Openfwi Fwi Eval

openfwi-fwi-eval

Parameter Efficient Fine-Tuning for Deep Learning-Based Full-Waveform Inversion — Ghosal et al. (2024) (arXiv:2412.19510, 2024)

What this evaluates

Evaluates deep learning models for seismic full-waveform inversion (FWI) by predicting subsurface velocity models from seismic wavefield data. It probes the model's ability to generalize across varying geological complexities and out-of-distribution scenarios using parameter-efficient fine-tuning.

Datasets

• OpenFWI — total ?; splits: train (-1), test (-1); repo https://github.com/Kaustav546/FWI-PEFT.git

Metrics

• MAE — range: other
  • Mean Absolute Error: average of absolute differences between predicted and ground truth velocity values.
• RMSE — range: other
  • Root Mean Squared Error: square root of the average of squared differences between predicted and ground truth velocity values.
• SSIM (primary) — range: [-1, 1]
  • Structural Similarity Index Measure: evaluates perceived change in structural information between predicted and reference velocity maps, ranging from -1 to 1 where 1 indicates perfect similarity.

Input / output format

Input: Seismic wavefield recordings (shot gathers or time-series data)

Output: Predicted 2D subsurface velocity model/map

Scoring recipe

def compute_metrics(pred, gt):
    mae = np.mean(np.abs(pred - gt))
    rmse = np.sqrt(np.mean((pred - gt) ** 2))
    ssim = compute_ssim(pred, gt)  # Standard SSIM implementation
    return {'MAE': mae, 'RMSE': rmse, 'SSIM': ssim}

Common pitfalls

• SSIM measures structural similarity rather than pixel-wise accuracy, so a high SSIM does not guarantee low MAE or RMSE.
• Performance heavily depends on dataset complexity (spatial information/gradient magnitude); models may underperform on simple geological structures while excelling on complex ones.
• Out-of-distribution (OOD) evaluation requires strict train/test splits where the test dataset has a significant statistical distribution shift from the training data.

Evidence (verbatim from paper)

Three evaluations used for this study are: MAE, RMSE, and SSIM.

Citation

@misc{ghosal2024parameter,
  title={Parameter Efficient Fine-Tuning for Deep Learning-Based Full-Waveform Inversion},
  author={Ghosal et al. (2024)},
  year={2024},
  note={arXiv:2412.19510}
}

• arXiv: 2412.19510