Physics-Informed Structural Attention Unlocks Efficient and Interpretable Inverse-Designed Photonic Devices

Abstract Photonic inverse design has attracted widespread attention due to its advantages in realizing high-performance and compact optical devices. Analog metamaterials offer design flexibility through freeform geometries, but their fabrication requires additional enforcement of minimum feature size constraints. Digital metamaterials provide a more manufacturable alternative; however, the discrete nature of the design problem imposes a “curse of dimensionality” that renders conventional search methods computationally prohibitive. Moreover, existing methods operate as black boxes, lacking interpretability. To address these limitations, we develop the adjoint-attention optimization method, a digital-metamaterial-based inverse design framework that leverages physics-informed structural attention to achieve both high efficiency and interpretability. At its core, the structural attention evaluates the salience of each design unit by acting on adjoint sensitivity, assigning higher scores to design variables whose state transitions most strongly improve device performance. We validate the framework through the design and experimental demonstration of multimode demultiplexers. Numerical results show that the score-based attention concentrates the discrete search within a low-dimensional subspace, substantially reducing computational cost compared with conventional discrete search methods. Structural attribution further identifies the functional hotspot regions that govern device functionality, revealing the low-rank property of the design space. Experimental results confirm the high performance of the fabricated devices across the C-band. The low-loss, low-crosstalk four-mode mode-division multiplexing circuit enables on-chip optical interconnect with single-wavelength transmission of 4 modes × 224 Gb/s. This approach establishes a general framework for efficient, interpretable inverse design and marks a step from blind automation to physics-guided, physically interpretable photonic device design.