rohithp-alt2/fno-3D-source-unmixing

3D FNO Source Unmixing Dataset

Its a dataset of 3D signals designed for seperating complex non linearly mixed fields using fnos.

Dataset Summary

• File Name: fno_3D_source.pt
• Total Samples: 20,000
• Tensor Shape: (20000, 2, 16, 16, 16)
• Data Type: torch.float32
• Uncompressed Size: ~625 MB
• Boundary Condition: I've applied a 3D Hann window envelope so the signal smoothly decays to zero at the spatial boundaries.

Signal Generation & Physics

The observed signal S is a non linear combination of 2 hidden complex Gaussian fields:

1. X̂ (Low-frequency macro drift): Fourier mode cutoffs (M_t, M_x, M_y = 2, 4, 4).
2. Ŷ (High-frequency wave burst): Fourier mode cutoffs (M_t, M_x, M_y = 6, 6, 6).

The mixing equation is defined as:

$$S=\text{Re}(\widehat{X}\cdot e^{i\cdot \text{Im}(\widehat{Y})})+\tanh (\text{Re}(\widehat{X}))\odot \sigma (\text{Re}(\widehat{Y}))S\; =\; \backslash text\{Re\}\backslash left(\backslash hat\{X\}\; \backslash cdot\; e^\{i\; \backslash cdot\; \backslash text\{Im\}(\backslash hat\{Y\})\}\backslash right)\; +\; \backslash tanh(\backslash text\{Re\}(\backslash hat\{X\}))\; \backslash odot\; \backslash sigma(\backslash text\{Re\}(\backslash hat\{Y\}))$$

How to Load

Using huggingface_hub and torch

import torch
from huggingface_hub import hf_hub_download

file_path = hf_hub_download(repo_id="YOUR_HF_USERNAME/YOUR_REPO_NAME",filename="fno_3D_source.pt",repo_type="dataset")
S = torch.load(file_path)