Add FloodForecaster: Domain-Adaptive GINO Framework for Flood Forecasting by MehdiTaghizadehUVa · Pull Request #1271 · NVIDIA/physicsnemo

MehdiTaghizadehUVa · 2025-12-10T03:19:40Z

PhysicsNeMo Pull Request

Description

This PR integrates FloodForecaster, a domain-adaptive Geometry-Informed Neural Operator (GINO) framework for rapid, high-resolution flood forecasting. The framework enables accurate, real-time flood predictions by learning from source domain data and adapting to target domains through adversarial training.

Key Features

Domain-Adaptive Training: Three-stage pipeline (pretraining → domain adaptation → rollout evaluation) for transfer learning from data-rich to data-scarce domains
Gradient Reversal Layer (GRL): Adversarial domain adaptation using CNN-based domain classifier
GINO Integration: Combines Graph Neural Operators (GNO) and Fourier Neural Operators (FNO) for irregular terrain processing
Physics-Informed Metrics: Volume conservation, arrival time, inundation duration, CSI, and FHCA
PhysicsNeMo Module Compliance: All components inherit from physicsnemo.Module with checkpointing support

Components Added

Training modules: domain adaptation trainer, pretraining pipeline
Data processing: GINO wrapper, preprocessing/postprocessing utilities
Datasets: Custom dataset classes for flood data loading
Inference: Rollout evaluation and visualization pipeline
Configuration: Hydra-based configuration system
Documentation: Complete README with usage examples

Checklist

I am familiar with the Contributing Guidelines.
New or existing tests cover these changes.
The documentation is up to date with these changes.
The CHANGELOG.md is up to date with these changes.
An issue is linked to this pull request.

Dependencies

No new dependencies required. All packages are either already in PhysicsNeMo or standard scientific computing libraries:

neuralop>=2.0.0 (existing)
hydra-core>=1.2.0 (existing)
wandb>=0.12.0 (optional, for logging)
Standard packages: matplotlib, tqdm, numpy, torch, omegaconf, pandas, h5py

Review Process

All PRs are reviewed by the PhysicsNeMo team before merging.

Depending on which files are changed, GitHub may automatically assign a maintainer for review.

We are also testing AI-based code review tools (e.g., Greptile), which may add automated comments with a confidence score.
This score reflects the AI’s assessment of merge readiness and is not a qualitative judgment of your work, nor is
it an indication that the PR will be accepted / rejected.

AI-generated feedback should be reviewed critically for usefulness.
You are not required to respond to every AI comment, but they are intended to help both authors and reviewers.
Please react to Greptile comments with 👍 or 👎 to provide feedback on their accuracy.

- Moved sample_animation.gif to docs/img - Removed functions using pickle due to security concerns - Removed duplicate import of KolmogorovArnoldNetwork in __init__.py

mnabian · 2025-12-11T01:15:39Z

+
+```yaml
+model:
+  model_arch: 'gino'


Is this actually configurable?

mnabian · 2025-12-11T01:18:08Z

+
+For questions, feedback, or collaborations:
+
+- **Mehdi Taghizadeh** – <jrj6wm@virginia.edu>


Please mention that you are the code contributor and maintainer.

mnabian · 2025-12-11T01:19:01Z

+- **Jonathan L. Goodall** – <jlg7h@virginia.edu>
+- **Negin Alemazkoor** (Corresponding Author) – <na7fp@virginia.edu>
+
+## License


You don't need this section. Please remove it.

mnabian · 2025-12-11T01:19:30Z

@@ -0,0 +1,338 @@
+# FloodForecaster: A Domain-Adaptive Geometry-Informed Neural Operator Framework for Rapid Flood Forecasting


Thanks for the comprehensive and very well structured readme!

mnabian · 2025-12-11T01:20:51Z

+  wireup_info: 'mpi'
+  wireup_store: 'tcp'


Would you be able to use physicsnemo distributed utilities instead?

CharlelieLrt · 2025-12-11T01:21:43Z


            # Load state dict after closing archive
-            model_dict = torch.load(io.BytesIO(model_bytes), map_location=device)
+            model_dict = torch.load(io.BytesIO(model_bytes), map_location=device, weights_only=False)


That's BC breaking, we can't accept this change

mnabian · 2025-12-11T01:22:33Z

+  wireup_store: 'tcp'
+  model_parallel_size: 1
+  seed: 123
+  device: 'cuda:0'


If we set use_distributed=True, how should the device filed be specified? Would it be a list of devices? Please clarify. It would be nice to add some details about distributed computing configs in the readme.

mnabian · 2025-12-11T01:23:56Z

+  resume_from_dir: null
+  resume_from_source: null
+  resume_from_adapt: null


Are these needed? If som please describe how these should be used.

mnabian · 2025-12-11T01:24:54Z

+  n_epochs_adapt: 2
+  learning_rate: 1e-4
+  adapt_learning_rate: 1e-4
+  training_loss: 'l2'


List the available options in a comment

mnabian · 2025-12-11T01:24:59Z

+  training_loss: 'l2'
+  testing_loss: 'l2'
+  weight_decay: 1e-4
+  amp_autocast: false


List the available options in a comment

mnabian · 2025-12-11T01:38:49Z

+
+        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
+
+    def save_checkpoint(self, save_folder: str, save_name: str) -> None:


I'm curious, how is GiNO's save and load checkpoint methods different from torch's save/load utilities? Couldn't you use the physicsnemo save/load utilities for better integration?

