SUMMARY

Materials design is a process of identifying material compositions and microstructures to achieve desirable properties. The challenge is the efficiency of searching solutions in the high-dimensional design space, where the evaluations of solutions usually involve costly experiments or simulations. Bayesian optimization (BO) is a new surrogate-based global optimization scheme which can reduce the number of observations required to obtain a desirable solution. In BO, the unknown objective function is usually approximated by a Gaussian process surrogate. The surrogate helps predict the mean and variance for each unsampled candidate. New samples are used to update the surrogate in the active learning process. To improve the efficiency of BO for multi-attribute materials design with the simultaneous consideration of multiple properties, in this research, a new multi-attribute BO method is proposed to efficiently find the desirable solution based on user preference. A multi-fidelity BO method is also proposed to improve the cost-effectiveness of searching with the consideration of observation costs, where low-fidelity and high-fidelity samples are effectively combined. The proposed new BO methods will be applied to design different materials including polymers and ceramics. A universal polymer descriptor is further proposed to enable systematic search in the discrete design space of compositions with repeating units. Full-atomistic and coarse-grained molecular dynamics simulations will be applied to demonstrate simulation-based materials design.