Adding supporting code for blog post #9
Conversation
| @@ -0,0 +1,97 @@ | |||
| classdef spectralConvolution1dLayer < nnet.layer.Layer ... | |||
There was a problem hiding this comment.
I expect we could use the same version here and in the 1d FNO example - maybe we should consider a directory of shared tools for the examples.
That's something we'll have to look into in future I think, not something to try deal with in this PR.
phiml-blog-supporting-code/physics-informed-ml-blog-supporting-code.prj
Outdated
Show resolved
Hide resolved
…g m files and md files for each subdirectory
… add startup script
…xample so that motion remained bounded
There was a problem hiding this comment.
I think there might be some open conversations still, but this is looking good to me.
I forget if I said elsewhere, but just in case, we should "Squash and merge" when merging this. That seems to be the default when I look, probably because we've used that elsewhere. My understanding is that will prevent those large data files, hidden in the historic commits, won't end up in the main branch since a later commit deletes them, and squashing those together will remove the files from the git history.
phiml-blog-supporting-code/README.md
Outdated
| 2. **Fourier Neural Operator**: learns the solution operator that maps input functions (e.g. forcing function) to the solution of a differential equation, utilizing Fourier transforms to parameterize the integral kernel and efficiently capture global dependencies. | ||
|
|
||
| ## Setup | ||
| Open the project file physics-informed-ml-blog-supporting-code.prj to correctly set the path. |
There was a problem hiding this comment.
I don't really like using a project. I think the path management is simple enough, when I opened the project the only path I noticed added was the /data, which we can probably reference with relative paths in most examples load(fullfile("..","data","datafile.mat"))
The project adds a bunch of .xml files without human readable names, and it's not obvious what their purpose is. That makes it harder to understand the state of local copies of the repo and harder to review changes.
I think we shouldn't use a project and should remove all the resources and metadata that it adds.
There was a problem hiding this comment.
We shouldn't add binary files to the repo where possible, particularly data files that are "large" - MB is large for git, e.g. this page says 50MB will give you warnings.
These files add a noticeable impact on the speed of pulling the repo, and it's not possible to review changes to these files as a diff (arguably the same is true of .mlx. At least visdiff locally could let me see a human readable diff).
I can message on Teams about how I'm planning to add some data separately.
| @@ -1,13 +1,13 @@ | |||
| # Physics Informed Machine Learning Methods and Implementation supporting code | |||
| This repository provides examples demonstrating a variety of Physics-Informed Machine Learning (PhiML) techniques, applied to a simple pendulum system. The examples provided here are discussed in detail in the accompanying blog post "Physics-Informed Machine Learning: Methods and Implementation". | |||
| This collection provides examples demonstrating a variety of Physics-Informed Machine Learning (PhiML) techniques, applied to a simple pendulum system. The examples provided here are discussed in detail in the accompanying blog post "Physics-Informed Machine Learning: Methods and Implementation". | |||
|
|
|||
|  | |||
There was a problem hiding this comment.
I feel like I'm missnig something, why's the alt text "hippo" 😅?
There was a problem hiding this comment.
Similar to above, it'd be nice if we can change the code so that this blank plot isn't needed.
| **Physics\-Informed Loss** | ||
|
|
||
|
|
||
| The physics loss enforces the differential equation by penalizing deviations from the governing pendulum ODE. Functions like [`dllaplacian`](https://www.mathworks.com/help/releases/R2025a/deeplearning/ref/dlarray.dllaplacian.html?searchPort=60735) and [`dljacobian`](https://www.mathworks.com/help/deeplearning/ref/dlarray.dljacobian.html) make it simple to define derivative terms appearing in ordinary and partial differential equations for use in a PINN. |
There was a problem hiding this comment.
I don't think you need that "?searchPort" part in the dllaplacian link.
phiml-blog-supporting-code/README.md
Outdated
| Created and tested with MATLAB® R2025a | ||
| - [Deep Learning Toolbox™](https://www.mathworks.com/products/deep-learning.html) |
There was a problem hiding this comment.
The generatePendulumData function uses sgolayfilt which needs Signal Processing Toolbox. I guess we won't need that when we get the data on supportfiles.
| @@ -76,6 +66,8 @@ for i = 1:numPlots | |||
| ylabel("$\theta(t)$",Interpreter='latex') | |||
| end | |||
| ``` | |||
|
|
|||
|  | |||
There was a problem hiding this comment.
It's nice if the alt text here is descriptive, like "Solution Plots" or something. Not sure if we have any control over that for MLX to MD export.
There was a problem hiding this comment.
Could these plots be made such that the y axis labels don't overlap, and the legend isn't blocking the plot?
Supporting code for the blog post Physics-Informed Machine Learning: Methods and Implementation