Mixed-Variable Adaptive Design Framework with Bayesian Optimization for Multi-Property Optimization of Anisotropic Nanodielectric Materials using “on-demand” Simulations

Design and development of novel polymer nanodielectrics still remains a challenge due to high number of parameters and properties from different domains involved with the process. Specifically, the mixed-variable multi-scale design space formed by interface choices, and microstructural variations creates a bottleneck in the design optimization. We propose a physics-based adaptive nanodielectrics design framework for multi-property optimization. The framework adopts a descriptor based anisotropic microstructure design technique that controls the non-spherical filler design variables, namely volume fraction, orientation, dispersion, and aspect ratio to generate microstructures, and incorporates attractive extrinsic and intrinsic interfaces as supportive mechanisms into the design. A novel and interpretable machine learning model, Latent Variable Gaussian Process (LVGP), is trained to capture and understand the underlying relationship between mixed-variable design space, interface (qualitative) and microstructural (quantitative) variables, and the properties. Finally, a multi-property Bayesian optimization is performed with the help of “on demand” finite element simulations to find the Pareto Front between mechanical and dielectric properties of anisotropic nanodielectric materials, while providing interpretability about the design process through global sensitivity analysis.