README.md

Property Prediction

1.Introduction

Property Prediction (PP) targets rapid, first-principles-level estimation of key crystalline properties—formation energy, band gap, elastic moduli, ionic conductivity, and more—without performing new density-functional-theory calculations. The workflow mirrors modern ML interatomic-potential pipelines but shifts the label space from forces to scalar and tensor observables. Starting from crystal structure files (CIF), an automated converter builds atom–bond graphs enriched with chemical descriptors and symmetry-aware positional encodings. Equivariant graph neural networks, or transformer-based variants, are then trained on tens of thousands of reference entries. By collapsing months of high-throughput DFT time into minutes of GPU inference, PP empowers data-driven discovery of semiconductors, catalysts and functional

2.Models Matrix

Supported Functions	MEGNet	Comfomer	GemNet	DimeNet++
Forward Prediction · Materials Properties
Formation energy	✅	✅	🚧	✅
Band gap	✅	✅	🚧	✅
Bulk modulus	✅	✅	🚧	✅
Shear modulus	✅	✅	🚧	✅
Young’s modulus	✅	✅	🚧	✅
Adsorption energy	🚧	🚧	🚧	🚧
Electron density	—	—	—	—
ML Capabilities · Training
Single-GPU	✅	✅	🚧	✅
Distributed training	✅	✅	🚧	✅
Mixed precision (AMP)	—	—	—	—
Fine-tuning	✅	✅	🚧	✅
Uncertainty / Active Learning	—	—	—	—
Dynamic→Static graphs	—	—	—	—
Compiler (CINN) opt.	—	—	—	—
ML Capabilities · Predict
Distillation / Pruning	—	—	—	—
Standard inference	✅	✅	🚧	✅
Distributed inference	—	—	—	—
Compiler-level inference	—	—	—	—
Datasets
Materials Project
MP2024	✅	✅	—	—
MP2020	✅	✅	—	—
MP2018	✅	✅	🚧	—
JARVIS
dft_2d	✅	✅	—	✅
dft_3d	✅	✅	—	—
Alexandria
pbe_2d	✅	✅	🚧	—
ML2DDB🌟	✅	✅	✅	✅

Notice:🌟 represent originate research work published from paddlematerials toolkit