Ensemble model for evaluating uncertainties

ASE data reader

Neighbor list implemented via PyTorch. The correctness is verified by comparing results to ASE and asap3. This class enables the direct calculatio

Rescale/shift outputs with per-property HeadConfig. Domain-specific handling is provided by MultiDomainRescaleShift.

Represents a diagonal kernel matrix, where get_column() and get_sq_dists() is not implemented. :param g: Diagonal of the kernel matrix.

r""" Fully connected linear layer with activation function. .. math:: y = activation(x W^T + b)

Base class for neural network potentials.

Batch neighbor list

Orchestrate feature extraction, aggregation, and optional distance scoring.

Extract features and gradients from linear-like layers.

Neighbor list backed by CURATOR's native C++ cell-list kernel.

Compute and cache features over datasets and ensembles.

Append-only HDF5 store for model features.

ML model calulator used for ASE applications

Copmute a one-hot floating point encoding of atoms' discrete atom types. Args: set_features: If ``True`` (default), ``node_features`` wil

Add ZBL pair-repulsion energy to the total energy before scale/shift.

This class is used to process neighbor list computed from asap3 or compute neighbor list using the internal neighbor list algorithm. It has three

Ziegler-Biersack-Littmark (ZBL) pair repulsion with polynomial cutoff.

Agnesi transform used for radial distance preprocessing.

Shift with atomwise control: - atomwise_shift=True: apply a per-atom shift (broadcast) to atomwise property. - atomwise_shift=False: appl

Configuration settings for CURATOR-LAMMPS integration.

TorchScript-friendly wrapper that adapts CURATOR models to OpenMM inputs. The wrapper expects OpenMM-style inputs (positions in nm, box vectors i

Lightweight wrapper to call CURATOR models inside torchsim graphs. - Accepts either an `ase.Atoms` instance or a tensor dictionary with

Distance metrics over feature tensors.

Parse a feature spec and compute the final feature representation.

input: m x n x p matrix m: number of models n: number of entries p: dimensionality of features

Function class to run GPAW calculation.

Compute features, build kernel matrices, and select structures.

Args:

CURATOR integration for LAMMPS using the MLIAP interface.

MACE model.

Run an official MACE model on CURATOR batches for feature-based selection. The adapter intentionally keeps the official MACE runtime in charge of

Adapter that lets MatGL models run on CURATOR atoms batches.

Utility wrapper that combines train datasets across domains while keeping domain-specific validation and test loaders.

Domain-aware wrapper around a base neural network potential.

Domain-aware wrapper holding one GlobalRescaleShift per domain. Selects by properties.domain at forward.

NequIP-style representation aligned with NequIP 0.17 defaults.