README.md

MLIP-Machine Learning Interatomic Potential

1.Introduction

Machine-learning interatomic potentials (MLIP) bridge the gap between quantum-level accuracy and classical molecular-dynamics speed. Traditional force fields rely on fixed functional forms and hand-tuned parameters, limiting transferability. In contrast, MLIP learn the energy-force landscape directly from high-fidelity density-functional-theory data, capturing many-body and chemical effects without explicit equations. Modern frameworks embed rigorous physical priors—permutation, rotation and translation invariance, smoothness, locality—into expressive models such as equivariant graph neural networks, message-passing networks, Gaussian process regressors and deep neural descriptors. A typical workflow begins by sampling diverse atomic configurations, computing reference energies, forces and stresses, then training the model with loss terms that balance all three quantities. Active-learning loops iteratively enrich the dataset where prediction uncertainty is high, minimizing human intervention. Once trained, an MLIP delivers near-DFT accuracy for million-atom, nanosecond-scale simulations at a small fraction of the cost, enabling studies of crack propagation, phase transitions, ion diffusion and catalytic reactions that were previously intractable. As datasets grow and architectures mature, MLIP are poised to become standard tools for predictive, large-scale materials and molecular modeling.

2.Models Matrix

Supported Functions	CHGNet	MatterSim
Forward Prediction
Energy	✅	✅
Force	✅	✅
Stress	✅	✅
Magmom	✅	-
ML Capabilities · Training
Single-GPU	✅	✅
Distributed Train	✅	✅
Mixed Precision	-	-
Fine-tuning	✅	✅
Uncertainty / Active-Learning	-	-
Dynamic→Static	-	-
Compiler CINN	-	-
ML Capabilities · Predict
Distillation / Pruning	-	-
Standard inference	✅	✅
Distributed inference	-	-
Compiler CINN	-	-
Molecular Dynamic Interface
ASE	✅	✅
Dataset
MPtrj	✅	🚧
ML2DDB🌟	✅	-

Notice:🌟 represent originate research work published from paddlematerials toolkit