Machine Learning-Assisted High-Throughput Semi-empirical Search of OFET Molecular Materials

TL;DR

This paper introduces a machine learning-guided search method combining DFST and LightGBM to efficiently explore and identify promising OFET molecules from a vast chemical space, validated by DFT calculations.

Contribution

It presents a novel integration of DFST and LightGBM for high-throughput molecular screening in OFET research, achieving efficient exploration of millions of molecules with high validity.

Findings

Generated over 2.8 million molecules with 100% chemical validity.

Screened 184 promising molecules for OFET applications.

Demonstrated high efficiency and accuracy of the ML-guided search process.

Abstract

Machine learning has been widely verified and applied in chemoinformatics, and have achieved outstanding results in the prediction, modification, and optimization of luminescence, magnetism, and electrode materials. Here, we propose a deepth first search traversal (DFST) approach combined with lightGBM machine learning model to search the classic Organic field-effect transistor (OFET) functional molecules chemical space, which is simple but effective. Totally 2820588 molecules of different structure within two certain types of skeletons are generated successfully, which shows the searching efficiency of the DFST strategy. With the simplified molecular-input line-entry system (SMILES) utilized, the generation of alphanumeric strings that describe molecules directly tackle the inverse design problem, for the generation set has 100% chemical validity. Light Gradient Boosting Machine (LightGBM) model's intrinsic Distributed and efficient features enables much faster training process and higher training efficiency, which means better model performance with less amount of data. 184 out of 2.8 million molecules are finally screened out with density functional theory (DFT) calculation carried out to verify the accuracy of the prediction.