AI/ML+Physics Part 3: Designing an Architecture

I use optimization to find the fewest library elements that describe dynamics. This is an architecture - a space of functions to describe observed data. There is a loss function and optimization algorithm to find the best function in the search space parameterized by the architecture. There are two architectures that relate to physics. One example is compression, assuming low-dimensional physics, using the SINDy library procedure to get a differential equation. This is an example of architectures promoting physics, as outlined in a paper by Kathleen Champion, Nathan Kutz, and myself. The paper combines a deep neural network autoencoder to learn a low-dimensional coordinate system for physics.