Built a groundwater prediction model on PhysicsNeMo, looking for feedback on the port

[Rekin226]

Hi all,

I've been using PhysicsNeMo to build a model that predicts groundwater levels across a network of monitoring wells in Taiwan, and I wanted to share it here to get some feedback from people who know the framework better than I do.

The short version: instead of the usual approach where you calibrate a separate model for each well, I trained one physics-informed model across all of them at once, running on CUDA. Mixed precision training came out around 14x faster than CPU, which is what made it practical to iterate on.

Repo: https://github.com/Rekin226/HydroPhysicsAI
Live demo: https://huggingface.co/spaces/Rekin226/HydroPhysicsAI-demo

I'm still fairly new to PhysicsNeMo, so I'd really value some input:

1. Is there an idiomatic way in PhysicsNeMo to condition one shared model on per-item parameters? Right now I hand-rolled a small hypernetwork and I suspect there's a cleaner built-in pattern I'm missing.
2. Is anything in my setup working against how PhysicsNeMo is meant to be used? I'd like the port to read as a good example rather than a workaround.
3. Are there similar PhysicsNeMo projects I could learn from, and would anyone be open to contributing?

Thanks for building this. It made porting the model a lot smoother than I expected.

[Rekin226]

Small follow-up: to make the "how do I use the framework more idiomatically" part concrete, I opened a tracking issue on my side with a proposed roadmap for deepening the integration:

Rekin226/HydroPhysicsAI#33

Right now the port is honest but container-level: the network is a real physicsnemo.Module with .mdlus checkpointing, but training still runs my own PyTorch loop, so the ModelMetaData(amp=True) flag isn't actually exercised yet. The plan is to route training through StaticCaptureTraining / StaticCaptureEvaluateNoGrad (and DistributedManager), then benchmark that against my hand-rolled AMP loop at identical accuracy.

If anyone has done attribute-conditioned hypernetworks over a shared ODE this way, or can sanity-check that approach, I'd love the input. Happy to take contributions on the issue too.