Error Control and Loss Functions for the Deep Learning Inversion of Borehole Resistivity Measurements

Mostafa Shahriari, David Pardo, J. A. Rivera, Carlos Torres-Verdín, Artzai Picón, Javier Del Ser, S. Ossandón, Victor M. CaloarXiv 2020 · OpenReview

TL;DR

This paper tackles the challenge of accurately inverting borehole resistivity measurements using deep learning (DL) for real-time geosteering applications. The authors demonstrate that the choice of loss function and error control mechanisms are crucial for achieving reliable inversion results, proposing novel approaches that outperform traditional data misfit methods.

What Problem Does This Paper Address?

Geosteering, a critical operation in oil and gas exploration, relies on real-time inversion of borehole resistivity measurements to guide drilling trajectories. Traditional inversion methods are often too slow for this real-time requirement. Deep learning (DL) offers a promising solution due to its speed after training, but its application to inverse problems like resistivity inversion faces significant challenges. Specifically, inverse problems are often ill-posed, meaning multiple subsurface models can produce the same measurements. Standard DL approaches using simple data misfit loss functions can lead to DNNs that approximate the inverse operator poorly, yielding unreliable results and potentially compromising geosteering decisions. This paper addresses the need for robust and accurate DL-based inversion by investigating the impact of loss functions and error control strategies.

Methods

The authors propose and evaluate several deep neural network (DNN) approaches for borehole resistivity inversion. Key methods include:

  • Encoder-Decoder Architecture: This architecture is used to simultaneously learn both the forward and inverse operators.
  • Two-Step Approach: This method first trains a DNN to approximate the forward function and then uses this trained forward model to train a separate DNN for the inverse operator.
  • Loss Functions: The paper compares several loss functions, including:
  • Data Misfit: A standard loss function that minimizes the difference between predicted and actual measurements. The authors show this is insufficient for ill-posed inverse problems.
  • Measurement Misfit: A loss function that aims to ensure the forward model's output, when passed through the inverse model, reconstructs the original measurements.
  • Encoder-Decoder Loss: A combined loss that trains both forward and inverse DNNs simultaneously.
  • Two-Step Loss: A sequential training approach as described above.
  • Regularization: A regularization term, based on the data misfit loss, is introduced to guide the DNN towards solutions present in the training dataset.
  • Data Preprocessing: Techniques like logarithmic coordinate changes and outlier removal are employed to improve data quality and comparability.
  • Error Control: The paper emphasizes controlling the relative error of rescaled variables to ensure comparable contributions to the loss function across different geophysical quantities.

Datasets

The study utilizes synthetic datasets generated for two main examples:

  • Example A: Model problem with known analytical solution: A simplified problem where the forward function is squaring a number and the inverse has two branches (square root). This is used for theoretical illustration.
  • Example B: Inversion of borehole resistivity measurements: This example uses a more realistic 1D layered Earth model.
  • Example B.1: Uses measurements from a single logging position.
  • Example B.2: Uses measurements from a full logging trajectory (65 positions).

For training, validation, and testing, the authors use 80%, 10%, and 10% of the generated samples, respectively. The dataset for Example B.1 consists of 1,000,000 samples.

Key Insights and Ideas

  • Data Misfit is Insufficient for Ill-Posed Problems: A simple data misfit loss function, common in many DL applications, fails to produce accurate inversions for ill-posed problems like borehole resistivity inversion. It can lead to DNNs that satisfy the loss but don't represent physically meaningful solutions.
  • Loss Function Choice is Critical: The selection of the loss function significantly impacts the quality and reliability of the inversion results. Encoder-Decoder and two-step approaches, which consider the relationship between forward and inverse operators, perform much better.
  • Regularization Can Be a Double-Edged Sword: While regularization can improve R² scores and provide more confident-looking results, it can also mask alternative valid solutions and hinder uncertainty quantification.
  • Training Data Sufficiency is Paramount: The performance of DL inversion is highly dependent on the size and representativeness of the training dataset. Insufficient data can lead to DNNs that are accurate on the training set but fail on unseen data.
  • Error Control is Essential for Comparability: Proper preprocessing and error control mechanisms are necessary to ensure that different geophysical variables (e.g., resistivity, distance) contribute appropriately to the loss function, preventing one variable from dominating the optimization.

Key Results

  • The data misfit loss function resulted in poor inversion accuracy for the analytical model problem (Figure 5).
  • The measurement misfit loss function recovered one of the possible solutions for the analytical problem but had critical limitations during training (Figure 6).
  • The Encoder-Decoder and two-step loss functions, when applied to the analytical problem, yielded accurate inversion results (Figures 7 and 8).
  • For the realistic borehole resistivity inversion (Example B.1), the Encoder-Decoder approach without regularization showed poor performance for geosignal forecasts, while with regularization, results were better but potentially misleading (Table 2, Figure 10).
  • The two-step approach without regularization showed improved results for geosignal forecasts compared to the Encoder-Decoder without regularization, but still highlighted the challenge of non-unique solutions.
  • Cross-plot analysis (specifically type 3) revealed significant discrepancies when the DNN predicted solutions were outside the training data space, even if the overall loss was low.
  • The paper demonstrated that the proposed DL methods can achieve real-time inversion with limited online computational cost, making them suitable for geosteering applications.

Limitations and Future Work

Limitations:

  • Data Sufficiency: The performance of the DL approach is highly dependent on the availability of a sufficiently large and representative training dataset. This becomes particularly challenging for more complex 2D and 3D Earth models.
  • Discontinuous Functions: DNNs struggle to accurately reproduce discontinuous functions, which can occur in geophysical data.
  • Out-of-Distribution Performance: DNNs can produce incorrect solutions when evaluated on data samples outside their training distribution.
  • Difficulty in Identifying Poor Results: It can be challenging to pinpoint the exact cause of poor inversion results, as it could stem from various factors like loss function choice, architecture, regularization, or data quality.

Future Work:

  • Investigate different DNN architectures, potentially using automated architecture search techniques.
  • Design optimal sampling techniques for training data.
  • Develop new regularization techniques.
  • Explore Bayesian DNNs for uncertainty quantification.
  • Utilize transfer learning for higher spatial dimensions to reduce data requirements.
  • Evaluate the performance of the approach with noisy measurements.
  • Design proper measurement acquisition systems and adequate Earth model parameterizations.

Summary updated

Abstract

Deep learning (DL) is a numerical method that approximates functions. Recently, its use has become attractive for the simulation and inversion of multiple problems in computational mechanics, including the inversion of borehole logging measurements for oil and gas applications. In this context, DL methods exhibit two key attractive features: a) once trained, they enable to solve an inverse problem in a fraction of a second, which is convenient for borehole geosteering operations as well as in other real-time inversion applications. b) DL methods exhibit a superior capability for approximating highly-complex functions across different areas of knowledge. Nevertheless, as it occurs with most numerical methods, DL also relies on expert design decisions that are problem specific to achieve reliable and robust results. Herein, we investigate two key aspects of deep neural networks (DNNs) when applied to the inversion of borehole resistivity measurements: error control and adequate selection of the loss function. As we illustrate via theoretical considerations and extensive numerical experiments, these interrelated aspects are critical to recover accurate inversion results.
BibTeX
@misc{shahriari2020error,
  author = {Mostafa Shahriari and David Pardo and J. A. Rivera and Carlos Torres-Verdín and Artzai Picón and Javier Del Ser and S. Ossandón and Victor M. Calo},
  title = {{Error Control and Loss Functions for the Deep Learning Inversion of Borehole Resistivity Measurements}},
  year = {2020},
  eprint = {2005.08868},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.08868},
}
Figure