Neural Networks vs Traditional Numerical Methods

Click For Summary
SUMMARY

Neural networks (NNs) are established as 'universal approximators' for nonlinear functions, demonstrating superior performance in certain applications compared to traditional numerical methods for nonlinear partial differential equations (PDEs). Specifically, NNs have shown promise in modeling complex equations such as the Navier-Stokes equations, potentially offering comparable accuracy with reduced computational intensity. The discussion highlights the need for further exploration into the effectiveness of NNs versus conventional methods in solving PDEs, as well as the importance of understanding the nuances between fitting models and solving equations.

PREREQUISITES
  • Understanding of neural networks and their function as universal approximators
  • Familiarity with nonlinear partial differential equations (PDEs)
  • Knowledge of traditional numerical methods for solving PDEs
  • Basic concepts of computational efficiency in numerical modeling
NEXT STEPS
  • Research applications of neural networks in solving Navier-Stokes equations
  • Explore the universal approximation theorem in depth
  • Learn about specific numerical methods for nonlinear PDEs, such as finite element analysis
  • Investigate case studies comparing neural networks and traditional methods in computational fluid dynamics
USEFUL FOR

Researchers, data scientists, and engineers interested in the intersection of machine learning and numerical analysis, particularly those focused on optimizing solutions for nonlinear PDEs.

BWV
Messages
1,667
Reaction score
2,012
As neural networks are 'universal approximators' for nonlinear functions, in general how do they perform in comparison to traditional numerical methods for nonlinear PDEs? Just googling, I can find papers on applications to Navier Stokes and other problems, but I don't really have the background to judge how potentially useful they are. For example, can NNs perform better (i.e. comparable accuracy but less computationally intensive) than current numerical methods for modelling the NS equations?

(this may be better in the Computer Science forum)

https://en.wikipedia.org/wiki/Universal_approximation_theorem
 
Physics news on Phys.org
The question is a little confusing to me. I am not an expert in neural networks. Neural networks does have the ability to fit a model through data points and in the examples I have seen they did a better job than other curve fitting or statistical regression algorithms. That agrees with how I would interpret the term "universal estimator". I'm not sure that finding the solution of a PDE is the same thing. But maybe I am missing something.

PS. It only took a quick Google search to find articles on solving PDEs using neural networks, so I will leave this to others with more expertise.
 

Similar threads

Undergrad Deep Learning the new key to nonlinear PDEs?
  • · Replies 2 ·
Replies
2
Views
2K
Neural networks as mathematical models of intelligence
  • · Replies 4 ·
Replies
4
Views
2K
Undergrad Tunneling with Gaussian Wave Packet
  • · Replies 2 ·
Replies
2
Views
2K
Forex trading from engineer with good background in maths-NN
  • · Replies 26 ·
Replies
26
Views
4K
MHB Implementation of numerical method
  • · Replies 1 ·
Replies
1
Views
2K
Numerical Analysis vs Applied Analysis
  • · Replies 5 ·
Replies
5
Views
4K
Requesting suggestions for languages, libraries, and architectures for parallel (and sometimes non parallel) numerical and scientific computations
  • · Replies 8 ·
Replies
8
Views
4K
Programs How good is this applied math program?
  • · Replies 13 ·
Replies
13
Views
4K
Graduate How Can Neural Networks Control an Inverted Pendulum on a Cart?
  • · Replies 2 ·
Replies
2
Views
7K
Courses Partial Differential Equations vs Classical Mechanics 2?
  • · Replies 2 ·
Replies
2
Views
3K