MLIP Model¶
A machine learning interatomic potential (MLIP) surrogate built around the
open-source mlip package. Accepts
structure dictionaries stored in Candidate.data and predicts per-structure
energies, making it suitable for active learning over atomistic systems.
With MLIPModelConfig.model_path=None the model trains from scratch. Set
model_path to a pretrained checkpoint to finetune. MLIPModelConfig.model_type
selects the underlying mlip architecture ("mace", "nequip",
"visnet", or "esen"), and network_config accepts the corresponding
mlip network config object (for example Mace.Config(...)). When finetuning,
weights are reinitialised to the pretrained values at the start of every train()
call, so each active-learning iteration starts from the same foundation model rather
than the previous iteration’s fit.
Key properties:
Input:
Candidate.datadictionaries compatible withmlip.data.ChemicalSystemkeyword arguments, such asatomic_numbersandpositions. Force labels for training are supplied asCandidate.features["forces"].Output: Standard
Predictionswith per-structure energy means.Optimizer: Native
mlipoptimizer config objects are accepted viaMLIPTrainConfig.optimizer_configLoss: Native
mliploss objects are accepted viaMLIPTrainConfig.loss; the default isMSELoss.Architectures: MACE, NequIP, ViSNet, and eSEN via
MLIPModelConfig.model_type.Reference energies: Scratch training computes per-element reference energies from the training split. Finetuning derives target-domain reference energies from the training split and merges them with the pretrained species table so checkpoint parameter shapes remain compatible while the absolute energy zero is retargeted.
Sampling: Not supported — raises
NotImplementedError.
Note
This model requires the optional mlip dependency. Install the CPU/default
build with the [mlip] extra:
pip install "alf-tools[mlip]"
To opt into the CUDA 13 build, use the [mlip-cuda13] extra instead:
pip install "alf-tools[mlip-cuda13]"