Digital twins for neuromorphic photonic edge devices
André Röhm, Rie Sai, Takatomo Mihana, Ryoichi Horisaki, Kazutaka Kanno, Atsushi Uchida · 2025
Photonic edge devices are promising candidates for autonomous vehicles or sensing applications close to demand. However, once deployed, reprogramming them can require computational powers or bandwidth not typically available. We investigate methods for constructing digital twin models of these photonic edge devices via explicit models or machine learning approaches. This allows efficient digital training, where only the optimized parameters need to be transmitted to the edge device. Furthermore, unit-to-unit variance of edge devices can be tolerated in such a system, if individualized twins are stored. We demonstrate the principle by constructing digital twins of an opto-electronic delay-based reservoir computer. We compare the fidelity and stability of the digital twin models and their performance when used to train the real experimental system.