Machine Learning Surrogates for Optimal 2D Spatial Packaging of Interconnected Systems with Physics Interactions (SPI2)

This work proposes new methods and applications of neural networks in automated design optimization for spatial packaging of interconnected systems with physics interactions (SPI2). The experiments are demonstrated on 2D electrothermal systems with heat-generating components connected via fluid-based cooling pipes. We seek to minimize the bounding box area of the system, while accounting for multiple physics and geometric constraints. We train a multilayer perceptron (MLP) on data generated through state-of-the-art physics-based optimization models, constructed with a finite element method (FEM)-based 2D steady state thermal conduction model and a lumped parameter pipe flow model. With the physics-based optimal designs as the ground truth, the model is tasked with mapping various configuration data, including design starting points, and boundary conditions of the physics model to the optimal system layout. Through training the model for this task, the network effectively acts as a surrogate for both physics-driven simulations and numerical optimization. Capable of real-time design synthesis, the MLP exhibits good generalization performance by generating designs that closely match those attained with the iterative physics-based optimization techniques. Relative to the ground truth physics-based designs, our trained surrogate achieves ∼ 91% accuracy while performing the same design tasks orders of magnitude faster. Further, we impose loss functions with respect to the geometric constraints of the physics-based model to more consistently produce feasible designs and match the characteristics of the physics data.