Out-of-sample Transition Temperature Predictions via Physics-informed Transfer Learning

Transition temperatures, such as melting, boiling, and decomposition temperatures are important physical properties for material synthesis and scale-up. For energetic materials, safety and suitability constraints make accurate estimation of these properties especially important. First-principles calculations for estimating transition temperatures are computationally expensive and often require sample-specific knowledge, such as crystal structure. For energetics, this sample-specific knowledge is time-consuming to obtain. While machine learning (ML) models can bypass many of the limitations of first-principles, they require large datasets which are lacking for energetics. In order to overcome this challenge, we develop a physics-informed transfer learning approaches by exploiting known chemical information contained in the graph structure of the molecule and relationships between predicted physical properties. First, we train a message passing neural network on large non-specific datasets to predict transition temperatures of out-of-sample test sets by weighting the data in the training dataset using multiple similarity metrics. Finally, we leverage accurate ML chemical property models to improve less accurate chemical property models for related physical properties. Results are compared against experiment and other modeling approaches.