| Abstract Scope |
We discuss the use of uncertainty quantification and scientific machine learning to efficiently perform engineering design. We present a tool that enables an optimizer’s search space to be restricted to only valid solutions through the use of constraining information (physical, statistical, data, uncertainty, expert knowledge, and assumptions) as “coordinate” transforms in a scientific learning problem. The application of all design requirements as constraints facilitates automation and greatly simplifies the design problem to be solved. The use of a quality metric, such as expected lifetime, as an objective automates solution improvement even in the presence of unknowns. This approach has been used to automate the generation of high-fidelity surrogates for physically-relevant response surfaces of interest. We have applied this approach to the real-time engineering design of quantum sensors and the optimal design of mixture models for additive manufacturing, and have seen three-to-four orders-of-magnitude reductions in the time-to-solution without loss of fidelity. |