11/05/2026
By Gad Licht

The Kennedy College of Sciences, Department of Physics, invites you to attend a master’s thesis defense by Gad Licht, titled “Exploring the Impact of Wave-Like Weight Variations on Physics and Data Losses in Physics-Guided Neural Networks for Photonic Funnels.”

Details

  • Date: Friday, Nov. 13, 2026
  • Time: 4–6 p.m.
  • Location: Zoom; contact Gad Licht for the link

Committee Members

  • Advisor: Viktor Podolskiy, Ph.D., Department of Physics and Applied Physics, University of Massachusetts Lowell
  • Hugo Riberio, Ph.D., Department of Physics and Applied Physics, University of Massachusetts Lowell
  • Timothy Cook, Ph.D., Department of Physics and Astronomy, University of Massachusetts Lowell

Abstract

Artificial neural networks (ANNs) have advanced many fields, including data science and programming, and are increasingly making an impact in hard sciences. However, ANNs require large training sets, and they can sometimes identify spurious patterns in data. Physics-guided neural networks (PGNNs) address these weaknesses by adding physics-based constraints to training, reducing the required size of the training sets and enforcing agreement between predictions and theory. A key challenge, however, is choosing the balance between data-based loss and physics-based loss (DBL vs. PBL).

This thesis investigates a dynamic alternative to fixed weighting: cyclically changing the relative DBL and PBL weights during training. The objective is to explore the effect of cyclical weighting on ANN predictions of the optical response of photonic funnels, composite-based structures that can confine light beyond the diffraction limit.

Multiple cycle designs are explored, including sine and absolute-sine schedules with and without the frequency increasing exponentially over time, with the performance of PGNNs assessed in terms of convergence speed, loss and prediction accuracy. Overall, cyclic loss-weight schedules generally improved simulation RMSE, total loss and convergence, while reducing sensitivity to initial parameter settings.