Cleavage Energy Prediction with MLIPs

This repository focuses on benchmarking Machine Learning Interatomic Potentials (MLIPs) for predicting cleavage energy using a DFT-calculated dataset. The project explores fine-tuning and transfer learning techniques to adapt pre-trained ML models for this specific task.

By embedding physics principles directly into the model's training process, such as ensuring symmetry invariance, energy extensivity, and surface area scaling, this work demonstrates how scientific principles can enhance ML accuracy and interpretability.

---

Project Overview

Objectives

• Benchmark MLIPs for cleavage energy prediction.
• Fine-tune and transfer pre-trained models for accurate results.
• Embed physics principles like rotational invariance and extensivity into ML training.
• Analyze model performance, outlier detection, and generalization.

Key Topics

• Benchmarking MLIPs: Compare predictive performance of MLIP models.
• DFT Dataset: Utilize a dataset of cleavage energies derived from Density Functional Theory.
• Fine-tuning and Transfer Learning: Adapt ML models for the specific cleavage energy prediction task.
• Physics-embedded Training: Ensure physical principles guide ML predictions.
• Scientific Insights: Demonstrate how physics principles improve ML generalization and accuracy.

---

Dataset

The dataset contains:

• Structural information (bulk and slab configurations).
• DFT-calculated cleavage energies.
• Additional features like surface area and Miller indices.

---

Methodology

1. Model Selection: Start with pre-trained MLIPs (e.g., GemNet-OC, MACE).
2. Fine-tuning: Adjust weights using the cleavage energy dataset.
3. Physics Embedding:
   • Ensure rotational and translational invariance.
   • Include surface area and energy extensivity.
4. Evaluation:
   • Benchmark models using MAE, MAPE, and parity plots.
   • Analyze outliers and model generalization.

---

Results

• Improved cleavage energy predictions with embedded physics.
• Identification of outliers and their impact on model performance.
• Benchmark results for MLIP models.

---

How to Use

1. Clone the Repository

git clone https://github.com/your-repo-name.git
cd your-repo-name

2. Install Dependencies

pip install -r requirements.txt

3. Run Fine-tuning

python main.py --config config.yml --mode train

4. Evaluate Results

python main.py --config config.yml --mode evaluate

---

Future Work

• Investigate outliers and challenging surfaces.
• Combine bulk and surface datasets for improved generalization.
• Develop new physics-embedded MLIP architectures.

---

Contributors

• Your Name

---

References

1. Schindler, P., et al., "Discovery of Stable Surfaces with Extreme Work Functions," Adv. Funct. Mater., 2024. DOI: 10.1002/adfm.202401764.
2. Open Catalyst Project Datasets (OC20, OC22).