Speaker: Mohammad Mahdi Rajabi
Date: Friday 8th of November 2024, 1:15pm.
Abstract:
To address common challenges in neural networks—such as large data requirements, poor generalization, overfitting, lack of transparency, and physically unrealistic outputs—incorporating physical intuition into different stages of neural network design has proven to be highly effective. This approach leverages the strengths of neural networks while ensuring their outputs adhere, fully or partially, to established physical laws. As a result, it improves the reliability, interpretability, and practicality of neural networks and can reduce the need for vast training datasets. This methodology is particularly useful for modeling physical systems, such as those in solid and fluid mechanics, as well as cyber-physical systems like smart grids and autonomous vehicles. With the increasing number of techniques and publications in this area, a clear and structured review of these methods has become essential. I will present an overview of current methods, terminology, and best practices for integrating physics into neural networks, providing a detailed classification of approaches including pre-training integration, in-training integration, and architecture-level embedding. I will also discuss the limitations of existing methods and highlight promising directions for future research.