A multi-task deep learning neural network for predicting flammability-related properties from molecular structures

It is significant that hazardous properties of chemicals including replacements for banned or restricted products are assessed at an early stage of product and process design. This work proposes a new strategy of modeling quantitate structure-property relationships based on multi-task deep learning for simultaneously predicting four flammability-related properties including lower and upper flammable limits, auto-ignition point temperature and flash point temperature. A multi-task deep neural network (MDNN) has been developed to extract molecular features automatically and correlate multiple properties integrating a Tree-LSTM neural network with multiple feedforward neural networks. Molecular features are encoded in molecular tree graphs, calculated and extracted without manual actions of the user or preliminary molecular descriptor calculation. Two methods, joint training and alternative training, were both employed to train the proposed MDNN, which could capture the relevant information and commonality among multiple target properties. The outlier detection and determination of applicability domain were also introduced into the evaluation of deep learning models. Since the proposed MDNN utilized data more efficiently, the finally obtained model performs better than the multi-task partial least squares model on predicting the flammability-related properties. The proposed framework of multi-task deep learning provides a promising tool to predict multiple properties without calculating descriptors.

Automated summary

This study introduces a novel deep learning approach for predicting multiple flammability-related properties simultaneously with higher accuracy compared to traditional methods.