Ml Generative Diffcsp

DiffCSP++ Crystal Structure Generation

Goal

Generate novel crystal structures using DiffCSP++ (ICLR 2024), a diffusion model that leverages space group symmetry constraints for crystal structure prediction (CSP) and ab initio generation.

1. Prerequisites

[!IMPORTANT] GPU Required: DiffCSP++ inference is significantly faster on GPU.

• The diffcsp-agent conda environment must be installed.
• DiffCSP++ repo cloned to /home/bdeng/projects/DiffCSP-PP.
• Pre-trained checkpoints downloaded to checkpoints/ directory.

2. Available Models

3. Usage Modes

Mode 1: Single Composition via MCP Tool (Recommended)

Generate structures with exact composition using the generate_structures_with_symmetry MCP tool:

mcp_diffcsp_generate_structures_with_symmetry(
    spacegroup=58,                        # Space group number (1-230)
    wyckoff_letters="2a,2d,4g",           # Wyckoff positions (comma-separated or shorthand "adg")
    atom_types="Mn,Li,O",                 # Element per Wyckoff position
    model_name="mp_csp",                  # CSP model
    num_samples=5,                        # Number of structures to generate
    step_lr=1e-5,                         # Langevin step size
    output_dir="research/my_project"
)

Mode 2: Batch Generation from JSON File

Generate multiple structures from a JSON specification file. This is useful when you have many different compositions to generate at once.

JSON format (see examples/example.json):

[
    {"spacegroup_number": 58, "wyckoff_letters": ["2a","2d","4g"], "atom_types": ["Mn","Li","O"]},
    {"spacegroup_number": 194, "wyckoff_letters": "abff", "atom_types": ["Tm","Tm","Ni","As"]}
]

Run the batch generation script:

# Env: diffcsp-agent
python .agents/skills/ml-generative-diffcsp/scripts/batch_generate.py \
    --json_file .agents/skills/ml-generative-diffcsp/examples/example.json \
    --model mp_csp \
    --output_dir diffcsp_batch_output \
    --step_lr 1e-5

Mode 3: Ab Initio (Unconditional) Generation

Generate structures from the training distribution without specifying composition. Requires a generation model (mp_gen, perov_gen, or carbon_gen).

# Env: diffcsp-agent
python .agents/skills/ml-generative-diffcsp/scripts/unconditional_generate.py \
    --model mp_gen \
    --num_structures 100 \
    --output_dir diffcsp_gen_output \
    --step_lr 5e-6

4. Parameters

5. Output Files

• structure_XXXX.cif: Generated crystal structure files (pymatgen CIF format)
• generation_metadata.json: Generation parameters and statistics

6. Constraints

[!WARNING] Space Group Knowledge Required: You need to know the space group number and Wyckoff positions for your target composition. Use ICSD, Materials Project, or pyxtal to find these.

[!NOTE] Wyckoff Notation: Positions can be given as full labels ("2a,2d,4g") or shorthand letters ("adg"). The number prefix is the site multiplicity — it's automatically determined from the space group.

• Environments: All scripts require the diffcsp-agent conda environment.
• GPU: A CUDA GPU is recommended for reasonable generation speed.
• CSP vs Gen models: CSP models require atom_types; Gen models can generate without them.

7. Workflow Integration

DiffCSP++ works well in combination with:

• Structure relaxation: Use MLIP tools (MACE, FairChem, MatGL) to optimize generated structures
• Stability analysis: Use mat-stability to calculate E_hull
• Comparison: Generate structures with DiffCSP++, ADiT, and MatterGen for diversity

Author: Bowen Deng Contact: GitHub @learningmatter-mit