Prediction of viscosity of biomass-based molecules using atom modules and modularity as descriptors in neural network models

Abstract

This paper presents a new artificial neural network (ANN) model for viscosity prediction at different temperatures using atom modules identified on the molecular graphs and the corresponding modularity values from the application of community detection algorithms. The model is compared to an ANN model using functional group counts and modularity as inputs. A dataset covering 430 molecules and temperature range between 253 and 463 K was employed to develop the models. The application of community detection resulted into 177 modules, which captured a wide range of molecular fragments and features, including polyfunctional molecules. The model was then reduced to 158 modules to improve performance considering statistical significance, correlation among inputs, relative importance in the ANN model and identification of outliers. The final modules model demonstrated improved performance on the training and test sets compared to the functional groups ANN model and showed excellent predictions for all compound classes. The model was then used to predict the viscosity of molecules potentially derived from biomass used in existing models, which served for further model validation. The final modules model showed improved performance based on the root mean squared error in comparison to existing models. This demonstrates the advantage of employing modularity and the modules identified by the community detection algorithm as molecular descriptors. The approach extends the potential of the application of community detection to molecular graphs and modularity to capture structure-property relationships for predicting molecular properties.