Software
The following is an incomplete list of implementations of the ACE model or closely related models that we are aware of. Different implementations focus on different aspects, e.g. evaluation or parameterisation, linear vs nonlinear models, and may incorporate different physics (charge transfer, magnetism, polarisation, etc).
PACE : "Performant implementation of the atomic cluster expansion"
C++ evaluation-only code with LAMMPs interface
currently supports the original linear ACE model (Drautz, 2019)
fitting methodology provided through related python packages, and ACE.jl
-
Python package for fitting ACE potentials
compatible with PACE
using tensorpotential - an implementation of ACE in Tensorflow
-
open source Julia implementation for rapid prototyping
abstractions to enable applications in different domains
invariant as well and equivariant properties
WIP: general symmetry groups
WIP: integration with Flux.jl for general nonlinearities
ace-flow: WIP
Implementation of ACE in Tensorflow
implements original ACE and recursive ACE
ace-torch: WIP
builds on graph-NN facilities in torch
Related Models and Software
MTP : Moment Tensor Potentials
Alternative implementation of symmetric polynomials using moments instead of spherical harmonics; Shapeev, SIAM Multiscale Modelling and Simulation, 2016
gitlab repository; also available with QuantumATK