On the flipside, it relies heavily on nightly compiler features and has yet to conquer issues such as Even very basic broadcasting breaks inferability · Issue #147 · JuliaDiff/Diffractor.jl · GitHub. To my knowledge only forward mode is coming any time soon, so even if all that is addressed Diffractor may not be fit for the purpose in this thread. Given the history of promotion around this project (if anyone is curious, search for “PINN” and “Diffractor” on Discourse) and the still completely unspecified timelines, I think it’d be very hard to argue the community is overcompensating on the expectation management front.
Digressing a bit, I wonder how much crossover this thread has with Difficulties writing a program that computes PDEs involving Laplacians with AD, which doesn’t appear to have any answers yet.