Machine-Learning-Based Constrained Optimization of a Test Coupon Launch Using Inverse Modeling

This paper demonstrates the forward modeling and inverse design of a test coupon launch structure used in the board measurement practice known as the delta-L method. An inverse model is trained to synthesize a launch design to exhibit a desired electrical performance and to be physically realizable. A forward model is constructed and used to evaluate the electrical performance of the designs synthesized by the inverse model during training. The training of this inverse model is treated as a convex optimization with constraints on the synthesized designs. These constraints inspire a novel implementation of constraint loss by a pair of everywhere-differentiable barrier functions. The finished inverse model is applied to a swift multi-criteria design optimization and the forward model is used to perform uncertainty analysis about the synthesized design. Considerations for further applications and improvement of the procedure are discussed.