Digital twin optical computing system speeds development without hardware access

Researchers from Tsinghua University and the National University of Defense Technology have built a digital twin optical computing system that lets developers train, optimize and transfer models without constant access to physical hardware. The approach cut one training task from about 1.74 days to 10.7 minutes and matched physical-system accuracy across benchmark applications.

Why it matters: - DT-OCS could remove a major bottleneck in optical computing development by reducing reliance on scarce physical hardware. - The system supports parallel experimentation, faster optimization and more reproducible benchmarking. - The approach could make high-speed optical computing more practical for more developers and more applications.

What happened: - Researchers from Tsinghua University and the National University of Defense Technology developed the Digital Twin Optical Computing System, or DT-OCS. - The system was presented as a software-based counterpart to a specific physical optical computing system. - DT-OCS was implemented as a callable software module. - The team demonstrated DT-OCS on a 10 GHz silicon-based integrated optical computing system. - Once optimization is complete in the digital twin, configuration parameters can be transferred directly to the physical system without additional hardware retraining or post-deployment fine-tuning.

The details: - DT-OCS reproduces end-to-end computational behavior across different operating configurations. - The platform targets a workflow problem in optical computing where developers often need repeated access to physical devices for configuration, optimization and testing. - In handwritten digit recognition, the digital-twin model reached 98.09% classification accuracy. - The transferred physical model reached 98.11% accuracy on the same task. - In color clothing and footwear classification, both the digital and transferred models achieved 83.07% accuracy. - In a gold trading strategy test for temporal decision-making, training that previously took about 1.74 days on the physical optical system took 10.7 minutes with DT-OCS. - That training time was less than 0.4% of the original time. - Because DT-OCS is reproducible software, multiple instances can run at the same time without competing for access to the same physical hardware. - The framework is open source. - The work is positioned as a foundation for more complex hybrid optoelectronic computing systems. - The paper was published in Opto-Electronic Advances. - A link to the journal is available as More information. - The article is titled "Digital twin optical computing system" and was published by Sun R, Li YM, Fu TZ et al. in Opto-Electron Adv 9, 250254 (2026). - The DOI is 10.29026/oea.2026.250254.

Between the lines: - The core shift is from hardware-bound optical development to a software-first workflow. - That matters because optical systems are difficult to scale when each optimization cycle depends on physical access. - Matching digital and physical accuracy suggests the twin can be useful not just for simulation, but for deployment-oriented optimization. - The open-source model could lower the barrier for researchers trying to compare methods or build on the platform.

What's next: - The DT-OCS framework could be used to develop more complex hybrid optoelectronic computing systems. - The software approach may help move optical processors from specialized hardware toward more flexible computing resources. - Wider adoption will depend on how well the method transfers across additional hardware platforms and application types.

The bottom line: - DT-OCS offers a faster, hardware-light way to develop optical computing systems without sacrificing performance after transfer to physical devices.