different MLIP software frameworks, referred to as engines, for data preparation, model training, fine-tuning, evaluation, and deployment. A common YAML interface defines each experiment, while engine-specific environment profiles preserve compatible software stacks and avoid dependency conflicts. The platform currently supports five MLIP model architectures, namely SchNet, PaiNN, NequIP, Allegro, and MACE, through their corresponding software implementations, providing automated configuration mapping and standardized dataset handling across these frameworks. Benchmarks on the rMD17 and 3BPA datasets show that Orchestr.AI reproduces the expected performance trends while enabling controlled cross-engine comparisons. Repeated executions with fixed configurations and engine profiles produce highly consistent results, demonstrating practical workflow reproducibility. Beyond training, Orchestr.AI provides modular architectural interfaces that connect trained models to ASE-based simulations, batch inference, epistemic uncertainty estimation, calibration, and active-learning workflows. The platform therefore provides a reproducible and extensible infrastructure for developing, comparing, and deploying machine-learning interatomic potentials, which we validate here through cross-engine accuracy and repeatability benchmarks.

Paper details

vlnc:W_zdxvp76r
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…