TL;DR
This paper introduces PhResNet, a novel physics-driven deep learning approach for Direct Current (DC) resistivity inversion. By embedding physical laws into the network and using unsupervised learning, PhResNet overcomes the need for real-world resistivity models, making it more practical for field applications. The method demonstrates accurate geological target identification in both synthetic and real-world tunnel survey data.
What Problem Does This Paper Address?
Geophysical surveys using the Direct Current (DC) resistivity method are crucial for understanding underground geological structures. A key step in these surveys is "inversion," which reconstructs the subsurface resistivity model from measured data. However, this inversion process is notoriously "ill-posed," meaning small errors in data can lead to large errors in the model, making accurate interpretation difficult. Traditional methods often rely on initial guesses and can get stuck in local optima.
While deep learning (DL) offers powerful tools for solving inverse problems, existing DL methods for resistivity inversion are typically "data-driven" and require extensive "labeled" data – meaning they need pairs of survey data and their corresponding actual resistivity models. Obtaining these real-world resistivity models is extremely challenging and expensive, limiting the applicability of purely data-driven DL. This paper addresses this gap by developing a DL method that doesn't need real resistivity models for training, making it more practical for real-world geophysical surveys.
Methods
The core innovation is PhResNet, a physics-driven deep learning network for unsupervised resistivity inversion. Here's how it works:
- Physics-Informed Network: Instead of relying solely on data, PhResNet embeds the physical laws governing electric field propagation directly into the neural network architecture. This is achieved by incorporating a forward modeling process within the network. This means the network learns to predict a resistivity model, then uses physics to predict what the survey data should look like for that model.
- Unsupervised Learning: The network is trained using a data misfit as the loss function. This means it learns by minimizing the difference between the predicted data (from its forward model) and the observed survey data. This eliminates the need for ground-truth resistivity models during training.
- Dynamic Smoothness Constraint: To further stabilize the ill-posed inversion problem and improve convergence, a dynamic smoothness constraint is applied. This constraint's strength is adjusted during training, being stronger initially to guide the network and weaker later to allow for finer detail.
- Transfer Learning: To adapt the network trained on synthetic data to real-world scenarios where data is scarce, a novel transfer learning strategy is employed. This combines the benefits of "full fine-tuning" (updating all network layers) and "linear probing" (retraining only the final layer) to effectively leverage pre-trained knowledge and adapt to new, limited real-world data.
- Network Architectures (PhResNet-i and PhResNet-n): Two specific network architectures are designed:
- PhResNet-i (for image data): Uses a U-Net architecture, suitable for spatially correlated data like apparent resistivity from surface surveys.
- PhResNet-n (for non-image data): Employs a fully connected neural network (FCNN) for data like potential measurements from boreholes, incorporating specific data preprocessing steps to handle the complex electrode configurations.
Datasets
- Synthetic Datasets: The paper utilizes synthetic datasets generated through forward modeling for both 3D surface ERT and 2D cross-hole ERT scenarios. These datasets are used for training, validation, and testing the PhResNet models.
- Real-World Field Data: The proposed method is validated on real-world data collected from an advanced tunnel survey. This data is used in conjunction with the transfer learning approach.
Key Insights and Ideas
- Physics as a Regularizer: Embedding physical laws (forward modeling) into a DL network acts as a powerful regularizer, guiding the inversion process and mitigating the ill-posed nature of the problem, even without ground-truth models.
- Unsupervised Learning for Practicality: The shift to unsupervised learning, driven by data misfit, is a significant step towards making DL inversion methods practical for real-world geophysical surveys where labeled data is scarce.
- Transfer Learning for Data Scarcity: The proposed transfer learning strategy effectively bridges the gap between synthetic training data and limited real-world field data, enabling the application of the trained models to practical engineering problems.
- Differentiated Architectures: Recognizing that different types of geophysical data (image vs. non-image) require tailored network architectures (U-Net vs. FCNN) is crucial for effective feature extraction and inversion.
Key Results
- Accurate Target Localization: PhResNet successfully located and imaged two low-resistivity anomalies in a 2D cross-hole ERT synthetic test with an error of only 0.1 m, outperforming traditional linear inversion which struggled to differentiate them.
- Realistic Resistivity Values: The inversion results for both synthetic and real-world data showed resistivity values close to the actual values of the targets.
- Effective Field Application: The method was successfully applied to a challenging field survey in a tunnel excavation, where inversion results were consistent with geological conditions revealed by excavation, identifying potential water-related anomalies.
- Speed: Once trained, the network offers very fast inversion times (around 1 second), making it suitable for large-scale surveys.
Limitations and Future Work
- Data Noise and Resolution: The paper acknowledges that in the field survey, high data noise led to lower imaging resolution, limiting the ability to image anomalies in fine detail.
- Transfer Learning Limitations: The effectiveness of transfer learning was noted as being limited by the scarcity of data in the early stages of the project.
- Computational Cost: While inference is fast, training the network requires millions or even billions of forward models, which is computationally intensive compared to traditional linear methods.
- Future Work: The authors suggest accumulating more real survey data for future studies and improving the transfer learning by building a high-quality sample library. They also believe the method can be further applied to real survey scenarios.
Summary updated