Training optical neural networks for nonlinear logics: Towards analog simulators of artificial life

This research investigates the potential of optical systems to perform nonlinear computations using light, an inherently linear medium. By introducing nonlinearity into a physical neural network, we explore modulation strategies to identify the optimal architecture for the problem at hand. The initial phase involves modeling nonlinear material layers and analyzing their integration properties.

As a benchmark, we implement an optical decoder and an optical AND gate using Laguerre-Gaussian vortex beams and Gaussian beams to encode binary states. The phase masks within the system are designed with an adjoint algorithm, termed wavefront matching, which enables multiple input beams to interact and produce desired computational outputs based on predefined mapping rules.

To further enhance the system's performance, optimization is conducted in Fourier space. To handle increasingly complex tasks, we developed a parallel machine learning model to improve training efficiency and leverage gradient descent in discovering the optimal architecture for the continuous properties of optical components. Findings from this research confirm the feasibility of training optical neural networks for nonlinear computations, as well as advancing coherent light manipulation.

Beyond simple optical gates, this approach extends to simulating cellular automata (CA), typically created from simple rule sets. We investigate the system's ability to characterize state-update rules, laying the groundwork for future applications, such as translating Lenia — a continuous 2D cellular automaton model of artificial life — into an optical format. These advancements provide a foundation not only for optical computing but also for a novel framework to explore artificial life and cellular automata within resonating physical systems.