ReciNet: Reciprocal Space-Aware Long-Range Modeling for Crystalline Property Prediction

Official PyTorch implementation of ReciNet: Reciprocal Space-Aware Long-Range Modeling for Crystalline Property Prediction (TMLR 2026).

Citation

@article{nie2025recinet,
  title={ReciNet: Reciprocal Space-Aware Long-Range Modeling for Crystalline Property Prediction},
  author={Nie, Jianan and Xiao, Peiyao and Ji, Kaiyi and Gao, Peng},
  journal={arXiv preprint arXiv:2502.02748},
  year={2025}
}

Overview

ReciNet is the reciprocal space-based geometry network, a hybrid neural architecture for crystal property prediction that jointly models short-range chemical bonding and long-range periodic interactions. Crystals exhibit infinite periodic arrangements of atoms, requiring methods that capture both local and global information. ReciNet addresses this by leveraging reciprocal space, the natural domain for crystals. The ReciprocalBlock is plug-and-play, and the module is architecture-agnostic.

Installation

We recommend Python 3.10 and CUDA 11.6.

# Create the environment.
conda create --name recinet python=3.10
conda activate recinet

conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia # or higher version if you want
conda install pyg -c pyg

python -m pip install -e .

Datasets

All datasets are downloaded automatically through the JARVIS toolkit on the first run, there is no need to manually download the JARVIS or Materials Project datasets.

JARVIS-DFT

We use following train/val/test splits and evaluate on five crystal properties:

Property	Train	Val	Test	Metric
Formation energy	44,578	5,572	5,572	MAE (meV/atom)
Bandgap (OPT)	44,578	5,572	5,572	MAE (eV)
Total energy	44,578	5,572	5,572	MAE (meV/atom)
Ehull	44,296	5,537	5,537	MAE (meV)
Bandgap (MBJ)	14,537	1,817	1,817	MAE (eV)

The Materials Project

We use following train/val/test splits for formation energy, band gap, bulk modulus, and shear modulus. The bulk/shear modulus splits are available at figshare.

Property	Train	Val	Test	Metric
Formation energy	60,000	5,000	4,239	MAE (meV/atom)
Band gap	60,000	5,000	4,239	MAE (eV)
Bulk modulus	4,664	393	393	MAE (log(GPa))
Shear modulus	4,664	393	393	MAE (log(GPa))

MatBench

We evaluate on two MatBench tasks :

• e_form (132,752 crystals)
• jdft2d (636 2D crystals)

Training

Configure the run

Edit configs/recinet.yaml to select the dataset and target property:

dataset: dft_3d                          
target: formation_energy_peratom         

Dataset	Available targets
dft_3d (JARVIS)	formation_energy_peratom, mbj_bandgap, optb88vdw_bandgap, optb88vdw_total_energy, ehull
megnet (MP)	e_form, gap pbe, bulk modulus, shear modulus

Train

python main.py \
    --config configs/recinet.yaml \
    --output_dir runs/jarvis_formation_energy

Training on a Custom Dataset

ReciNet supports custom datasets in the JARVIS Leaderboard format. Prepare a directory containing:

• dataset_info.json — metadata describing the dataset
• id_prop.csv — crystal IDs and target values, listed in train → val → test order
• Crystal structure files readable by JARVIS tools

Then run:

python main.py \
    --config configs/recinet.yaml \
    --output_dir runs/custom_run \
    --data_root /path/to/custom_dataset

and set the corresponding target in recinet.yaml:

dataset: dft_3d
target: <your_target_name>

Use jarvis_populate_data.py as a reference for generating compatible datasets. Predefining the train/val/test split (in id_prop.csv) is required.

Benchmark Results

For benchmark results, see the paper for full tables results.

Acknowledgments

This codebase builds upon ComFormer and PotNet. We thank the authors for releasing their code. We also thank the maintainers of JARVIS-tools, Materials Project, and MatBench for providing the datasets and benchmarks.

License

This project is released under the MIT License — see LICENSE for details.