The development of theoretical science follows an observation-assumption-model approach, effective for simple systems but hindered by complexity in engineering. Artificial intelligence (AI) and machine learning (ML) offer a data-driven alternative for making inferences when direct solutions are elusive. Feature engineering extends dimensional analysis, revealing hidden physics from data. We present a physics-transfer (PT) framework to predict physics across digitally varied models, addressing the accuracy-performance trade-off in multiscale challenges. This is exemplified in modeling brain morphology development, essential for disease diagnosis and prognosis. Nonlinear deformation physics from basic geometries is encoded into a neural network and applied to complex brain models. Results agree with longitudinal magnetic resonance imaging (MRI) data, and learned variables correlate with physical descriptors, such as undetectable stress states and submicroscopic characteristics, demonstrating the effectiveness of PT in understanding multiscale problems.