README.md

---
license: apache-2.0
task_categories:
  - image-to-image
tags:
  - physics
  - cfd
  - digital-twin
  - datacenter
  - surrogate-model
  - neural-operator
  - nvidia
  - physicsnemo
pretty_name: "Boreas: Datacenter Digital Twin Surrogate Predictions"
size_categories:
  - 1K<n<10K
---

# Boreas Data — Datacenter Digital Twin Surrogate Predictions

Prediction outputs from five neural surrogate models trained on [NVIDIA PhysicsNeMo-Datacenter-CFD](https://huggingface.co/datasets/nvidia/PhysicsNeMo-Datacenter-CFD) for **Project Boreas** (SJSU MSDA capstone, Spring 2026).

This dataset powers the [Boreas Omniverse Kit operator console](https://github.com/irangareddy/kit-app-template) — a custom NVIDIA Omniverse application with an LLM agent (GPT-4o) and Apple Vision Pro AR streaming.

## Contents

| Folder | What | Count | Size each |
|--------|------|-------|-----------|
| test_data/inputs/ | 10-channel input tensors (80x96x960) | 10 | 282 MB |
| test_data/targets/ | 5-channel CFD ground truth (80x96x960) | 10 | 141 MB |
| outputs/predictions/ | Model predictions (5-ch, 80x96x960) | 50 | 141 MB |
| stl/ | OpenFOAM datacenter geometry | 6 | < 1 MB |
| results/ | Evaluation metrics JSON | 2 | < 1 MB |

**Total: 79 files, ~11 GB**

## Five surrogate models

| Model | Family | Params | T MAE (C) | Inference (ms) |
|-------|--------|--------|-----------|----------------|
| **U-Net** | Spatial (CNN) | 22.6M | **0.205** | 177 |
| PI-U-Net | Spatial + physics | 344K | 0.244 | 87 |
| FNO | Spectral (FFT) | 28.3M | 0.489 | 636 |
| PI-FNO | Spectral + physics | 28.3M | 0.505 | 674 |
| Transolver | Transformer | 545K | 0.598 | 4,972 |

Trained at full resolution (960x96x80 = 7.37M points) on NVIDIA GB10 (Grace Blackwell, 128 GB).

## File naming

Predictions: `outputs/predictions/sample_NNNN_MODEL.npy`

- NNNN = 0000-0009 (from PhysicsNeMo test split)
- MODEL = unet, fno, pifno, pi_unet, or transolver
- Shape: (5, 80, 96, 960) — channels: Ux, Uy, Uz, T, p (Z-score normalized)

## Denormalization

| Field | Mean | Std | Unit |
|-------|------|-----|------|
| T | 39.0 | 4.0 | C |
| Ux, Uy, Uz | 1.5984 | 1.3656 | m/s |
| p | 6.1227 | 4.1660 | Pa |

To convert: `T_celsius = prediction[3] * 4.0 + 39.0`

## Quick start with Kit app

```bash
git lfs install
git clone https://huggingface.co/datasets/irangareddy/boreas-data ~/boreas-data
git clone https://github.com/irangareddy/kit-app-template.git
cd kit-app-template
.\repo.bat build
.\run_streaming.bat
```

## Quick start in Python

```python
import numpy as np

gt = np.load("test_data/targets/sample_0000.npy")
pred = np.load("outputs/predictions/sample_0000_unet.npy")

T_gt = gt[3] * 4.0 + 39.0
T_pred = pred[3] * 4.0 + 39.0
print(f"Temperature MAE: {abs(T_pred - T_gt).mean():.3f} C")
```

## Source

Derived from [nvidia/PhysicsNeMo-Datacenter-CFD](https://huggingface.co/datasets/nvidia/PhysicsNeMo-Datacenter-CFD) (Apache 2.0).

## Citation

```bibtex
@misc{boreas2026,
  title={Evaluating Neural Operator Architectures for Datacenter Digital Twin Surrogate Models},
  author={Nukala, Sai Ranga Reddy and Kankanala, Shaila Reddy and Singh, Sadhvi and Jonnalagadda, Akshith Reddy and Shim, Simon},
  year={2026},
  institution={San Jose State University},
  url={https://github.com/irangareddy/298AB}
}
```

## Links

- Research: [github.com/irangareddy/298AB](https://github.com/irangareddy/298AB)
- Kit app: [github.com/irangareddy/kit-app-template](https://github.com/irangareddy/kit-app-template)
- Source data: [nvidia/PhysicsNeMo-Datacenter-CFD](https://huggingface.co/datasets/nvidia/PhysicsNeMo-Datacenter-CFD)