Physics-constrained deep-learning framework for MRI metasurfaces

Metasurfaces resolve limited signal-to-noise ratio (SNR) and radio frequency (RF) inhomogeneity bottlenecks in magnetic resonance imaging (MRI). However, strong unit-cell near-field coupling creates an "ill-posed"' design problem, where conventional optimization struggles with computation costs and non-unique mappings. We propose a physics-constrained bidirectional deep learning framework for MRI metasurfaces. A Spectral-Feature U-Net surrogate model substitutes time-consuming simulations. To address multi-valued mapping in inverse design, a tandem training strategy with physics-consistency constraints guides a residual multi-layer perceptron (Res-MLP) toward physically equivalent solutions. In simulations incorporating a human arm phantom, an AI-designed metasurface achieved precise 63.8 MHz resonance and excellent field homogeneity, keeping specific absorption rate (SAR) within safety limits. This establishes a robust "AI for Science" toolchain for automated, patient-adaptive MRI hardware design.