Accelerating Band Gap Prediction and High-Throughput Screening of Covalent Organic Frameworks Based on Transfer Learning

Abstract

Covalent organic frameworks (COFs) are a vast class of materials with nearly infinite structural possibilities, making it challenging to quickly identify COFs with specific properties, particularly their electronic properties. In this study, we apply transfer learning to overcome these limitations by fine-tuning a pretrained model (PMTransformer) on various COF data sets, enabling the rapid prediction of COF band gaps. Our model accurately predicts COF band gaps with a mean absolute error of 0.23 eV and R² of 0.89, outperforming the crystal graph convolutional neural network model. We validate the model’s predictions using density functional theory (DFT) calculations on a separate COF data set, confirming the consistency of predicted and DFT-calculated band gaps. By applying the model to a large COF database, we identify promising (sim) conductive COFs, demonstrating the model’s potential as an efficient screening tool for discovering COFs with optimized electronic properties for future applications in electronics and optics.