Researchers introduce model-free training
Optical computing has emerged as a powerful approach for high-speed and energy-efficient information processing. Diffractive optical networks, in particular, enable large-scale parallel computation through the use of passive structured phase masks and the propagation of light. However, one major challenge remains: systems trained in model-based simulations ofen fail to perform optimally in real experimental settings, where misalignments, noise, and model inaccuracies are difficult to capture.
In a new paper published in Light: Science & Applications, researchers at the University of California, Los Angeles (UCLA) introduce a model-free in situ training framework for diffractive optical processors, driven by proximal policy optimization (PPO), a reinforcement learning algorithm known for stability and sample efficiency. Rather than rely on a digital twin or the knowlege of an approximate physical model, the system learns directly from real optical measurements, optimizing its diffractive features on the hardware itself.
“Instead of trying to simulate complex optical behavior perfectly, we allow the device to learn from experience or experiments,” said Aydogan Ozcan, Chancellor’s Professor of Electrical and Computer Engineering at UCLA.
AI Designs Optical Processors Without Human Programming
Researchers at the University of California, Los Angeles (UCLA) have developed a new approach to designing optical processors using artificial intelligence. This method allows for the creation of processors without the need for traditional, manual design processes, potentially revolutionizing the field of optical computing.
The team utilized in situ reinforcement learning with proximal policy optimization, a technique where an AI agent learns to optimize a system directly within its operating environment. This allowed the AI to design optical processors capable of performing complex mathematical operations, specifically matrix multiplication, which is fundamental to many computing tasks.
Unlike conventional methods that rely on human expertise and iterative refinement, the AI autonomously explored various processor configurations, learning through trial and error to identify designs that maximized performance. The resulting processors are “model-free,” meaning they don’t require pre-existing models of light propagation,simplifying the design process and potentially leading to more innovative solutions.
“Our approach bypasses the need for complex simulations and human intuition, allowing the AI to discover designs that might not have been conceived otherwise,” explained Yuhang Li, a lead author of the study.
The research, published in Light: Science & Applications, demonstrates the potential of AI to accelerate the advancement of advanced optical computing technologies. This could lead to faster, more energy-efficient processors for a wide range of applications, including machine learning, image processing, and scientific computing.
More information:
Yuhang Li et al, Model-free optical processors using in situ reinforcement learning with proximal policy optimization, Light: Science & Applications (2026). DOI: 10.1038/s41377-025-02148-7
Provided by
UCLA Engineering Institute for Technology Advancement