AI model focused on optics developed

A research team at Shanghai Jiao Tong University has unveiled Optics GPT, described as the world's first large language model built specifically for the field of optics, offering what developers call "virtual optical expert-level" support for research, design, education and industrial use.

Unlike general artificial intelligence tools that are trained on broad, mixed data, Optics GPT, which was released on Sunday, was developed using an "optics-native" training approach, meaning it was trained from the ground up on materials focused only on optics — the science of light and how it is generated, transmitted and controlled.

The research team said the model can deeply understand optical principles and assist work in areas that include optical communication, optical design, optoelectronics and optical computing. In simple terms, this covers technologies behind fiber-optic networks, lenses and imaging systems, light-based electronic devices and new forms of computing that use light instead of electricity.

Experts described the development as an important breakthrough at the intersection of AI and so-called hard technology — fields rooted in physical science and engineering. They said it addresses a common weakness of general-purpose AI models, which often perform well in language tasks but struggle with the precise calculations and specialized knowledge required in technical disciplines such as optics.

While general AI models are powerful, experts noted that they often lack the depth needed for professional engineering work. Optics GPT was trained on large volumes of optics-specific data, allowing it to learn core theories, design logic and problem-solving methods used by optical engineers and researchers.

To evaluate the model's capabilities, a team led by Yi Lilin, a professor at Shanghai Jiao Tong University's School of Integrated Circuits, created a testing system covering six major areas: optical physics, quantum optics, optical design, nonlinear optics, optical computing and optical communication. These fields range from the basic behavior of light to advanced technologies used in communications and computing.

The team compared Optics GPT with several mainstream general-purpose and open-source AI models. The results showed that the specialized model performed better across all major categories, demonstrating stronger professional understanding and practical engineering ability in optics.

"This confirms a new technological path," Yi said. "With focused and structured training, smaller specialized models can outperform much larger general models in specific fields."

Yi is also the deputy director of the State Key Laboratory of Photonics and Communications.

Researchers said Optics GPT has broad potential in education, research and industry. In scientific research, for example, the model can serve as a round-the-clock assistant, helping researchers quickly review academic papers, generate ideas, run simulations and design experiments, accelerating the process from theory to testing.

At an industry level, Shenzhen Wanliyan Technology, a company founded in 2023 that develops high-end electronic measurement instruments, has integrated the model into its hardware systems. The company said researchers can use voice commands to interact with the model, which then guides tasks such as setting up instruments, calibrating systems, collecting data and analyzing results.

This automated process reduces the need for manual adjustments and allows researchers to focus more on innovation rather than equipment operation, the company said.

Yi said the model is expected to strengthen key parts of the optical industry chain. It could improve the intelligence and user experience of high-end domestic instruments, support fault detection and maintenance in data center optical networks, and enable faster and more accurate factory calibration of optical components.

In laser manufacturing, he said, the model could help core equipment automatically adjust operating parameters, predict performance issues and carry out maintenance with less human intervention.

By deploying a fully self-developed domestic model in these applications, Yi said, Optics GPT could provide new tools and infrastructure to support independent research and the upgrading of China's hard technology sectors, particularly in optics.

zhouwenting@chinadaily.com.cn