Materials Science Adsorption Agent

A Python agent for automated materials science workflows, focusing on surface adsorption analysis using Materials Project data, ASE/pymatgen tools, and FairChem UMA machine learning potentials.

Features

1. Materials Project Integration: Retrieve crystal structures using MP API
2. Structure Conversion: Convert pymatgen structures to ASE Atoms objects
3. Surface Cleaving: Generate surface slabs with user-specified Miller indices and terminations
4. Adsorbate Selection: Choose from common adsorbates (H, O, CO, OH, N, etc.)
5. Adsorption Site Analysis: Find optimal adsorption sites using pymatgen algorithms
6. ML-Accelerated DFT: Calculate adsorption energies using FairChem UMA models
7. Results Export: Generate CIF files and comprehensive JSON reports

Installation

Requires Python 3.11–3.13 (driven by fairchem-core 2.x).

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

pip install agent-materials-science

# Install the package (geometry workflow: slabs + adsorption sites)
pip install .

# Optional: ML adsorption energies (installs torch ~=2.8, large download)
pip install ".[ml]"

# Optional: AdsorbML production workflows
pip install ".[adsorbml]"

# Set up Materials Project API key
export MP_API_KEY="your_api_key_here"
# Or create a .env file with: MP_API_KEY=your_api_key_here

UMA checkpoints are gated on Hugging Face: request access to facebook/UMA and run huggingface-cli login once.

Quick Start

Command Line Interface

# Basic usage - analyze silicon (111) surface with H adsorbate
python -m agent_materials_science.cli --material Si --miller 1,1,1 --adsorbate H

# Using Materials Project ID
python -m agent_materials_science.cli --mp-id mp-149 --miller 1,1,0 --adsorbate CO

# Full workflow with relaxed energy calculations
agent-materials-science \
    --material SrTiO3 \
    --miller 1,0,0 \
    --adsorbate O \
    --layers 6 \
    --vacuum 15 \
    --supercell 2,2 \
    --calculate-energies \
    --relax --fix-layers 2 \
    --device auto \
    --output-dir outputs/srtio3_analysis

The agent-materials-science console script is installed with the package; python -m agent_materials_science.cli works too.

Python API

from agent_materials_science import MaterialsScienceAgent, AgentConfig

# Create agent configuration
config = AgentConfig(
    material="Si",
    miller_indices=(1, 1, 1),
    adsorbate="H",
    n_layers=6,
    vacuum=15.0,
    supercell=(2, 2),
    calculate_energies=True,
    output_dir="outputs"
)

# Run agent
agent = MaterialsScienceAgent(config)
result = agent.run()

# Access results
print(f"Material: {result.formula} ({result.material_id})")
print(f"Best adsorption site: {result.best_site}")
print(f"Output files: {result.output_files}")

Configuration

Environment Variables

Explicit constructor/CLI arguments always take precedence over environment variables.

Supported Adsorbates

27 species, including atoms (H, O, N, C, S, F, Cl), diatomics (H2, O2, N2, CO/OC, OH/HO, NO/ON, HF, HCl) and molecules (H2O, CO2, N2O, NH3, CH4, C2H2, C2H4, HCOO, CH3OH). Run agent-materials-science --list-adsorbates for the authoritative list with descriptions.

FairChem Models

UMA models require FairChem v2 and a Hugging Face account with access granted to the UMA model repository (huggingface-cli login).

• uma-s-1p2 - UMA small v1.2 (latest small model; recommended, default)
• uma-s-1p1 - UMA small v1.1
• uma-m-1p1 - UMA medium v1.1 (higher accuracy, slower)
• esen-* - eSEN models

uma-s-1 was removed from the FairChem registry and can no longer be downloaded. Run agent-materials-science --list-models to query the live registry when fairchem-core is installed. A local checkpoint file path can be passed via --model /path/to/checkpoint.pt.

Pick the task head to match the problem: oc20 for surface adsorption / catalysis (the default here), omat for bulk inorganic materials, omol for isolated molecules. Mixing task heads between the slab and the reference makes adsorption energies meaningless.

Adsorption-site backend

Site finding uses pymatgen's AdsorbateSiteFinder when pymatgen is installed (--site-finder auto, the default). It is symmetry-aware, so it returns only the distinct sites and avoids redundant energy evaluations. A self-contained, periodicity-correct geometric finder is used as a fallback (--site-finder builtin). Pass --no-symm-reduce to keep every site.

Project Structure

agent_materials_science/
├── __init__.py           # Package exports
├── agent.py              # Main agent orchestration
├── cli.py                # Command line interface
├── config.py             # Configuration management
├── core/
│   ├── __init__.py
│   ├── adsorption.py     # Adsorption site finding
│   ├── surface.py        # Surface/slab generation
│   └── workflow.py       # Workflow orchestration
├── tools/
│   ├── __init__.py
│   ├── materials_project.py  # MP API wrapper
│   ├── ase_tools.py      # ASE utilities
│   ├── fairchem_calc.py  # FairChem calculator
│   └── converters.py     # Structure converters
├── outputs/              # Default output directory
├── requirements.txt
└── README.md

Output Files

The agent generates the following output files:

• {material_id}_slab_{hkl}.cif - Clean slab structure
• {material_id}_slab_{hkl}.vasp - VASP POSCAR format
• {material_id}_sites.json - Adsorption sites data
• {material_id}_best_site.cif - Structure with adsorbate at best site
• {material_id}_results.json - Complete analysis results

Methodology notes & limitations

• Adsorption-energy convention. E_ads = E(slab+ads) − E(slab) − E_gas(ads), where E_gas follows the OC20 convention: a linear combination of per-element reference energies (H −3.477, C −7.282, O −7.204, N −8.083 eV, from the OC20 paper) that pairs consistently with UMA oc20 total energies — this is the scheme recommended by the FairChem documentation. Adsorbates containing other elements fall back to relative energies (site ranking is still valid) unless you provide per-element overrides via get_adsorbate_reference_energy(..., overrides=...).
• Consistent relaxation. With --relax, the clean slab is relaxed once (before site finding) and each slab+adsorbate system is relaxed, so both totals in E_ads come from relaxed geometries. Use --fix-layers 2 (or more) to hold the bottom layers bulk-like during relaxations.
• ML accuracy. Energies come from a machine-learned potential (UMA); expect ~0.1 eV-scale deviations from explicit DFT depending on the system.
• Single initial placement per site. Each site is evaluated from one initial geometry (optionally relaxed). The state-of-the-art recipe is AdsorbML: generate many initial configurations, ML-relax each, and take the minimum — available via pip install "fairchem-core[adsorbml]" (or higher-level libraries such as quacc slab_to_ads_flow and atomate2) for production studies.
• Surface-normal assumption. Site finding assumes the surface lies in the xy-plane with vacuum along z (the convention produced by the slab builder).

License

MIT License