Forecasting the beta-FPUT Recurrences with Machine Learning Models

Note: Meeting ID: 918 3071 5048
Speaker: Matthew Litton, Physics Department, BA Candidate

When: April 29, 2021 (Thu), 11:30AM to 12:30PM (add to my calendar)

This event is part of the Thesis Defenses.

The Fermi-Pasta-Ulam-Tsingou (FPUT) problem is a cornerstone in the study of nonlinear systems, statistical mechanics, and experimental/computational mathematics. We will approach this problem by developing machine learning models to approximate the beta-FPUT dynamics with a focus on correct predictions of future recurrences to initial conditions. The beta-FPUT chain is a high-dimensional nonlinear system which in general proves challenging for machine learning algorithms and so serves as a test of the capabilities of modern deep neural networks. The base architecture for our model is a Long Short-Term Memory/Dense Neural Network (LSTM-DNN) hybrid model. We will further address some modifications to this architecture, primarily with a focus on regularization, using both model independent L2 regularization and a Physics Guided Neural Network model (PGNN) which incorporates an unsupervised energy conserving penalty term to the loss function. The models are then evaluated by three metrics, the root mean squared error (RMSE) between the predicted and the true solution, recurrence times, and energy difference.

Time: Apr 29, 2021 11:30 AM Eastern Time (US and Canada)

Join Zoom Meeting https://bostonu.zoom.us/j/91830715048?pwd=L3FvamRQTmtrcUZqUDlaWnByTFlYUT09

Meeting ID: 918 3071 5048 Passcode: 659173