Abstract

Optical computing systems (OCS) are promising accelerators for artificial intelligence due to their high bandwidth, low latency, and inherent parallelism. However, in existing OCS implementations, task development often requires direct participation of physical hardware during training and optimization, tightly coupling development workflows to device access and limiting offline design, reproducible benchmarking, and parallel exploration.

Here, we propose a Digital Twin Optical Computing System (DT-OCS), a system-level, measurement-driven digital surrogate that emulates the end-to-end input-output behavior of a specific physical OCS under different operating configurations. DT-OCS is implemented as a differentiable software module, enabling fully offline task training and configuration optimization. The optimized configuration parameters can be directly transferred to the physical OCS without hardware-in-the-loop retraining. We validate DT-OCS on a high-speed optical computing system operating at 10 GHz and equipped with a silicon-based integrated computing chip. Representative tasks, including image classification and temporal strategy generation, are evaluated.

Across all tasks, the transferred models closely match the performance of their digital-twin counterparts after direct parameter transfer. These results demonstrate that DT-OCS enables a hardware-decoupled and fully offline development paradigm for optical computing systems. The DT-OCS implementation is released as open-source code to facilitate reuse and further development.