CDS&E: Physics-driven computational tools for photonic design

Device innovation today typically requires scientists or engineers to perform a time-consuming iterative procedure involving design and simulation. The design process is based on the application of expert knowledge to identify plausible device layouts, which are validated with a general physics-based algorithm and iteratively improved. New advances in machine learning have the potential to be disruptive in this innovation cycle, due to the ability for such algorithms to learn and process data in entirely new ways.

This proposal focuses on the development of new machine learning tools that can automate and dramatically accelerate the design and simulation procedure by orders of magnitude faster speeds. These concepts will be based on a new class of algorithms that bring together conventional concepts in the data sciences with physics. Optical devices that can serve as miniaturized optical systems will be used as a testbed to benchmark the performance of these algorithms, though the concepts are ultimately general to scientific computing problems.

If successful, these algorithms will serve as the foundation for a new class of computer-aided design tools that will help scientists and engineers innovate new classes of devices and systems with great expediency. The education goal of this project is to develop and disseminate new curricula that inspires high school students to consider STEM as a career pathway.

The research objective of this proposal is to create an algorithmic platform for the global optimization of freeform photonic devices that can scale to large area, three-dimensional, multi-functional devices. The fundamental roadblock that is targeted is the inability of existing global freeform optimization methods to practically scale to complex three-dimensional systems, due to scaling limits in the sampling and simulation of devices within the global design space.

These fundamental scaling limits will be addressed by creating data-driven and physics-driven neural network electromagnetic surrogate solvers that can couple with new global search and design space evaluation tools based on deep network training. The proposed concepts will build on a recent discovery that population-based global freeform optimization can be performed by training a generative neural network using physics-based calculations.

The expected outcomes are the development of new concepts and broadly applicable algorithms that will enable the global optimization of three-dimensional dielectric electromagnetic devices.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.