Significantly Improving the Prediction of Molecular Atomization Energies by an Ensemble of Machine Learning Algorithms and Rescanning Input Space: A Stacked Generalization Approach

Machine learning holds the promise of learning the energy functional via examples, bypassing the need to solve complicated quantum–chemical equations and realizing efficient computing of molecular electronic properties. However, a single machine learning model may reach an upper limit of prediction accuracy even with optimal parameters. Unlike the “bagging” and “boosting” approaches, which can only combine the machine learning algorithms of a same type, the stacking approach can combine several distinct types of algorithms through a meta-machine learning model to maximize the generalization accuracy. Here, we present a stacked generalization approach for predicting the ground state molecular atomization energies. The results suggest prediction error from stacked generalization frameworks are significantly reduced by 38%, compared to the best level 0 individual algorithm that construct the stacked generalization framework. Furthermore, compared to conventional stacked generalization framework, it shows that rescanning the original input space by level 1 meta model can extract more information, which can further improve the prediction accuracy.