dACE.jl learns many-body interaction rules of collectives from trajectory data using the directional atomic cluster expansion. Its rotation-equivariant models separate contributions by body order, making learned interactions interpretable in 2D and 3D.
Requires Julia 1.12 or later.
The package is a companion to the paper and is not yet registered.
using Pkg
Pkg.add(url="https://github.com/TSGut/dACE.jl")For examples, see the documentation.
The paper detailing the theoretical background of the dACE approach is the primary reference for this package: (TODO)
If you find this package useful, please also cite the software itself through its Zenodo record, 10.5281/zenodo.23762609, which resolves to the latest release.
dACE.jl is built on Polynomials4ML.jl, ACEfit.jl and NeighbourLists.jl from ACEsuit, with benchmark collectives (in the separate helper lib/dACESwarms) simulated via Agents.jl.