TL;DR
This paper introduces a novel machine learning approach for detecting geological features directly from raw seismic data, overcoming the computational bottlenecks of traditional methods. By combining kernel ridge regression with a randomized Nyström approximation, the authors achieve significant speedups (up to 1000x) while maintaining comparable accuracy, making real-time geological feature detection a possibility.
What Problem Does This Paper Address?
Detecting subsurface geological features, like faults, from seismic data is crucial for various energy applications (geothermal, carbon sequestration, oil & gas). Traditional methods often rely on physics-driven models (like full-waveform inversion) which are computationally expensive, require significant human interpretation, and can be limited by data coverage. Existing machine learning approaches often work on processed seismic data (post-stack), requiring an initial inversion or migration step, which can lose information and still be computationally demanding. The core challenge is to efficiently and accurately extract geological features directly from the vast amounts of raw, pre-stack seismic data.
Methods
The authors propose a data-driven approach using Kernel Ridge Regression (KRR). KRR is a powerful supervised learning technique that can model non-linear relationships between seismic measurements and geological features. However, KRR's computational cost (specifically, forming and inverting an $n \times n$ kernel matrix, where $n$ is the number of data samples) becomes prohibitive for the millions of data points typically found in seismic datasets.
To address this, they incorporate a randomized matrix sketching technique, specifically the Nyström method. The Nyström method approximates the large kernel matrix with a smaller, low-rank matrix. This is achieved by randomly sampling a subset of columns from the original kernel matrix, effectively reducing the dimensionality of the problem. This allows for a significant reduction in computational time and memory requirements without substantially sacrificing accuracy. The overall algorithm involves training a KRR model using this Nyström approximation on pre-stack seismic data to learn a mapping function that predicts geological features.
Datasets
The authors created a synthetic dataset for training and evaluation. This dataset consists of:
- 60,000 velocity models: Each model is a 100x100 grid representing subsurface properties. These models vary in:
- Horizontal offset of fault zones (30-70 grids)
- Dipping angle (25°-165°)
- Number of layers (3-5)
- Layer thickness (5-80 grids)
- Layer velocity (3000-5000 m/s)
- Synthetic seismic measurements: Generated by forward modeling on these velocity models. This includes:
- Common-shot gathers with 32 receivers.
- Ricker wavelet source with a 25 Hz center frequency.
- Time-series data of length 1,000 for each receiver.
- Feature dimension of seismic data: $3.2 \times 10^4$.
- Total volume of synthetic seismic data: $1.92 \times 10^9$.
The dataset was split, with 75% used for training and 25% for testing. Both acoustic and elastic seismic data were generated and tested.
Key Insights and Ideas
- Directly learning from pre-stack seismic data is feasible and efficient: This bypasses the need for computationally intensive pre-processing steps like migration or inversion, which are often required by other methods.
- Randomized Nyström approximation is a powerful tool for scaling kernel methods: It effectively tackles the $O(n^3)$ time and $O(n^2)$ memory complexity of traditional KRR, making it applicable to massive datasets.
- The trade-off between accuracy and computational cost is manageable: The Nyström method allows for a controlled sacrifice of accuracy for significant gains in speed and memory efficiency, with accuracy improving as the number of sampled columns ($s$) increases.
- Geological features can be characterized by offset and dipping angle: The paper demonstrates the ability to predict these key parameters directly from seismic data.
- Elastic seismic data provides more information but requires more data: While elastic models lead to more accurate results due to richer physics, they also necessitate larger training datasets to achieve comparable performance to acoustic models.
Key Results
- Significant Speedup: The proposed randomized KRR (R-KRR) method achieved speed-up ratios of ~10² to ~10³ compared to conventional KRR in a multi-core environment. For the full dataset, training time was reduced to approximately 10 seconds.
- Comparable Accuracy: The R-KRR method demonstrated comparable accuracy to conventional KRR, especially when dealing with large datasets. While conventional KRR was slightly more accurate on very small datasets, it became computationally infeasible as the dataset size grew.
- Robustness: The method showed robust performance across different Nyström sample sizes ($s=3000$ and $s=6000$) and was not significantly affected by the randomness inherent in the Nyström approximation.
- Accurate Feature Estimation: A visual example showed that the estimated geological fault parameters (angle and offset) were very close to the ground truth. For instance, an angle of 79.1° was estimated as 79.0°, and an offset of 49.0 grids was estimated as 50.0 grids.
Limitations and Future Work
- Parameter Tuning: The performance of the R-KRR method depends on tuning three parameters: the regularization parameter ($\lambda$), the kernel width parameter ($\sigma$), and the number of random samples ($s$). While heuristics and cross-validation are suggested, finding optimal values can still be challenging.
- Dataset Size and Accuracy: While the method scales well, achieving very high accuracy still requires a substantial amount of training data, especially for more complex elastic models.
- Future Work: The paper suggests that the method has great potential for real-time or near real-time detection of subsurface geological features. Further exploration into optimizing parameter selection and applying the method to more complex geological scenarios would be valuable.
Summary updated