ml-cluster-expansion

To train a cluster expansion model for disordered alloys, supply a disordered primordial structure, generate initial ordered structures, relax and label them via MLIP MCP tools, and train using mcp_smol_train_cluster_expansion to produce a cluster_expansion.json file.

Building a cluster expansion for compositionally disordered materials requires a disordered primordial structure, extensive total energies in the training data with consistent fixed-cell mapping, and cluster-orbit feature definitions cut off by pair and triplet radii.

Yes, you can run lattice Monte Carlo simulations to sample finite-temperature configurations from a trained cluster expansion model using the mcp_smol_run_monte_carlo MCP tool, which also supports extracting candidate structures for continued training.

Yes, the cluster expansion workflow supports active learning through an agent-driven iterative loop that handles sampling, labeling, and training via MCP tools to iteratively refine the model for disordered lattice materials.

Yes, you can apply custom regularization when fitting a cluster expansion model by using the optional direct feature-matrix fitting capability, which supports advanced regularization techniques such as sparse group lasso.

Cluster expansion for multi-sublattice disorder requires consistent fixed-cell mapping across all training data and depends on accurately defined cluster-orbit feature cutoffs, meaning incomplete or inconsistently mapped energy datasets will produce unreliable models.