mlx-atomistic

Apple Silicon-native alpha molecular dynamics and density-functional-theory runtime, built directly on MLX and Metal —
it runs the GPU on your Mac, with no CUDA, server, or cloud.

Documentation · Overview · Molecular mechanics · DFT · API reference · Benchmarks

---

What is mlx-atomistic?

mlx-atomistic is an experimental Apple Silicon-native runtime for molecular dynamics (MD) and density functional theory (DFT), built directly on Apple's MLX array framework and the Metal GPU backend. It runs the simulation kernels on the GPU in your Mac — no CUDA, no remote cluster, no cloud. The 0.0.1 package is a strict alpha preview: early plane-wave DFT building blocks, molecular-mechanics force terms, prepared-system imports, and Jupyter-first visualization.

mlx_atomistic is the primary trajectory generator and product runtime in this repo. OpenMM, LAMMPS, and the source trees under vendors/ are reference and validation surfaces only — they never replace the MLX runtime path.

Features

• Apple Silicon native — MLX arrays execute through the Metal backend on your machine. No CUDA, no server, no cloud.
• Plane-wave DFT building blocks — Γ-point Kohn–Sham SCF, LDA plus public-alpha PBE-PZ81 GGA diagnostics, non-SCF k-point/band diagnostics, Davidson/preconditioned-residual eigensolver diagnostics, and GTH/UPF pseudopotentials with proof-level local + nonlocal projector diagnostics.
• Molecular-mechanics building blocks — Lennard-Jones, Coulomb, harmonic bonds and angles, periodic + Ryckaert-Bellemans torsions, NVE/Langevin NVT, bounded PME proof surfaces, and proof-level barostat/NPT diagnostics.
• Prepared-system imports — AMBER prmtop/inpcrd, CHARMM PSF/parameter, and GROMACS .top/.gro subsets, with explicit physical-unit metadata.
• Reference-validation ready — OpenMM and LAMMPS surfaces are opt-in local validation lanes, not package/runtime dependencies.
• Self-documenting — Google-style docstrings generate the API reference and an llms.txt for agentic tools.

Quick start

Install the alpha package from PyPI into a Python 3.13 environment:

uv run --no-project --python 3.13 --with mlx-atomistic \
  python -c "import mlx_atomistic as ma; print(ma.__version__)"

For notebook or development work from a checkout:

uv venv --python 3.13
uv sync --extra notebook --extra prep --extra viz
uv run python -m ipykernel install --user --name mlx-atomistic --display-name "mlx-atomistic"
uv run jupyter lab

Plain uv sync uses the light test group by default. OpenMM and LAMMPS are installed only when you explicitly request the reference or dev group.

If uv cannot use the home cache in a sandboxed run, point it at a writable cache:

UV_CACHE_DIR=/tmp/mlx-atomistic-uv-cache uv sync --extra notebook --extra prep --extra viz

Benchmarks

uv run python -m mlx_atomistic.benchmarks.lj_md --particles 256 --steps 20
uv run python -m mlx_atomistic.benchmarks.lj_md --sizes 128,512,2048 --steps 20 --json
uv run python -m mlx_atomistic.benchmarks.mm_force_terms --evaluations 20 --json
uv run python -m mlx_atomistic.benchmarks.validation_gauntlet --json
uv run python -m mlx_atomistic.benchmarks.stability --json
uv run python -m mlx_atomistic.benchmarks.dft_scf --sizes 8,16,24,32 --iterations 5 --mixer both --json

Documentation

The full documentation — narrative guides plus an auto-generated API reference — lives at appautomaton.github.io/mlx-atomistic:

• Overview — what the runtime is and how the pieces fit.
• Molecular mechanics — topology, force-field terms, virtual sites, GBSA/OBC, soft-core λ, replica exchange.
• Density functional theory — plane-wave SCF, exchange-correlation, mixing, forces.
• API reference — generated directly from the package docstrings.
• Benchmarks — validation gauntlet, stability, and performance methodology.

The narrative source lives in docs/ and the site itself in site/; the API pages are regenerated from the package on every deploy, so they never drift from the code.

Runtime boundary

mlx_atomistic is the product runtime. Low-level MD kernels accept Lennard-Jones reduced-unit inputs unless a caller converts at the API boundary; prepared-system artifacts carry explicit physical-unit metadata. Sparse trajectory frames are kept separately from dense per-step diagnostics. NVE is available for energy-drift checks and Langevin NVT for seeded temperature-controlled runs. See docs/runtime-boundaries.md for the dependency roles and reference-engine provenance, and docs/units.md for the unit policy.

Part of App Automaton

mlx-atomistic is part of App Automaton — open skills, harnesses, and on-device tools for engineering with AI coding agents, with a focus on local-first, Apple-Silicon-native execution.

License

MIT.