(Upcoming) Constructing Surrogate Models with Constraints for Additive Manufacturing.
Date:
Machine learning models for additive manufacturing often result in unreliable predictions which violate physical or manufacturing constraints and require significant post-hoc correction. In this talk, we present an alternate approach to machine learning to strictly enforce constraints while avoiding pitfalls of traditional neural-network-based methods. We develop Newton-type methods for engineering scale problems, addressing the need for reliable constrained optimization in additive manufacturing. By discretizing the optimization statement as a quadratic program and solving a second-order KKT system via interior point (barrier) methods, we enforce constraints by construction rather than through penalization. Our methods are structure-preserving, enforcing any underlying governing laws, geometric constraints, or other application-dependent requirements. We demonstrate this approach on a range of test problems and identify further applications for these methods to find reliable, constraint-satisfying predictions in additive manufacturing