pde-solver

Hi,

I am pretty new to neurodiffeq, thank you very much for the excellent library.

I am interested in the way, and the computational speed, of computing partial derivatives w.r.t. the inputs.

Take forward ODE (1D, 1 unknown variable) solver for example, the input is x, a batch of coordinates, and the output of the neural network is y, the approximated solution of the PDE at these coo

Currently any field or operator can have a name manually assigned via its name attribute. If this is not specified at operator instantiation, it will be created based on the name of the operator and operands. When the operator is evaluated, the resulting field inherits the same name. This is useful for keeping track of how operators are created, but can make it confusing to identify the structur

I was having problems reading in a single frame of a solution using solution.Solution and finally tracked it down to the fact that although the first parameter of Solution is named frame you cannot call it using this as a kwarg, e.g. the two commands below should load the same frame but the second silently does nothing:

>>> from clawpack.pyclaw.solution import Solution

>>> frameso

We now have a set of example programs for the linear shallow water solver that are producing reasonable output. It's time to start looking into a more in-depth verification study. This will involve

• Verifying p-refinement convergence rates (varying polynomial degree) in cases where an analytical solution is known ( a new example may be needed )
• Verifying h-refinement convergence rates (var