High-Throughput Surface Cleavage Energy Prediction Using MLIPs

ML4MSD Final Project - Northeastern University
Author: Ardavan Mehdizadeh
Advisor: Prof. Peter Schindler (D2R2 Research Group)
Date: December 2025

Overview

This project develops a high-throughput pipeline for predicting metallic surface cleavage energies using Machine Learning Interatomic Potentials (MLIPs). Using MACE-OMAT-0 with GPU acceleration, we achieved ~10,000× speedup over traditional DFT calculations.

Dataset

• Bulk Materials: 22,441 metallic compounds from Materials Project
• Surface Structures: 778,105 surfaces (Miller indices ≤ 2)
• Valid Predictions: 672,806 converged calculations
• Computational Time: 4.5 hours on 4× NVIDIA L40S GPUs

Key Results

1. Bulk-Surface Stability: No correlation (r = -0.15) between bulk stability and surface cleavage energy
2. Method Comparison: Strong correlation (R = 0.93) between single-point and relaxed calculations
3. Crystal Systems: Cubic systems show 24% higher cleavage energies than triclinic

Repository Structure

01_bulk_selection/          # Materials Project query
02_surface_generation/      # Surface slab generation
03_cleavage_computation/    # MLIP cleavage energy calculations
04_analysis/                # Data analysis and visualization

Installation

conda create -n mlip_surface python=3.12
conda activate mlip_surface
pip install -r requirements.txt

Usage

1. Query Bulk Materials

cd 01_bulk_selection
python query_materials_project.py

2. Generate Surfaces

from unique_slabs_generator import UniqueSlabsGenerator
gen = UniqueSlabsGenerator(bulk_structure, miller_index=[1,1,1])
slabs = gen.get_unique_slabs()

3. Compute Cleavage Energies

cd 03_cleavage_computation
python production_cleavage_energy_fast.py --start 0 --end 1

4. Analyze Results

cd 04_analysis
python aggregate_results.py
python analyze_singlepoint_vs_relaxed.py

Methodology

Cleavage Energy Formula:

CE = (E_slab - N × E_bulk) / (2 × Area)

MLIP Configuration:

• Model: MACE-OMAT-0 (medium)
• Optimizer: FIRE
• Convergence: fmax = 0.4 eV/Å, max_steps = 500
• GPU: TorchSim + cuEquivariance (18× speedup)

Dependencies

• PyTorch 2.0+ (CUDA)
• MACE-torch 0.3.4+
• TorchSim 0.4.0 with cuEquivariance 0.7.0
• Pymatgen, ASE, mp-api

See requirements.txt for complete list.

Performance

Metric	Value
Time per surface	0.021 seconds
Convergence rate	86.8%
Speedup vs DFT	~10,000×

Citation

Surface Generation: Prof. Peter Schindler, D2R2 Research Group

Related Work: Mehdizadeh & Schindler (2025). Surface stability modeling with universal machine learning interatomic potentials. AI for Science, 1, 025002.

License

MIT License - See LICENSE file for details