Assessing Graph‐based Deep Learning Models for Predicting Flash Point

Abstract

Flash points of organic molecules play an important role in preventing flammability hazards and large databases of measured values exist, although millions of compounds remain unmeasured. To rapidly extend existing data to new compounds many researchers have used quantitative structure-property relationship (QSPR) analysis to effectively predict flash points. In recent years graph-based deep learning (GBDL) has emerged as a powerful alternative method to traditional QSPR. In this paper, GBDL models were implemented in predicting flash point for the first time. We assessed the performance of two GBDL models, messagepassing neural network (MPNN) and graph convolutional neural network (GCNN), by comparing against 12 previous QSPR studies using more traditional methods. Our result shows that MPNN both outperforms GCNN and yields slightly worse but comparable performance with previous QSPR studies. The average R ! and Mean Absolute Error (MAE) scores of MPNN are, respectively, 2.3% lower and 2.0K higher than previous comparable studies. To further explore GBDL models, we collected the largest flash point dataset to date, which contains 10575 unique molecules. The optimized MPNN gives a test data R ! of 0.803 and MAE of 17.8 K on the complete dataset. We also extracted 5 datasets from our integrated dataset based on molecular types (acids, organometallics, organogermaniums, organosilicons, and organotins) and explore the quality of the model in these classes.

Design Green Chemicals by Predicting Vaporization Properties Using Explainable Graph Attention Networks

“…The best model from training against the DIPPR database showed the MAEs of 6.4-7.1 K for training, validation, and test sets. These errors are comparable to the typical experimental errors of FP measurements using standard methods (5.0-8.0 K) 57,81,82 . On the other hand, the model for TM showed a higher test set MAE (21.7 K) than other properties, but it was not used for designing green chemicals.…”

“…Meanwhile, TC values of 7,362 molecules were collected from the same data source. FP of molecules were gathered from the Design Institute for Physical Properties (DIPPR) database 82 and other literature 57 . We removed the ambiguous FPs which are significantly different among multiple literature sources, leading to a total of 3,282 data points [46][47][48]50,[52][53][54]56,82 , 708 of which are from the DIPPR database.…”

Opportunities and Challenges for Machine Learning in Materials Science

“…For example, the deep convolutional individual residual network (40) was shown to achieve better performance from a long list of initial features than traditional methods such as RFDTs and ridge regression. Some graph-based deep learning CV: cross validation methods, which build feature maps from a very basic initial feature list, have shown comparable or better performance in organic molecule studies than human-crafted traditional features in QSAR/QSPR comparisons (e.g., message passing NN frameworks) (146,174). Similarly, for inorganic materials, the graph-based MatErials Graph Network (142), SchNet (147) and SchNetPack (148), and crystal-graph CNN (149,150) have shown performance comparable to or better than non-deep-learning approaches.…”

Neural Network Prediction Model of Cocrystal Melting Temperature Based on Molecular Descriptors and Graphs

“…Given node characteristics, it generates a node representation for each node that could be used for downstream tasks. It is a general summary of several GNN models selected by Gilmer et al through the messaging mechanism. ,− …”