Fast EUV diffraction modeling via physics-informed neural operators and structure decomposition

We present a hybrid method combining physics-informed neural operators (PINOs) with analytical multilayer modeling for efficient simulation of EUV nano-optical devices. Traditional rigorous electromagnetic solvers face prohibitive computational costs when modeling devices with thick multilayer structures, while the Fourier-domain plane-wave decomposition-based techniques require diffraction simulations from multiple illumination directions. Our approach decomposes the simulation into three components: PINO-based downward diffraction from the absorber, analytical plane-wave reflection from the multilayer stack, and PINO-based upward diffraction. The PINO model, trained across diverse unit cells and illumination directions, performs the required diffraction computations in milliseconds within a single inference pass, independent of the number of diffraction simulations, and reduces the simulation domain size by several factors compared with full-stack modeling. Validation against rigorous full-stack simulations shows excellent agreement. The approach is readily extensible to other spectral ranges, higher refractive index systems, and can naturally accommodate multiple round-trip back-reflections between absorber and multilayer structures without increasing the total simulation time.