Scientific AI for automated engineering design under uncertainty

We discuss a new tool that uses physics-informed AI to efficiently solve complex inverse problems under uncertainty, and enables the automated generation of high-fidelity surrogates for physically-relevant response surfaces of interest. By applying all constraining information as transforms, we simplify optimization problems, ensure all candidate solutions are valid, and facilitate automation. By incorporating certification and validation requirements as constraints, we have seen a dramatic increase in the efficiency and automation of the solution of engineering design problems. We recently utilized this approach to automate Rietveld refinement, where we were able to reduce the time-to-solution for textures from months to minutes. We have also applied our framework to the real-time calibration of superconducting quantum hardware and quantum sensors, as well as the real-time optimal design of finite element simulations of printed parts, process simulations, and mixture models for additive manufacturing. All applications of this framework have demonstrated similar orders-of-magnitude reductions in the time-to-solution without loss of solution fidelity.