U.S. Engineers Develop ChatGPT Algorithm for Solar Cell Design

A team of engineers from the University of Michigan has introduced OptoGPT, a groundbreaking algorithm designed to create optical multilayer film structures, which have a variety of applications including in solar cells.

Image: L. Jay Guo Laboratory, Michigan Engineering

OptoGPT: Transforming Optical Design

OptoGPT leverages the same underlying computer architecture as ChatGPT to reverse-engineer from desired optical properties to the required material structures. The algorithm can generate designs for multilayer film structures—comprising stacked thin layers of various materials—remarkably fast, within 0.1 seconds. These well-crafted multilayer structures can enhance light absorption in solar cells and improve designs for other optical components like telescopes. According to the developers, OptoGPT's designs typically use six fewer layers than previous models, making them simpler to manufacture.

The Design Process

Professor L. Jay Guo, an electrical and computer engineering expert at the University of Michigan, explained that designing these structures traditionally requires significant expertise, as it involves selecting the optimal materials and determining the appropriate thickness for each layer. OptoGPT simplifies this by treating materials of specific thicknesses as "words" and encoding their optical properties as inputs. The algorithm identifies correlations between these "words" to predict the next one, effectively creating a "sentence" that achieves the desired optical properties. "In a sense, we created artificial sentences to fit the existing model structure," Guo remarked.

Research and Development

Guo is the lead author of a paper titled "OptoGPT: A foundation model for inverse design in optical multilayer thin film structures," recently published in Opto-Electronic Advances. The paper highlights OptoGPT's capability to tackle the complex inverse design problem in multilayer structures. The model can accommodate various input targets and incident angles/polarizations, making it versatile for different structures and facilitating the fabrication process through its diversity and flexibility.

Future Directions

The researchers noted that while they trained OptoGPT with a large dataset of 10 million samples, this dataset only represents a small portion of the vast and complex design space for optical multilayer thin film structures. Consequently, OptoGPT might miss some potential designs outside of the sampled dataset. The team emphasized the need for close collaboration among multiple research groups to enhance the model for more generalized and intricate photonic inverse designs.

The ongoing work aims to expand OptoGPT's capabilities to handle more complex structures, further pushing the boundaries of optical design innovation.