Kirill Zubov

but yeah, I agree, it will be cool to have some solution wrapper that can make base manipulation to a solution. something like: ```julia sol.y(0.1) sol.approx(dx) plot(sol) sol.res ```

https://github.com/SciML/NeuralOperators.jl/issues/83

the first script, training neural operator for a family of Burger equations. training is mapping between functional space of some initial conditions {u(t0 x)} and solution of equation u(t,x) here...

fine-tuning optimization with PINNs here it's used already trained neural operator over the entire given some functional space of initial conditions how pre-trained state for PINNs solver in order to...


the problem is that points are generated for the loss function as a whole and not for each variable. In essence, we are solving the equation u(t, 1) ~ u(t,...

@mcabbott oh, sorry. yes, Flux case gets `ERROR: control flow.` The previous error was probably due to an overload of Chain from pinns.jl https://github.com/JuliaDiff/Diffractor.jl/blob/be4eeb59879b7f6773746f9f3b4ef8df38ac9f99/test/pinn.jl#L16

`QuadratureStrategy` is still unstable. Try ` NeuralPDE.QuadratureTraining(quadrature_alg = CubatureJLh())` I'm not sure that it is can be more stable it needs to dig into the problem. `GridTraining` is most stable...

I ran the script with Grid(predict) and quadrature strategy(predict2) it is quite the same ``` strategy = NeuralPDE.QuadratureTraining(quadrature_alg = CubatureJLh(), reltol = 1e-3, abstol = 1e-3, maxiters = 50) ```

the same that you, only I changed the strategy to ``` strategy = NeuralPDE.QuadratureTraining(quadrature_alg = CubatureJLh(), reltol = 1e-3, abstol = 1e-3, maxiters = 50) ```