Massively parallel universal nonlinear computing with diffractive surfaces
Abstract
We introduce a diffractive optical computing framework that enables universal approximation of nonlinear functions using linear, passive optical components. By encoding the input variable into the phase of an optical wavefront and processing it through cascaded spatially engineered diffractive surfaces, independently defined nonlinear functions are computed in parallel at diffraction-limited output pixels. We theoretically prove universality for arbitrary bandlimited nonlinear functions and numerically demonstrate the approximation of one million nonlinear functions in parallel with negligible error. The framework is experimentally validated through in situ optical learning, demonstrating the simultaneous implementation of 35 nonlinear functions using a compact spatial light modulator-based setup.