Active Learning for Machine-Learned Interatomic Potentials

This repository benchmarks active learning (AL) strategies for training machine-learned interatomic potentials (MLIPs) using data from the Materials Project and OQMD. Four AL strategies—Random, Uncertainty, Diversity, and Hybrid— are evaluated across C, Si, Fe, and Ti–O systems.

Key Result:
Diversity sampling gives the strongest performance, achieving a 10.9% improvement (p = 0.008) on the Ti–O system.

---

Features

• Multi-system benchmark (C, Si, Fe, Ti–O)
• Query-by-Committee uncertainty estimation
• Diversity-based sampling via k-means
• Ensemble NN model (17 → 128 → 128 → 1)
• Reproducible experiments (5 seeds per strategy)
• Runs on Google Colab (< 8 GB RAM)

---

Quick Start

Install

git clone https://github.com/vchoyal/25_MLIP_for_Constrained_enviorments.git
cd mlip-active-learning
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Repository Structure

1. References/ All references used in the paper
2. ML Model/ Ensemble NN + AL strategies
3. Results.png/ All figures
4. Report/ Full Docs & PDF of the paper