TL;DR
This paper introduces a novel "physics-driven deep-learning" (PhyDNN) framework to solve the challenging geosteering inversion problem in directional drilling. By integrating a physical forward model into a deep neural network, PhyDNN achieves faster, more accurate, and robust predictions compared to traditional methods and purely data-driven deep learning approaches.
What Problem Does This Paper Address?
Directional drilling relies heavily on geosteering to navigate wells within oil and gas reservoirs. This process involves solving a "geosteering inversion" problem, which aims to determine the subsurface formation properties (like resistivity) from measurements taken by logging-while-drilling (LWD) tools. Traditional methods for this inversion are often slow, sensitive to initial guesses, and can get stuck in suboptimal solutions (local minima). Industrial solutions like lookup tables offer speed but lack accuracy and scalability. Purely data-driven deep learning methods, while fast, can struggle with accuracy and robustness, especially when dealing with noisy data or complex geological structures. This paper addresses the need for a geosteering inversion method that is both fast and accurate, overcoming the limitations of existing techniques.
Methods
The core innovation is the Physics-Driven Deep-Learning Network (PhyDNN). This framework combines a deep neural network (DNN) with a physics-based forward model.
- DNN as an Encoder: The DNN acts as an "encoder," learning to map the measured data (d) to the earth model parameters (m). The architecture is inspired by the VGG16 network, utilizing 1D convolutional layers, pooling, and instance normalization.
- Physics-Based Forward Model as a Decoder: Instead of a learned decoder, PhyDNN uses a pre-defined, differentiable physical forward model (F) that simulates the LWD tool's response (d) given an earth model (m). This model is crucial for enforcing physical consistency.
- Physics-Driven Loss Function: The training objective is a combined loss function that includes:
- Model Misfit (Lml): Measures the difference between the predicted earth model (m̂) from the DNN and the true earth model (m).
-
Data Misfit (Ldl): Measures the difference between the actual measurements (d) and the synthetic measurements generated by the physical forward model (F) using the DNN's predicted earth model (F(D(d))). This dual-loss approach, guided by both data and physics, helps regularize the training and improve robustness.
-
Numerical Jacobian Estimation: For the forward model, which is differentiable but lacks an analytical Jacobian, numerical estimation is used to compute gradients for backpropagation.
Datasets
The study utilizes a synthetic dataset generated from a simulated 1D layered earth model. This model is described by resistivity values (R1, R2, R3) and boundary depths (Dup, Ddn). The dataset was created by sampling these parameters within defined ranges and using the physical forward model to generate corresponding tool responses. A large lookup table (approximately 1.61 GB) was generated for training and testing, with 90% of the data used for training and 10% for validation. No new benchmark datasets were introduced; rather, the generated synthetic dataset served as the basis for evaluation.
Key Insights and Ideas
- Physics as a Regularizer: Integrating a physical forward model into the deep learning framework acts as a powerful regularizer. This prevents the DNN from learning spurious correlations from the data alone and ensures that the predictions are physically plausible.
- Balancing Data and Physics: The dual-loss function (model misfit + data misfit) is key. It allows the network to learn from both the observed data and the underlying physical laws, leading to more accurate and reliable inversions.
- Robustness to Noise: The PhyDNN framework, especially when augmented with noise injection during training (NR-PhyDNN), demonstrates superior robustness to noisy measurements compared to purely data-driven DNNs. This is critical for real-world applications where data quality can vary.
- Avoiding Local Minima: The physics-driven approach helps guide the optimization process, making it less susceptible to getting trapped in local minima, a common problem with traditional iterative methods.
Key Results
- Accuracy: PhyDNN consistently achieved lower model and data misfits compared to lookup tables and data-driven DNNs across various test cases. Its performance was comparable to LMA (Levenberg-Marquardt algorithm) in terms of inversion accuracy.
- Speed: PhyDNN is significantly faster than LMA and lookup table methods. For interpreting 80 logging points, PhyDNN took only 0.011 seconds, compared to 62.315 seconds for a 2M lookup table and 169.754 seconds for LMA.
- Memory Efficiency: PhyDNN requires substantially less memory (7 MB) than large lookup tables (1.61 GB for 2M), making it more practical for deployment.
- Robustness: The noise-robust version of PhyDNN (NR-PhyDNN) showed excellent performance even with significant noise levels, outperforming its non-augmented counterpart and the data-driven DNN.
Limitations and Future Work
- Computational Cost of Forward Model: While PhyDNN is faster than iterative methods, the forward model's computation can still be a bottleneck. The paper mentions using OpenMP for parallelization to accelerate CPU computation.
- Generalizability to Complex Geometries: The current study focuses on 1D layered earth models. Extending the framework to more complex 3D geological structures would be a significant future direction.
- Real-World Data Validation: The experiments were conducted using synthetic data. Future work should involve validating the PhyDNN on actual field data to confirm its performance in real-world scenarios.
- Downhole Computing Potential: The authors suggest that PhyDNN has the potential to enable downhole computing for LWD tools, which would allow for real-time inversion and decision-making directly at the drill bit. This is a promising avenue for future research and development.
Summary updated