Hello everyone! I wanted to resurrect this old topic. Has anyone found a better way to do higher order derivatives using PINNs over the last year? It seems as though the best option is still to do finite differences. I also noticed that the same equation I am working with (helmholtz equation) still has an open topic that was mentioned in the beginning of this thread.
You can do forward-over-reverse (no finite differences required) as described in this thread: Nested AD with Lux etc
Still nothing for Flux/zygote? I really like Flux and have put a lot of time of my research into it. It would be nice to have an option that works with Flux/Zygote. Other than PINNs, Flux has treated me very well.
Lux is still a solution, and it’s still the case that nested AD is a terrible way to do PINNs anyways.
2 Likes