SciML / ModelingToolkitNeuralNets.jl

Symbolic-Numeric Universal Differential Equations for Automating Scientific Machine Learning (SciML)
MIT License
23 stars 1 forks source link

CI fails #5

Closed SebastianM-C closed 6 months ago

SebastianM-C commented 6 months ago

Describe the bug 🐞

From the CI logs it seems that we're in a very unstable region of the parameter space as I spotted a few "instability detected". I think that the contribution from the neural network is way too high (I saw a loss value of over 100), so providing a better initialization is needed.

Edit: https://github.com/SciML/UDEComponents.jl/actions/runs/8584659765/job/23525159777?pr=7#step:6:643 shows that the NaN test failed, so that confirms my suspicion.

Expected behavior

Tests pass.

SebastianM-C commented 6 months ago

Fixed by #6