[ICLR 2026] LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery

Official implementation of “LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery”

🧠 What is LLEMA?

LLEMA is a unified framework that uses large language models (LLMs) + chemistry-informed evolutionary rules + surrogate predictors to discover novel, stable, synthesizable materials faster. It tackles the challenge of balancing conflicting objectives (e.g., bandgap vs. stability, conductivity vs. transparency) by combining reasoning, evolution, and prediction.

🚀 Key Contributions

• LLM-driven candidate generation under property constraints
• Evolutionary memory loop with chemistry-informed operators
• Multi-objective optimization using surrogate models
• Benchmark suite of 14 materials discovery tasks across electronics, energy, aerospace, coatings, and optics
• Empirical results: higher hit rates, stronger Pareto fronts, and broader diversity.

🔧 Getting Started

Requirements:

• Python 3.11+

Steps:

1. Clone this repository

git clone https://github.com/your-org/LLEMA.git
cd LLEMA/

2. Create and activate an environment

conda env create -f environment.yml  # creates env defined in file
conda activate llema

3. Install Python dependencies

pip install --upgrade pip
pip install -r requirements.txt

To run surrogate models locally, clone their repos (see below).

API Keys and Configuration

You must provide API keys before running the agent:

• OPENAI for LLM calls: OPENAI_API_KEY
• Materials Project for structure/property queries: MATERIALS_PROJECT_API_KEY

Recommended: Copy the example environment file and fill in the values.

cp env.example .env
# edit .env and set OPENAI_API_KEY and MATERIALS_PROJECT_API_KEY

Environment variables read by LLEMA (subset):

• OPENAI_API_KEY – used by the agent LLM interface
• LLM_MODEL – optional, defaults to gpt-4o-mini
• MATERIALS_PROJECT_API_KEY – used by property extraction utilities
• SURROGATE_MODELS_DIR – optional, defaults to src/surrogate_models

If not using a dotenv loader, you can also export them in your shell before running:

export OPENAI_API_KEY=...
export MATERIALS_PROJECT_API_KEY=...

Note: The src/agent/config.py file contains run-specific information such as iteration limits, memory settings, and multi-island configuration parameters.

Surrogate Models

LLEMA integrates fast surrogate models to estimate materials properties during the search loop.

ALIGNN (Atomistic Line Graph Neural Network)

• Pretrained models from JARVIS-DFT are to be downloaded and stored under src/surrogate_models/alignn/alignn/ as .zip archives.
• For details on which archives are included and local customizations, see src/surrogate_models/README.md.

cd src/surrogate_models
git clone https://github.com/usnistgov/alignn.git

CGCNN (Crystal Graph Convolutional Neural Network)

• CGCNN can be used as an alternative or complementary surrogate.
• LLEMA includes minor output-format changes for clearer, property-specific CLI output.

cd src/surrogate_models
git clone https://github.com/txie-93/cgcnn.git

See src/surrogate_models/README.md for more details on supported properties and output formats.

⚙️ Quick Start

Run the full benchmark suite via a bash script:

cd src
bash run_all_tasks.sh

📊 Evaluation

LLEMA provides tools to evaluate CIF files for validity (property constraints) and stability analysis on the taks in LLEMABench. This section describes how to use these evaluation scripts.

Validity Analysis

The calculate_validity.py script evaluates CIF files against task-specific property constraints to determine if they meet the requirements for a given materials discovery task.

Usage:

cd src
conda activate llema  # Ensure the mat_sci environment is activated
python calculate_validity.py --tasks <task_name> [options]

Examples:

# Evaluate CIF files for a specific task
python calculate_validity.py --tasks "Hard, Stiff Ceramics"

# Evaluate for all available tasks
python calculate_validity.py --tasks all

Arguments:

• --tasks: Task name(s) to evaluate. Use "all" to process all tasks, or specify one or more task names.
• --cif-dir: Directory containing CIF files to process (default: example)
• --output-dir: Output directory for results (default: auto-generated with timestamp in validity_output/)

Output:

• Results are saved in validity_output/property_output_<timestamp>/ directory
• Each task generates a results_<task_name>.jsonl file containing:
  • Compound formula
  • Calculated property values (band gap, formation energy, bulk modulus, etc.)
  • Categorical constraint results (earth_abundant, non_toxic, etc.)
  • Successful and failed constraint checks
  • Materials API usage flag

Stability Analysis

The calculate_stability.py script analyzes the thermodynamic stability of candidates from validity analysis results by calculating energy above hull and formation energy.

Usage:

cd src
conda activate llema  # Ensure the mat_sci environment is activated
python calculate_stability.py --task <task_name> [options]

Arguments:

• --task or -t: Specific task name to analyze (required)
• --max-samples or -n: Maximum number of samples to process per task
• --quiet or -q: Reduce output verbosity (only show summary statistics)
• --output-dir: Specific validity output directory to process (default: latest)

Output:

• Summary statistics are saved in stability_output/stability_summary_<timestamp>.json
• The JSON file contains:
  • Overall statistics: total candidates, valid/invalid counts, stability breakdown (stable/marginally stable/unstable/unknown)
  • Task-specific breakdown with detailed statistics
  • Energy above hull calculation success rates
  • Materials API and surrogate model usage statistics

Note: The stability analysis script automatically searches for results in validity_output/property_output_* directories and maps CIF files from the example directory (or the directory specified during validity analysis).

📚 Citation

@inproceedings{abhyankar2026llema,
        title={LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery},
        author={Abhyankar, Nikhil and Kabra, Sanchit and Desai, Saaketh and Reddy, Chandan K},
        booktitle={The Fourteenth International Conference on Learning Representations (ICLR)},
        year={2026},
        url={https://openreview.net/forum?id=TIqzhBvCNB}
}

📄 License

This repository is licensed under the MIT License.

📬 Contact Us

For any questions or issues, you are welcome to open an issue in this repo or contact us at nikhilsa@vt.edu and sanchit23@vt.edu.