The AI tool, named OptoGPT, leverages the same architecture as ChatGPT to reverse-engineer desired optical properties into the necessary material structures. This innovation is aimed at improving the design of optical multilayer film structures, which consist of stacked thin layers of different materials.
These multilayer structures are crucial for various applications, such as maximizing light absorption in solar cells, optimizing reflection in telescopes, enhancing semiconductor manufacturing with extreme UV light, and regulating building heat with smart windows that adjust transparency and reflectivity based on temperature.
OptoGPT is capable of producing designs for these multilayer film structures in just 0.1 seconds. Moreover, its designs typically feature six fewer layers than previous models, simplifying the manufacturing process.
"Designing these structures usually requires extensive training and expertise as identifying the best combination of materials, and the thickness of each layer, is not an easy task," said L. Jay Guo, U-M professor of electrical and computer engineering and corresponding author of the study published in Opto-Electronic Advances.
The research team adapted a transformer architecture-the machine learning framework used in large language models like OpenAI's ChatGPT and Google's Bard-to automate the design process for optical structures.
"In a sense, we created artificial sentences to fit the existing model structure," Guo said.
The model treats materials at a specific thickness as words, encoding their associated optical properties as inputs. By identifying correlations between these "words," the model predicts the next word to create a "phrase"-a design for an optical multilayer film structure-that achieves the desired optical property, such as high reflection.
To validate OptoGPT's performance, researchers used a dataset of 1,000 known design structures, including their material composition, thickness, and optical properties. OptoGPT's designs differed from the validation set by only 2.58%, a lower variance than the 2.96% found in the training dataset.
OptoGPT is trained on a vast amount of data, enabling it to handle general optical design tasks across the field. For focused tasks, like designing high-efficiency coatings for radiative cooling, researchers can use local optimization to fine-tune the thickness and improve accuracy. This method improved accuracy by 24% during testing, reducing the difference between the validation dataset and OptoGPT's responses to 1.92%.
The researchers further analyzed OptoGPT's decision-making process using a statistical technique to map out associations. They discovered that materials clustered by type, with dielectrics converging around a central point as thickness approached 10 nanometers, reflecting similar light behavior regardless of material at small thicknesses. This finding further validated OptoGPT's accuracy.
As an inverse design algorithm, OptoGPT starts with the desired effect and works backward to create a material design. It offers greater flexibility than previous algorithms designed for specific tasks, allowing for the design of optical multilayer film structures for a wide range of applications.
Research Report:OptoGPT: A foundation model for inverse design in optical multilayer thin film structures
Related Links
University of Michigan
Space Technology News - Applications and Research
| Subscribe Free To Our Daily Newsletters |
| Subscribe Free To Our Daily Newsletters |
