Abstract
Machine learning (ML) techniques have recently gained significant attentionfor solving compliance minimization (CM) problems. However, these methodstypically provide poor feature boundaries, are very expensive, and lack asystematic mechanism to control the design complexity. Herein, we address theselimitations by proposing a mesh-free and simultaneous framework based onphysics-informed Gaussian processes (GPs). In our approach, we parameterize thedesign and state variables with GP priors which have independent kernels butshare a multi-output neural network (NN) as their mean function. Thearchitecture of this NN is based on Parametric Grid Convolutional AttentionNetworks (PGCANs) which not only mitigate spectral bias issues, but alsoprovide an interpretable mechanism to control design complexity. We estimateall the parameters of our GP-based representations by simultaneously minimizingthe compliance, total potential energy, and residual of volume fractionconstraint. Importantly, our loss function exclude all data-based residuals asGPs automatically satisfy them. We also develop computational schemes based oncurriculum training and numerical integration to increase the efficiency androbustness of our approach which is shown to (1) produce super-resolutiontopologies with fast convergence, (2) achieve comparable compliance and lessgray area fraction compared to traditional numerical methods, (3) providecontrol over fine-scale features, and (4) outperform competing ML-basedmethods.