…self

…nfig-name are defaults)

… update distributed docs

… coverage - Rename trainer_physicsnemo.py to trainer.py and update all references - Update GINOWrapper.load_state_dict to handle PhysicsNeMo checkpoint format - Fix checkpoint loading in inference.py to support both PhysicsNeMo and neuralop formats - Add comprehensive unit tests for data processing, training, trainer, and inference - Reduce test suite size to align with other PhysicsNeMo examples - Remove test coverage analysis markdown file - Fix various test failures related to model structure and checkpoint loading

MehdiTaghizadehUVa · 2026-01-13T19:15:47Z

Hi @mnabian, All review comments are addressed:

Deleted untrained_checkpoint.mdlus
Removed weights_only=False from module.py
Updated README: clarified dataset link, removed redundant flags, added multi-GPU training notes, noted single-device inference, updated contact section, removed License section
Updated config.yaml: added comments for device (ignored in distributed mode), resume_from_source/resume_from_adapt (checkpoint paths), and listed options for training_loss/testing_loss ('l1', 'l2')
Migrated to PhysicsNeMo checkpointing in data_processor.py

All changes are committed and ready for review.

…tion - Add tqdm progress bars for epoch, batch, and evaluation loops - Fix autoregressive skip connection to handle correct tensor shapes (B, out_channels, n_out) -> (B, n_out, out_channels) - Add autoregressive config option to model config - Update pretraining and inference to handle autoregressive parameter correctly

Align the GINOWrapper latent path with native neuralop GINO semantics so mesh-to-grid projection uses latent query points, not mesh coordinates. Also filter loss-only fields from trainer and domain-adaptation model forwards and add regressions covering latent-grid shape handling and target-key filtering.

copy-pr-bot · 2026-04-27T16:55:05Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

MehdiTaghizadehUVa and others added 30 commits April 3, 2025 18:54

Added HydroGraphNet and its corresponding files

313d4d0

README file updated.

5371dc6

Create animations

24c7c51

Add files via upload

29ec0ce

Delete examples/weather/flood_modeling/hydrographnet/outputs directory

3ef4bab

Create sample_animation.gif

b02726b

Add files via upload

26d700c

README file updated.

c1e764a

Merge remote-tracking branch 'origin/main'

07592ee

README file updated.

0ec972e

Update README.md

3b9d7ce

Merge branch 'main' into main

cf4137c

Address review comments:

1d38b30

- Moved sample_animation.gif to docs/img - Removed functions using pickle due to security concerns - Removed duplicate import of KolmogorovArnoldNetwork in __init__.py

Merge remote-tracking branch 'origin/main'

a73cf67

formatting

cc01767

Merge branch 'main' into main

151c830

add unit tests for HydroGraphDataset, KAN layer, and MeshGraphKAN

9ca0b8f

formatting

0c5dbe1

fix meshgraphcan

5aa7b62

update test data dir

453fa55

fix timeout and tests

87b18ee

fix imports

aadd743

fix tests.

61d6605

file for fix tests.

2cbc592

Merge branch 'main' into main

037e4e0

Merge branch 'main' into main

06bb3c5

formatting

625405d

Merge branch 'NVIDIA:main' into main

b3502da

Merge branch 'NVIDIA:main' into main

ac98fb4

Merge branch 'NVIDIA:main' into main

b60f9bf

mnabian reviewed Dec 11, 2025

View reviewed changes

CharlelieLrt requested changes Dec 11, 2025

View reviewed changes

mnabian reviewed Dec 11, 2025

View reviewed changes

MehdiTaghizadehUVa added 5 commits January 9, 2026 17:52

Remove weights_only=False from torch.load calls in module.py

1539d9f

Clarify that data generation scripts are provided, not the dataset it…

0c4944b

…self

Remove redundant Hydra flags from README commands (config-path and co…

01b5540

…nfig-name are defaults)

Address review comments: add config comments, make loss configurable,…

fb1d67b

… update distributed docs

MehdiTaghizadehUVa requested a review from mnabian January 14, 2026 22:52

MehdiTaghizadehUVa added 2 commits January 14, 2026 17:53

Merge branch 'main' into main

2e427ac

Fix HydroGraphNet data flow and training correctness

cdf23d4

MehdiTaghizadehUVa requested a review from coreyjadams as a code owner April 8, 2026 02:33

MehdiTaghizadehUVa added 5 commits April 17, 2026 16:10

Harden FloodForecaster data, training, and checkpoint workflows

e74cc72

Restore HydroGraphNet README and correct dataset references

c94d7ab

Merge FloodForecaster hardening into main

292a40e

Consolidate FloodForecaster tests and docs

7b40e8c

MehdiTaghizadehUVa added 2 commits April 28, 2026 16:13

Fix FloodForecaster smoke pipeline regressions

e637881

Restore FloodForecaster README animations

98139df


		For questions, feedback, or collaborations:

		- Mehdi Taghizadeh – <jrj6wm@virginia.edu>

		@@ -0,0 +1,338 @@
		# FloodForecaster: A Domain-Adaptive Geometry-Informed Neural Operator Framework for Rapid Flood Forecasting


		raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")

		def save_checkpoint(self, save_folder: str, save_name: str) -> None:

Conversation

MehdiTaghizadehUVa commented Dec 10, 2025

PhysicsNeMo Pull Request

Description

Key Features

Components Added

Checklist

Dependencies

Review Process

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MehdiTaghizadehUVa commented Jan 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

copy-pr-bot Bot commented Apr 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

MehdiTaghizadehUVa commented Jan 13, 2026 •

edited

Loading