Deep learning for chemical compound stability prediction
This paper explores the idea of using deep neural networks with various architectures and a novel initialization method, to solve a critical topic in the field of materials science. Understanding the relationship between the composition and the property of materials is essential for accelerating the course of materials discovery. Data driven approaches using advanced machine learning to derive knowledge from that of existing compounds, and/or from simulations of nonexisting ones, have only started to play a crucial role. We demonstrate an application with a large-scale data set containing 300K organic and inorganic compounds. Deep multilayer perceptrons are used to capture nonlinear mappings between chemical composition and compound stability characterized by a continuous value, known as the formation energy. It is surprising to see that input features as raw and sparse as the compositional fractions of elements can lead to a remarkably accurate modeling of a far-fetched regression prediction. The performance is shown to be outperforming state-of-the-art predictions by as much as 54%.