A Data Fusion Approach for Quantitative Property Prediction Using Natural Language Processing

Machine learning is increasingly used in materials science to predict molecular properties due to its ability to be trained directly using source data. Recent developments in natural language processing, particularly in the biosciences and materials sciences, have shown value in automating the process of reading academic literature. This is because natural language in academic papers can encode significant knowledge about materials and their properties. Quantifying such knowledge could enable new ways of exploiting the insights encoded in unstructured text for quantitative tasks such as property prediction. This work presents a novel data fusion approach to develop features that fuse quantitative material properties with semantic information from unstructured text to predict the detonation velocity of energetic materials. The effectiveness of the proposed method is demonstrated through experiments on a diverse set of energetic compounds.