High-Precision Inverse Design of Multi-Category Grating Couplers Based on KAN-Transformer

Deep learning algorithms have demonstrated significant potential in designing photonic-integrated devices, yet challenges persist in achieving high-precision predictions across wide wavelength ranges and multiple incident angles. Here, we propose a neural network-based method for rapid, high-accuracy design of multi-category grating couplers. The proposed architecture integrates the Kolmogorov-Arnold Network (KAN) module with a Transformer structure, enabling prediction of discrete grating coupler coupling efficiencies over incident angles ranging from 0<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{\circ }$</tex-math></inline-formula> to 25<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{\circ }$</tex-math></inline-formula> (5<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$^{\circ }$</tex-math></inline-formula> increments) and wavelengths spanning 1300–1700nm. Post-training prediction accuracy exceeds 99.5%. By coupling this framework with inverse design algorithms, we achieve functional requirement-specific grating coupler designs within approximately 5 minutes. Experimental validation confirms the feasibility of method. Notably, the approach exhibits exceptional scalability, requiring minimal adjustments for application to arbitrary photonic device designs.