Predicting brain morphogenesis via physics-transfer learning.

Brain morphology emerges from the interplay of genetic programming and mechanical forces, yet its fractal-like folding patterns make quantitative analysis and prediction difficult, especially when labeled data are scarce. Here we introduce a theory-grounded physics-transfer learning framework that enables generalization analysis and prediction in complex physical systems by leveraging consistent governing laws across levels of complexity. Specifically, mechanistic insights into the nonlinear elasticity of simple, analytically tractable geometries are embedded into neural networks and successfully transferred to brain models through a sequence of physics-anchored domains, integrating explicit physics constraints into learning theory. Building on the physics-transfer theory, we derive a generalization bound that explains the strong performance in brain feature characterization and morphogenesis prediction. Beyond predictive accuracy, the framework yields reduced-dimensional evolutionary representations that distill the essential physics of brain morphogenesis. Validation against medical imaging data demonstrates the promise of physics-aware digital-twin technologies for understanding, diagnosing and intervening in the developing and diseased brain.