Constructing hidden differential equations using a data-driven approach with the alternating direction method of multipliers (ADMM)
This paper adopted the alternating direction method of multipliers (ADMM) which aims to delve into data-driven differential equations. ADMM is an optimization method designed to solve convex optimization problems. This paper attempted to illustrate the conceptual ideas and parameter discovery of the...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
AIMS Press
2025-02-01
|
| Series: | Electronic Research Archive |
| Subjects: | |
| Online Access: | https://www.aimspress.com/article/doi/10.3934/era.2025040 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This paper adopted the alternating direction method of multipliers (ADMM) which aims to delve into data-driven differential equations. ADMM is an optimization method designed to solve convex optimization problems. This paper attempted to illustrate the conceptual ideas and parameter discovery of the linear coupled first-order ODE. The estimation of the coefficients of the underlying equation utilized a combination of algorithms between physics-informed neural networks (PINNs) and sparse optimization. Both methods underwent a sufficient amount of looping during the search for the best combinations of coefficients. The PINNs method took charge of updating weights and biases. The updated trainable variables were then fetched to the sparse optimization method. During the sparse optimization process, ADMM was used to restructure the constrained optimization problems into unconstrained optimization problems. The unconstrained optimization problem usually consists of smooth (differentiable) and non-smooth (non-differentiable) components. By using the augmented Lagrangian method, both smooth and non-smooth components of the equations can be optimized to suggest the best combinations of coefficients. ADMM has found applications in various fields, such as signal processing, machine learning, and image reconstruction, which involve decomposable structures. The proposed algorithm provides a way to discover sparse approximations of differential equations from data. This data-driven approach provides insights and a step-by-step algorithm guide to allow more research opportunities to explore the possibility of representing any physical phenomenon with differential equations. |
|---|---|
| ISSN: | 2688-1594 |