Add EddyFormer

[mrlazy1708]

PhysicsNeMo Pull Request

Description

This PR implements the EddyFormer model (https://arxiv.org/abs/2510.24173) from NeurIPS 2025, together with a three-dimensional isotropic turbulence benchmark dataset.

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

scipy is added as an additional dependency. It is used to obtain quadrature weights.

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.

[greptile-apps]

Greptile Summary

• Implements EddyFormer model architecture from NeurIPS 2025 paper for 3D isotropic turbulence simulation with spectral element method
• Adds new scipy dependency and complete example implementation including dataset download, training script, and configuration
• Contains several critical implementation issues including incorrect tuple extension, unused parameters, and potential performance bottlenecks

Important Files Changed

Filename	Overview
physicsnemo/models/eddyformer/_basis.py	New file implementing Legendre basis functions; contains critical tuple extension bug that could cause runtime errors
physicsnemo/models/eddyformer/eddyformer.py	New file implementing main EddyFormer model class; includes unused scale parameter and hardcoded interpolation method
physicsnemo/models/eddyformer/sem_conv.py	New file implementing SEM convolution operations; complex implementation with caching and multiprocessing optimization
examples/cfd/isotropic_eddyformer/train_ef_isotropic.py	New training script with anomaly detection enabled during training which could significantly impact performance
physicsnemo/models/eddyformer/_datatype.py	New file implementing SEM data structures; contains device handling inconsistencies and complex tensor operations

[greptile-apps]

_{12 files reviewed, 17 comments}

_{Edit Code Review Agent Settings | Greptile
React with 👍 or 👎 to share your feedback on this new summary format}

[mrlazy1708]

What's new since last review:

1. C0-continuous basis is added. Now the model reproduces full results on the 3D HIT case;
2. Torch compile, DDP, and AMP training are all supported by model and enabled using DistributedManager and StaticCaptureTraining. AMP uses the default torch.float16 is able to accelerate the training by 2x on four A100s using DDP.

@CharlelieLrt