Optical Neural Network for Scientific PDEs

Partial differential equations (PDEs) are vital for modeling scientific phenomena but are computationally intensive with traditional solvers. Machine learning (ML) offers a data-driven alternative, yet electronic ML hardware struggles with scalability and energy efficiency. Optical computing, harnessing photons' parallelism and low energy use, presents a powerful solution. We introduce the Optical Neural Engine (ONE), a reconfigurable architecture integrating diffractive optical neural networks (DONNs) [1] for Fourier space [2] and optical crossbars (XBARs) [3] for real space processing. ONE excels in efficiency and accuracy for diverse PDEs, enabling scalable, high-performance scientific computations. We propose an optical neural engine (ONE) combining diffractive and crossbar optical systems to solve diverse PDEs, including Darcy flow, Navier-Stokes, and Maxwell's equations. ONE achieves energy-efficient, scalable, and real-time computation for scientific applications.