Kirill Zubov

@finmod try run the code from the test https://github.com/SciML/NeuralPDE.jl/blob/2e9bac04111bb6a0049da1f26939b09984a53bcb/test/NNPDE_tests.jl#L480

The difference is that we train on more points in the tests. ```julia ts = [domain.domain.lower:dt/5:domain.domain.upper for domain in domains][1] function getData(sol) data = [] us = hcat(sol(ts).u...) ts_ =...

I wonder what it looks like so the learning behavior is different (example in the docs) depending on the operating environment or something else

also, try updating all packages to the latest versions and running the version from the documentation. To make sure it doesn't depend on package versions

I tried to write this a long time ago, here is a draft(not work and overdate) ```julia ## Shrödinger Equation ##Let h = u + iv where \[u,v\] = phi(x,t),...

Not very clear with the 3rd order ODE about speed. but the accuracy of the solution is clearly highe. The numerical method is fast because it is not accurate. That...

There require 2 dim. for KS. I suppose that it will work well with AD. I will write a hardcode 2 dim version tomorrow for presentation but actually it parser...

So far, AD is significantly slower than the numerical derivative. Poisson's eq, 2D maxiters=100 AD BFGS 39.180 s (33577736 allocations: 68.19 GiB) ADAM 14.158 s (12608940 allocations: 24.47 GiB) numeric...

I think it is eventual to speed up significantly AD runtime, for example, https://juliadiff.org/ForwardDiff.jl/v0.10.2/user/advanced.html#Configuring-Chunk-Size-1

Kuramoto–Sivashinsky equation looks good with ForwardDiff AD but still too long BFGS, 300 iterations