Gridap.jl VS ModelingToolkit.jl

Another I've been working on learning is Julia, which aims to use a syntax very similar to how you'd write it mathematically, and I like being able to include units in calculations using the unitful.jl package, and there are FEM packages available like Gridap.

PINN methods are absurdly slow (DeepXDE is about 10,000x slower than an ODE solver for example, while using implicit parallelism vs serial ODE solver) but they are flexible. So ModelingToolkit.jl has alternative options, like DiffEqOperators.jl takes the same specification and generates ODESystem and NonlinearSystem problems via finite difference discretizations (known as "method of lines"). There's a (pseudo-)spectral part of the interface coming relatively soon as well, with GridAP.jl integration for FEM coming soon. So this is made to be a universal arbitrary PDE -> approximate solution interface which is generic to the method and solving process.

I'd use a modeling tool like https://mtk.sciml.ai/dev/ Using the standard library, you wouldn't need to come up with all equations yourself. Depending on the details of your use case, system identification as suggested before might be a faster approach though.

In this case, yes, this should just be done numerically. But using symbolic transformations to optimize numeric code is also a really neat application of symbolic computing that doesn't get enough attention, imo. [This library](https://github.com/SciML/ModelingToolkit.jl), for example, uses symbolics to do sparsity detection, automatic derivative/gradient/jacobian/hessian calculations, index reduction, etc. to speed up numerical differential equation solving.

PDEs are difficult because you don't have a simple numerical definition over all PDEs because they can be defined by arbitrarily many functions. u' = Laplace u + f? Define f. u' = g(u) * Laplace u + f? Define f and g. Etc. To cover the space of PDEs you have to go symbolic at some point, and make the discretization methods dependent on the symbolic form. This is precisely what the ModelingToolkit.jl ecosystem is doing. One instantiation of a discretizer on this symbolic form is NeuralPDE.jl which takes a symbolic PDESystem and generates an OptimizationProblem for a neural network which represents the solution via a Physics-Informed Neural Network (PINN).

There's a very clear momentum for Julia here in this domain of modeling and simulation. With JuliaSim funding an entire modeling and simulation department within Julia Computing dedicated to building out an ecosystem that accelerates this domain and the centralization around the SciML tooling, this is an area where we absolutely have both a manpower and momentum advantage. We're getting many universities (PhD students and professors) involved on the open source side, while building out different commercial tools and GUIs on top of the open numerical core. The modeling and simulation domain itself is soon going to have its own SciMLCon since our developer community has gotten too large to just be a few JuliaCon talks: it needs its own days to fit everyone! Not only that, in many aspects we're not just moving faster but have already passed. Not in every way, there's still some important discussion in controls that needs to happen, but that's what the momentum is for.

Julia has great support for modeling, have a look at ModelingToolkit.jl. From the README: