Benchmarking ML models for prediction of second-harmonic generation coefficients

The aim of the project is to use active learning to screen hypothetical compounds from OPTIMADE-enabled databases for those with high SHG coefficients for a given band gap. This procedure generated a DFT-computed SHG dataset, SHG-25, which was then used to benchmark a series of ML models across different classes. The results and methods of this benchmarking are found in this repository under ./benchmarks, along with a series of utilities in the shg-ml-benchmarks Python package under ./src.

This repository accompanies the preprint:

V. Trinquet, M. L. Evans, G-M.R. Rignanese, Accelerating the discovery of high-performance nonlinear optical materials using active learning and high-throughput screening (2025) arXiV:2504.01526 DOI: 10.48550/arXiv.2504.01526

The resulting dataset is archived on the Materials Cloud Archive:

V. Trinquet, M. L. Evans, G-M.R. Rignanese, Accelerating the discovery of high-performance nonlinear optical materials using active learning and high-throughput screening (2025) DOI: 10.24435/materialscloud:wk-qm

with associated OPTIMADE API access at https://optimade.materialscloud.org/archive/wk-qm/v1/info.