Neural Networks for Structure-Informed Prediction of Formation Energy (employed in SIPFENN)

pySIPFENN Documentation: pysipfenn.org

pySIPFENN GitHub: git.pysipfenn.org

Original SIPFENN Paper: 10.1016/j.commatsci.2022.111254

Network Changelog:

V 0.10 - All models moved to the open ONNX format for improved interchangeability; NN30 neural network similar to NN20 but accepting the new KS2022 feature vector; Python code migrated to public GitHub repository.

V 0.9 - Python code updated to the release version; paper published

V 0.8 - Python code (beta) to run models included

V 0.7 - Original upload of development models 

Selected works with SIPFENN alongside DFT and experiments:

- 10.1016/j.actamat.2021.117448

- 10.1038/s41598-021-03578-0

SIPFENN Abstract (original publication, 2021):

In recent years, numerous studies have employed machine learning (ML) techniques to enable orders of magnitude faster high-throughput materials discovery by augmentation of existing methods or as standalone tools. In this paper, we introduce a new neural network-based tool for the prediction of formation energies based on elemental and structural features of Voronoi-tessellated materials. We provide a self-contained overview of the ML techniques used. Of particular importance is the connection between the ML and the true material-property relationship, how to improve the generalization accuracy by reducing overfitting, and how new data can be incorporated into the model to tune it to a specific material system.
    
    In the course of this work, over 30 novel neural network architectures were designed and tested. This lead to three final models optimized for (1) highest test accuracy on the Open Quantum Materials Database (OQMD), (2) performance in the discovery of new materials, and (3) performance at a low computational cost. On a test set of 21,800 compounds randomly selected from OQMD, they achieve mean average error (MAE) of 28, 40, and 42 meV/atom respectively. The second model provides better predictions on materials far from ones reported in OQMD, while the third reduces the computational cost by a factor of 8.
    
    We collect our results in a new open-source tool called SIPFENN (Structure-Informed Prediction of Formation Energy using Neural Networks). SIPFENN not only improves the accuracy beyond existing models but also ships in a ready-to-use form with pre-trained neural networks and a user interface. 

Contacts:

- Adam Krajewski: ak@psu.edu

- Prof. Zi-Kui Liu: zxl15@psu.edu