Quantitative properties through semantic learning

Material properties have almost exclusively been studied via natural science-based approaches. However, Natural Language Processing (NLP) poses fertile new ground for research where linguistics plays a role in the characterization, design, and discovery of energetic materials. This paper will test if statistical relationships exist between the language used to discuss energetic materials and their fundamental physicochemical properties. Statistical distributions are developed for each separately – one based on language and the other based on known property data – and metrics are developed to compare the two distributions. Word embedding models based on unsupervised Skip-Gram techniques are used to generate n-dimensional word vectors, in an attempt to capture the semantic content in the text. A softmax function is then used to produce a statistical distribution. The chemical space embedding is based on the chemical and physical properties of energetic compounds and the statistical nature arises from the variabilities in measured values of properties as well as differences in the chemical composition of ostensibly the same materials. Impact sensitivity (h₅₀), detonation velocity (D), and detonation pressure (P) are considered for the material property. Statistical comparisons between the word space and the chemical space distributions are performed using limited sets of reference energetics: 5 for h₅₀ and 14 for D and P. A surprising and remarkable degree of statistical equivalence is found, in some cases showing >90% confidence levels. This work posits a new means for using automated machine-assisted approaches to learn from technical documents and facilitate the search and discovery of new materials.