MACHINE LEARNING IN SCARCE DATA REGIME FOR DESIGN AND DISCOVERY OF MATERIALS

Abstract

In recent years, data-driven approaches based on machine learning have emerged as a promising method for rapid and efficient estimation of the structure-property-performance relationships, leading to the discovery of advanced materials. However, the cost and time required to obtain relevant data have limited application of these methods to only a few classes of materials where extensive property data are available. Moreover, the material property prediction poses its own unique set of challenges, in part, due to: 1) the complex non-linear response of materials in different space and time domains, 2) inherent variability in material in terms of composition and processing conditions from the atomic to the macroscopic scales, and 3) the need for accurate, rapid and less expensive predictive models for accelerated material discovery. This dissertation aims to develop three novel machine learning frameworks for constructing targeted learning frameworks and discovering novel materials when dealing with limited available data. The dissertation also highlights the future directions and challenges of such approaches. In the first approach, we develop data-driven methods to estimate the material properties under shock compression. A novel featurization approach combining synthetic and physical features was developed showing substantial improvements in the machine learning model performance. The effects of feature engineering, model choices, and uncertainty in the experimental data were investigated. In the second approach, we develop a novel joint embedding framework that enables transfer learning, with the target of locally optimizing the shock wave properties of nitrogen-rich molecules. This work is motivated by a need to overcome challenges associated with the translation of machine learning approaches to domains where there is a relative lack of domain-specific data. However, the properties studied in the second approach do not consider factors needed to assemble a complete material system. Therefore, in the third and final approach, we investigate material systems whose properties at system level are determined by various upstream design factors, such as the composition of raw materials, manufacturing variability, and considerations involved while assembling the system. We propose a stacked ensemble learning framework to make statistical inferences about the system properties.