AI/ML+Physics Part 3: Designing an Architecture

Prof. Steve Brunton

Slide at 00:21

PHYSICS INFORMED MACHINE LEARNING
PHYSICAL MODELS FROM DATA VIA OPTIMIZATION
1. DECIDE ON PROBLEM
(What are we modeling?)
2. CURATE DATA
(WHAT DATA WILL INFORM THE MODEL?)
3. DESIGNAN ARCHITECTURE
(RNN, AUTOENCODER, DMD, SINDY?)
4. CRAFT A Loss FUNCTION
(WHAT MODELS ARE "GOOD"?)
5. EMPLOY OPTIMIZATION
(WHAT ALGORITHMS TO TRAIN MODEL?)

Share slide

Summary (AI generated)

Physics Informed Machine Learning involves building models from data that either have a physical basis or are used to discover new physics or incorporate physics into the machine learning process. Today, we will focus on stage three, which involves designing and architecture.

Designing and architecture is a crucial part of the machine learning pipeline. There are various architectures that can be used to discover physics and embed physics into machine learning. One popular area in physics informed machine learning is the neural network zoo, which is a figure from Nathan Kutz and the book Data Driven Science and Engineering.

The neural network zoo provides an overview of different types of neural network architectures that can be used for specific tasks in machine learning. Some examples include auto encoder networks, Gans, deeper current networks, and many more. These architectures are created by combining different neural network building blocks.