ace-jax

Build, fit and evaluate Atomic Cluster Expansion (ACE) interatomic potentials in Python with JAX, on CPU or GPU. One pip install gives all of these functions.

Documentation: https://acesuit.github.io/ace-jax/. It contains the installation, a quickstart, tutorials, how-to guides, and the CLI and API reference.

What it does

• Build a basis and fit it with one command. aj fit --order 3 --max-degree 10 --train train.xyz --out fit builds the symmetry-adapted ACE basis and fits it. The fit.yaml file that it writes runs the fit again.
• Bayesian fits. Linear ACE uses Bayesian linear regression. The fit selects the energy, force and virial noise levels and the prior scale by maximising the evidence, so you do not set weights by hand.
  • An experimental hybrid ACE + Gaussian-process model (--m-per-species > 0) adds a calibrated set of uncertainty approximations (MAP, Laplace, Pathfinder, VI, NUTS). Most fits use the linear model, the default.
  • The linear model has POPS and ARD uncertainties.
  • ARD gives a conformally calibrated per-atom force uncertainty (forces_std, forces_q, a 3x3 forces_cov). aj calibrate recalibrates it on new labelled cells.
  • The fit can also learn the radial basis (--learn-radial).
• Fast evaluation. ACECalculator and GPCalculator are ASE calculators, fast enough for molecular dynamics. GP and ARD models also predict energy_std and forces_std. export_lammps deploys a model to LAMMPS with lammps-jax. On CPUs, linear ACE runs in LAMMPS as a C++ kernel (aj export --platform cpu).
• PACE potentials. ace-jax loads, evaluates and writes pacemaker .yace files.
• Data in, data out. Training data is extended XYZ or a list of ase.Atoms. ace-jax converts stress labels to virials. aj eval writes predictions as extended XYZ, with an RMSE table for each configuration type.

Install

pip install ace-jax

Extras: "ace-jax[cuda]" (CUDA 12 JAX), "ace-jax[cuda13]" (CUDA 13 JAX and nvcc for the GPU kernel; NVIDIA driver 580 or later), "ace-jax[gp]" (the Pathfinder rung), "ace-jax[native]" (C++ code on CPUs: a neighbour list and the kernel), "ace-jax[lammps]" (LAMMPS export with lammps-jax) and "ace-jax[jaxmd]" (MD with JAX-MD).

To build a new basis, ace-jax needs the ace-jax-coupling wheel. This core dependency is available for Linux x86_64 and aarch64, macOS arm64 and Windows x64. On other platforms, fit and evaluate from an existing basis file. See Installation.

Quickstart

Fit a linear ACE model to labelled data, check it on a test set, and evaluate it:

aj fit --order 3 --max-degree 10 --train train.xyz --test test.xyz \
    --e0 lsq --opt lbfgs --out fit
aj eval --model fit/model.npz --data test.xyz --out predictions.xyz

Use the fitted model from Python:

import jax
jax.config.update("jax_enable_x64", True)

from ase.build import bulk
from ace_jax import ACECalculator

atoms = bulk("Si", "diamond", a=5.43, cubic=True)
atoms.calc = ACECalculator("fit/model.npz")
print(atoms.get_potential_energy(), atoms.get_forces())

The Quickstart runs this on a small silicon dataset. The tutorials are notebooks that run on a laptop CPU. Some of them come from the MLIP School 2026. For coding agents, skills/ace-jax/SKILL.md is a compact usage guide.

Performance and validation

Fits agree with the ACEfit design matrix and least-squares solution to 1e-8. PACE evaluation agrees with the ML-PACE C++ code and python-ace. CI checks both. The throughput benchmarks against LAMMPS ML-PACE are in docs/dev/benchmarks.md.

Contributing

See CONTRIBUTING.md for the development setup, tests and the reference-parity jobs. See the changelog for releases. ace-jax has the MIT licence and is part of ACEsuit.