Figure 1.

The framework of Generative Inverse Design Networks (GIDNs). GIDNs consist of two DNNs: the predictor and the designer. In the predictor, the weights and biases are learning variables. Those values are optimized to minimize the difference between the ML predictions and ground truth. In the designer, the values of the weights and biases are adopted from the predictor and set to be constants. Initial designs generated with values from a Gaussian distribution are fed into the designer as inputs and optimized designs are generated as outputs. In the feedback loop, the optimized designs are verified by a physics‐based model and can be added to previous training data for the next iteration of training and design processes.