Deep residual network for interpolation and inverse problems
Part Of
Journal of Physics: Conference Series
Journal Volume
2766
Journal Issue
1
Start Page
012093
ISSN
17426588
Date Issued
2024
Author(s)
DOI
10.1088/1742-6596/2766/1/012093
Abstract
In this paper, we introduce the Power-Enhancing Residual Network, a simplified version of the highway network. This novel neural network architecture aims to enhance interpolation capabilities. By incorporating power terms in residual elements, this architecture enhances the network's expressive capacity, leading to new possibilities in deep learning. We explore key design aspects such as network depth, width, and optimization techniques, showcasing its adaptability and performance advantages. Results highlight its precision and demonstrate superiority over conventional networks in accuracy, convergence speed, and computational efficiency. Additionally, we investigate deeper network configurations and apply the architecture to solve the inverse Burgers' equation, illustrating its effectiveness in real-world problems. Overall, the Power-Enhancing Residual Network represents a versatile and transformative solution, pushing the boundaries of machine learning. The codes implemented are available at: https://github.com/CMMAi/ResNet_for_PINN.
Event(s)
9th European Thermal Sciences Conference, EUROTHERM 2024, 10 June 2024 - 13 June 2024, Bled
Publisher
Institute of Physics
Description
Article number 012093
Type
conference paper