Physics-Informed Descriptors Base Machine Learning Model for High-Precision (Molecular Property Dipole Moment($\mu$)) Prediction on QM9

Abstract Dipole Moment(µ) is important property which helps to understand how molecules behave and support to design the new material. Therefore, predicting a molecule’s dipole moment property has high importance. In this paper, we introduce a MLP Model based on physics-informed descriptors which help to predict dipole moment µ on QM9 dataset with a test MAE = 0.0231Debye. Our selection of descriptor (atom counts, bond lengths, bond angles, and approximate partial charge) helps this model to learn more and produces highly precise results without using 3D methods. Above mentioned results show this model is fast, precise and accurate alternative to existing 3D − GNN models which are complex and required complete 3D geometry. This model performs better then GNN models such as ALIGNN(MAE = 0.0248Debye), SphereNet(MAE = 0.025Debye), DimNet + + (MAE = 0.030Debye), and SchNet(MAE = 0.033Debye). Performance of our model is near to state of the art model TorchMD − Net(MAE = 0.011Deybe) and PainNN(MAE = 0.012Deybe) which make this model computationally efficient and alternative to 3D − GNN models.