Fig. 2: Network architecture.

a The structure of multi-fidelity DeepONet, where Net1 predicts \({S}_{E}\) (S-parameter by single EM simulation) of geometry \(G\) at a specific \(f\) and Net2 predicts the residual between \({S}_{E}\) and \({S}_{M}\) (S-parameter by multi-physics simulation). FC represents the fully connected layer. \({L}_{E}\) and \({L}_{R}\) represent the loss in Net1 and Net2, respectively. b The structure of inverse design model, where \({L}_{G}\), \({L}_{S}\), \({L}_{T}\), and \({L}_{R}\) represent the loss by geometry data, EM forward model, thermal forward model, and analytical model. c The latent dynamics networks developed on the frequency flow, where Net3 is used to update the latent space \(a(f)\) and Net4 is used to predict the average temperature \(T(f)\). \({L}_{T}\) represents the loss in Net3 and Net4.