Opening the black box for data efficiency and inverse design in photonics

Supervised neural networks are rising as an algorithm of choice for surrogate models in photonics, because they are versatile, fast to evaluate, easily differentiable, and perform well in high-dimensional problems. However, the drawback of this black box approach is that it requires a lot of data. Unfortunately in the context of photonics, data is generated through expensive full solves of Maxwell’s equations. This talk will present ways to open the black box for better data efficiency and performance of deep surrogate models. The first part of this talk will present how active learning can reduce the need for data by at least an order of magnitude by adapting the data generation to the model learning. The second part will present how information about the physics can be incorporated into the neural network for more efficiency.