By leveraging transformative neural networks that power large language models, engineers can obtain material recipes with the optical properties they need.
Study: OptoGPT: A Basic Model for Inverse Design of Thin-Film Multilayer Optical Structures (DOI: 10.29026/oea.2024.240062)
Makers of solar cells, telescopes and other optical components may be able to design better devices faster using AI.
OptoGPT, developed by engineers at the University of Michigan, leverages the underlying computer architecture of ChatGPT to work backwards, from desired optical properties to the hardware structure that can provide them.
The new algorithm enables the design of multilayer optical film structures (stacked thin layers of different materials) that can serve a variety of purposes. Well-designed multilayer structures can maximize light absorption in a solar cell or optimize reflection in a telescope. They can improve semiconductor manufacturing with extreme UV light and improve heat regulation in buildings with smart windows that become more transparent or more reflective depending on the temperature.
OptoGPT produces designs for multilayer film structures in 0.1 seconds, almost instantly. Additionally, OptoGPT designs contain an average of six fewer layers than previous models, meaning its designs are easier to manufacture.
“Designing these structures typically 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 Guoprofessor of electrical and computer engineering at UM and corresponding author of the study published in Opto-Electronic Advances.
For a novice in the field, it’s hard to know where to start. To automate the process of designing optical structures, 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 their own needs.
“In a sense, we created artificial sentences to fit the structure of the existing model,” Guo said.
The model treats materials of a certain thickness as words, also encoding their associated optical properties as inputs. By looking for correlations between these “words,” the model predicts the next word to create a “sentence”—in this case, a design for an optical multilayer film structure—that achieves the desired property, such as high reflectance.
The researchers tested the performance of the new model using a validation dataset containing 1,000 known design structures, including their material composition, thickness, and optical properties. Comparing the OptoGPT designs to the validation set, the difference between the two was only 2.58%, lower than the closest optical properties in the training dataset, at 2.96%.
In the same way that large language models are able to answer any textual question, OptoGPT is trained on a large amount of data and able to address general optical design tasks well across the entire domain.
If researchers are focused on a task, such as designing a highly efficient coating for radiative cooling, they can use local optimization variables by adjusting the boundaries to achieve the best possible result, to further refine the thickness to improve accuracy. In testing, the researchers found that fine-tuning improved accuracy by 24%, reducing the difference between the validation dataset and the OptoGPT responses to 1.92%.
Continuing the analysis, the researchers used a statistical technique to map the associations established by OptoGPT.
“The high-dimensional data structure of neural networks is a hidden space, too abstract to understand. We tried to poke a hole in the black box to see what was happening,” Guo said.
When mapped in 2D space, materials group together by type, such as metals and dielectrics, which are electrically insulating but can support an internal electric field. All dielectrics, including semiconductors, converge to a central point as the thickness approaches 10 nanometers. Optically, the model makes sense because light behaves similarly across materials as it approaches these small thicknesses, further validating OptoGPT’s accuracy.
Known as an inverse design algorithm because it starts from the desired effect and works backwards to the material design, OptoGPT offers more flexibility than previous inverse design algorithm approaches, which were developed for specific tasks. It allows researchers and engineers to design multilayer optical film structures for a wide range of applications.
This work was supported in part by the National Science Foundation (PFI-008513 and FET-2309403).
Additional co-authors: Taigao Ma and Haozhu Wang of the University of Michigan.
L. Jay Guo is also a professor of applied physics, macromolecular science and engineering, and mechanical engineering.
/Public dissemination. This content from the original organization/authors may be of a point-in-time nature and edited for clarity, style, and length. Mirage.News takes no institutional position or bias, and all views, positions, and conclusions expressed herein are solely those of the author(s). See the full story here.