It is important that the hazardous properties of chemicals, including substitutes for prohibited or restricted products, are assessed at an early stage in product and process design. In this thesis, a new strategy for modeling quantified structure-property relationships is proposed that is based on multi-task deep learning and simultaneously predicts four flammability-related properties, including lower and upper flammability limits, autoignition point temperature and flash point temperature. A deep multitasking neural network (MDNN) was developed to automatically extract features from molecular structures and correlate multiple properties without computing molecular descriptors. The neural tree LSTM network is integrated into several neural feedforward networks. Two methods, joint training and alternative training, were both used to train the proposed MDNN, which was used to capture the relevant information and commonalities between multiple target characteristics. The detection of outliers and the determination of the scope have also been introduced into the assessment of deep learning models. Because the proposed MDNN used data more efficiently, the final model obtained is better than the multi-task partial least squares model in predicting flammability-related properties. The proposed framework for multi-task deep learning offers a promising tool for predicting multiple properties without calculating descriptors.