Development of Convolutional Neural Network-Based Models for Efficient and Reliable Flashpoint Prediction

Abstract

The initial step in quantitative structure–property relationship (QSPR) modeling typically involves selecting relevant descriptors for predicting flammability characteristics. However, traditional selection methods face challenges and inefficiencies when dealing with large pools of molecular descriptors. To address these issues, this study proposes a convolutional neural network (CNN)-based method to automatically identify relevant descriptors, eliminating the need for preselection. This method is applied to predict the flashpoints of pure compounds and mixtures using two different sets of molecular descriptors. Bayesian Optimization is employed to tune the hyperparameters of the CNN models through 5-fold cross-validation. The results demonstrate that the developed CNN models achieve high accuracy, with R² values of 0.9723 for pure compounds and 0.9939/0.9951 for mixtures in the test data set, outperforming previously established models. Notably, the CNN models maintain high accuracy even when the number of valid molecular descriptors is increased from 125 to 861. Overall, the proposed approach leads to robust performance and accelerates QSPR model development through automated molecular descriptor selection.