HAPNet: Hierarchical Atomic Pathway Networks

Smart Hybrid Functionals via Hierarchical Atomic Pathway Networks

Installation

git clone https://github.com/SuthPhy2Ai/HAPNet.git
cd HAPNet
conda create -n hapnet python=3.10
conda activate hapnet
pip install -r requirements.txt

External Dependencies

Mat2vec pretrained embeddings are required but not included due to licensing. See src/mat2vec-master/README.md for download instructions.

The dielectric dataset (dbs/clean.db) must be built from the Materials Project. See dbs/README.md for instructions.

Project Structure

HAPNet/
├── models/              # GRU encoder + CLIP model
├── utils/               # Dataset, feature extraction, training utilities
├── scripts/             # Training scripts (CLIP + regression)
├── src/                 # Path representation, embeddings, auxiliary data
├── dbs/                 # ASE database (user-built, 7240 structures)
├── checkpoints/         # Saved model weights
└── conf.py              # Centralized hyperparameters

Training

Two-stage pipeline:

# Stage 1: CLIP contrastive pretraining
python scripts/train_clip.py

# Stage 2: Regression fine-tuning
python scripts/train_regression.py

Hyperparameters for CLIP are defined in scripts/train_clip.py. Regression hyperparameters are in conf.py.

Checkpoints are saved to checkpoints/.

Feature Representation

Each atom is encoded as a 384-dim vector:

Component	Dims	Source
Atomic embedding	92	One-hot atomic number
Mat2vec	200	Pretrained material embeddings
Symmetry	64	Point group features
Window distance	4	Local distance (Gaussian basis)
Radial structure	24	Radial distribution (Gaussian basis)

Atoms are ordered via a TSP-inspired nearest-neighbor path, then processed by a bidirectional GRU with attention.