majiayu000 name: mlip-guide description: Machine Learning Interatomic Potentials (MLIPs) (4 sub-skills: mace-advanced, mlip-validation, torchsim-batch, universal-mlip)
Machine Learning Interatomic Potentials (MLIPs)
Universal and specialized machine learning interatomic potentials for rapid atomistic simulation without DFT. Covers model selection, usage patterns, validation strategies, and known limitations.
Sub-Skills
| Sub-Skill | Directory | Description |
|---|---|---|
| Universal MLIPs | universal-mlip/ |
MACE-MP-0, CHGNet, M3GNet/MatGL, SevenNet-0: setup, usage, benchmarking, and validation against DFT |
| TorchSim Batch GPU | torchsim-batch/ |
GPU-accelerated batch MD and optimization with TorchSim: 10-100x speedup, auto-batching, parallel relaxation and screening |
Method Decision Guide
What do you need MLIPs for?
Quick geometry relaxation / screening many structures?
--> universal-mlip/ (MACE-MP-0 medium is fastest, CHGNet is a good alternative)
--> torchsim-batch/ (GPU available? 10-100x faster batch relaxation with TorchSim)
Molecular dynamics (phonons, thermal, diffusion)?
--> universal-mlip/ (MACE-MP-0 large for best accuracy)
--> torchsim-batch/ (GPU available? TorchSim for 10-100x faster MD on GPU)
Elastic constants / equation of state?
--> universal-mlip/ (any universal MLIP; validate against DFT for novel systems)
Band gaps / electronic properties / magnetic ordering?
--> MLIPs CANNOT predict these. Use Quantum ESPRESSO DFT instead.
System contains rare elements / extreme conditions?
--> Validate MLIP against DFT first; MLIPs may extrapolate poorly.
Pre-installed vs. Installable
| MLIP | Status | Install Command |
|---|---|---|
| MACE-MP-0 | Pre-installed | -- |
| CHGNet | pip install | pip install chgnet |
| M3GNet (MatGL) | pip install | pip install matgl |
| SevenNet | pip install | pip install sevenn |
| ORB Models | pip install | pip install orb-models |
| TorchSim | pip install | pip install torch-sim (requires PyTorch + CUDA for GPU) |