Global cross-attention Transformer for zero-shot metasurface inverse design

The integration of artificial intelligence with electromagnetic metasurfaces has inaugurated a new era of intelligent metasurfaces, enabling self-adaptive ability for various user demands and in complex environments. However, inverse design, as the core of intelligent metasurfaces, is typically trained based on an assumption of ideal input, thus failing to maintain robustness against complex real-world signal distortions. Here, we present a zero-shot inverse design approach based on the global cross-attention Transformer (GCAT), which is immune to complex environmental interferences despite being trained exclusively on pristine, interference-free data. Distinct from convolutional or recurrent architectures, GCAT treats inverse design as a query-based feature retrieval process, employing a set of learnable physical queries to interact with global far-field features. Through a cascaded cross-attention mechanism, the model learns to actively attend to dominant structural invariants while suppressing irrelevant interference components via soft-thresholding. Consequently, GCAT demonstrates exceptional zero-shot robustness, maintaining high reconstruction fidelity with Pearson correlation coefficient (PCC) larger than 0.97 across diverse unseen interference scenarios whereas baseline models collapse. Our work establishes a resilient algorithmic foundation for the development of intelligent metasurfaces, paving the way for reliable inverse design in dynamic real-world environments.