This project explores the challenge of misalignment between well logs and seismic data due to time shifts caused by inaccurate time-depth relationships (TDR). Such misalignment—where seismic peaks and troughs do not align with well-log events—poses a serious problem for supervised learning models, preventing them from learning correct input-label relationships.
To address this, we simulate realistic time-shift scenarios using synthetic data and train a neural network to estimate and correct these shifts.
- Model: Marmousi velocity model
- Domain: Depth-domain simulation
- A velocity model in the depth domain is used to compute a ground-truth TDR.
- The depth sampling interval is set to
1.25 km(unit illustrative). - Synthetic seismic traces are generated based on this true TDR.
- Perturb the true TDR to obtain a time-shifted version.
- Use the shifted TDR to generate synthetic seismic traces that exhibit realistic time shifts.
- The difference between the original and shifted TDR is used as the label for supervised learning.
This process generates a dataset of aligned and misaligned seismic traces to train a model for time-shift estimation and correction.
Figure 1. True TDR and shifted TDRs.
Figure 2. True record and shifted records.
To develop a neural network capable of learning the mapping from misaligned seismic traces to accurate time-shift corrections, enabling better well-to-seismic integration in real-world applications.
If you find this project useful, please give it a star!