Multi-output ensemble deep learning: A framework for simultaneous prediction of multiple electrode material properties

The development of new electrode materials plays an important role in enhancing the performance of batteries. Machine learning can provide powerful support for discovering, developing and designing new materials. In this paper, an accurate and extensible multi-output ensemble deep learning (MOEDL) framework is constructed for simultaneous prediction of multiple material properties. The base models of ensemble learning are based on the deep neural network (DNN), and Bayesian optimization (BO), attention mechanism (AM) and deep belief network (DBN) are utilized to solve the shortcomings of the DNN model. Root mean square prop (RMSprop) algorithm and Monte Carlo (MC) method are applied to improve the reliability and accuracy of the model. And, the ridge regression model is utilized to integrate the results of the three base models with avoiding the multicollinearity problem. The feature set extracted from the Materials Project database is used to verify the accuracy, effectiveness and robustness of the model based on the framework. Compared with the calculation results of DFT, the Pearson correlation coefficient (PCC) and coefficient of determination (R²) of the properties output at the same time reach above 0.97 and 0.93 respectively for 10 types of ion batteries. This work could help accelerate the discovery and design of materials. In addition, based on the previously proposed CHAIN architecture, the constructed framework is not only applicable to the research of material development, but also can be extended to the design, management and control within the battery full lifespan.