Software
Our group develops open-source computational tools for materials discovery and design. Below is a list of software packages developed and maintained by our group.
libfp — Crystal Fingerprint Library
A high-performance crystal fingerprint library with a C core and Python interface. libfp characterizes local atomic environments using Gaussian Overlap Matrix (GOM) eigenvalues and provides precise quantification of structural similarities. It features analytical gradients with respect to atomic positions and strain, enabling efficient use in optimization and molecular dynamics. libfp is integrated into several of our other software packages, including EOSnet, ReformPy, and PALLAS.
GitHub | J. Chem. Phys. 144, 034203 (2016)
EOSnet — Crystal Graph Neural Network
EOSnet is a machine learning framework based on crystal graph neural networks that integrates crystal fingerprints (Embedded Overlap Structures) to account for many-body interactions. This approach enables accurate prediction of a wide range of material properties, including formation energy, band gap, and elastic moduli, by providing a more comprehensive representation of complex structural information.
J. Phys. Chem. Lett. 16, 717 (2025) | arXiv:2411.02579
ReformPy — Fingerprint-Oriented Structural Optimization
ReformPy (Rational Exploration of Fingerprint-Oriented Relaxation Methodology) is a Python package that enhances structural optimization by integrating fingerprint-space biasing into the potential energy surface. By guiding optimization toward high-symmetry, low-energy structures through the intrinsic symmetry of atomic configurations, ReformPy significantly improves the probability of discovering energetically favorable structures in crystal structure prediction.
GitHub | J. Phys. Chem. Lett. 15, 3185 (2024)
PALLAS — Phase Transition Pathway Sampling
PALLAS is a computational method for predicting phase transition pathways between crystal structures without prior knowledge of the transition mechanism. It combines the solid-state dimer method, crystal fingerprints, swarm-intelligence structure searching, and graph theory to automatically identify low-energy transition pathways and elucidate the mechanisms driving phase transformations.
J. Phys. Chem. Lett. 10, 5019 (2019)
SigML — Machine-Learning-Accelerated DMFT
SigML uses graph neural networks to predict DMFT self-energies, providing a warm-start for dynamical mean-field theory calculations that reduces the number of expensive DMFT iterations by 2–4x. Combined with machine-learned interatomic potentials trained on DMFT-level forces, this framework enables large-scale simulations of strongly correlated materials at extreme conditions, such as iron at Earth's core pressures and temperatures.