Kirill Zubov

```julia using Flux, NeuralPDE, Test using Optimization, OptimizationOptimJL, OptimizationOptimisers using Integrals, IntegralsCubature using QuasiMonteCarlo import ModelingToolkit: Interval, infimum, supremum using DomainSets import Lux τ_min = 0.0 τ_max = 1.0 τ_span...

The first is just too big, PINNs in general are still quite sensitive and unstable. with too big FFN is usually underfitting.

Does it work with `Adam`?

it is working with adam and work with Flux. the problem in NeuralPDE, is somewhere is unsafe to convert into code. here is MWE, which shows that everything works directly,...

use something like "bcs=[ u(x) ~ u(x)]". but probably it needs to add the ability to use empty bcs = []

weights of NN in Float64, use Float32, that need less memory

or do the number of weights less `hls`

try Flux, probably it is trouble with GC in Lux

@marcofrancis It is only one batch by default now. Each batch must contain homogeneous data, i.e. a complete mesh, otherwise, it is a decomposition into subtasks https://neuralpde.sciml.ai/dev/tutorials/neural_adapter/#Domain-decomposition if the input...

would be nice to have a FEM full-time evolution solver for all examples