A Study on Distance Based Representation of Molecules for Statistical Learning

Statistical learning of molecular structure properties is gaining interests among researchers. These methods are faster compared to traditional QM based methods. In addition, the physical properties can be incorporated as feature sets and a properly trained model can predict the desired properties of a molecular system. For this, a number of machine learning regressors are used to predict molecular energies of Si − n (n = 1, 2, ⋯25) clusters, water, methane and ethane molecules. For the Si_n cluster, six out of eight regressors seem to predict the energies accurately. For other data sets, Decision Tree Regressor prediction resulted fairly good, in general, compared to others. However, through the addition of atomic charges as an extra feature improved the performance of other regressors, this did not cause any improvement for the Decision Tree regressor. Since calculating atomic charges in itself is an expensive task, we summarize that decision tree regressor is suitable for predicting molecular properties compared to other regressors tested here.