Multi-Scale Optimization of Materials Using Machine Learning

Optimizing materials is crucial for advancing modern technologies, requiring analysis from atomic to microstructural scales. Traditional methods, including trial-and-error experiments and standard computational approaches, often struggle due to high computational costs and uncertainties in multi-scale optimization. This dissertation presents a machine learning-integrated framework to streamline material optimization across these scales.

At the atomic and molecular levels, the framework combines high-throughput molecular dynamics simulations with ML models to rapidly identify high-performance polymers with superior thermal conductivities. For larger scales, it employs gradient-free optimization and JAX-based differentiable finite element methods to efficiently solve inverse problems, determining heterogeneous material properties from indirect measurements. This strategy overcomes the limitations of traditional numerical solvers, which often lack gradient information.

To bridge multiple scales, the research incorporates uncertainty quantification, modeling, and optimization, enhancing the accuracy and reliability of the process. This integrated approach significantly improves material discovery and optimization, addressing key challenges in computational efficiency and predictive accuracy.