Optimization of a design using a physics solver integrated with a neural network

Abstract

A computer-implemented physics solver operates within a neural network framework. A neural network accepts a coordinate for each location within a design domain and outputs a local composition for each location. A model of the design domain is formed in a physics solver. The model includes discrete elements that encompass the design domain. A solution of the model provides a value of a design objective. For a plurality of iterations, the following is performed: the neural network determines current local compositions for the locations in the design domain corresponding to the discrete elements; the current local compositions are input into the discrete elements of the physics solver to obtain a current value of the design objective; and the current value of the design objective is used to find a loss gradient of a loss function. The loss gradient is used to update the neural network during the iterations.

CPC Classifications