Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhanced 3D Gravity Inversion Using ResU-Net with Density Logging Constraints: A Dual-Phase Training Approach

Siyuan Dong, Jinghuai Gao, Shuai Zhou, Baohai Wu, Hongfa Jia

TL;DR

This work tackles non-uniqueness and depth-resolution challenges in gravity inversion by marrying physics-informed priors with deep learning. A ResU-Net-based dual-phase training framework is introduced: Net-I trains in a depth-weighted, weighted-density space to boost depth resolution, and Net-II is fine-tuned with sparse density well-logging constraints and 3-D total variation to enforce geological plausibility. Across synthetic benchmarks and a Bishop Model, the approach yields higher MA and DA than unconstrained DL and offers competitive advantages over focusing inversion methods, with stronger performance in field data from the San Nicolas area. The combination of depth weighting, well-logging masking, and a two-stage training scheme provides a robust, transferable workflow for integrating multi-physics constraints into gravity inversion, enhancing both accuracy and interpretability.

Abstract

Gravity exploration has become an important geophysical method due to its low cost and high efficiency. With the rise of artificial intelligence, data-driven gravity inversion methods based on deep learning (DL) possess physical property recovery capabilities that conventional regularization methods lack. However, existing DL methods suffer from insufficient prior information constraints, which leads to inversion models with large data fitting errors and unreliable results. Moreover, the inversion results lack constraints and matching from other exploration methods, leading to results that may contradict known geological conditions. In this study, we propose a novel approach that integrates prior density well logging information to address the above issues. First, we introduce a depth weighting function to the neural network (NN) and train it in the weighted density parameter domain. The NN, under the constraint of the weighted forward operator, demonstrates improved inversion performance, with the resulting inversion model exhibiting smaller data fitting errors. Next, we divide the entire network training into two phases: first training a large pre-trained network Net-I, and then using the density logging information as the constraint to get the optimized fine-tuning network Net-II. Through testing and comparison in synthetic models and Bishop Model, the inversion quality of our method has significantly improved compared to the unconstrained data-driven DL inversion method. Additionally, we also conduct a comparison and discussion between our method and both the conventional focusing inversion (FI) method and its well logging constrained variant. Finally, we apply this method to the measured data from the San Nicolas mining area in Mexico, comparing and analyzing it with two recent gravity inversion methods based on DL.

Enhanced 3D Gravity Inversion Using ResU-Net with Density Logging Constraints: A Dual-Phase Training Approach

TL;DR

This work tackles non-uniqueness and depth-resolution challenges in gravity inversion by marrying physics-informed priors with deep learning. A ResU-Net-based dual-phase training framework is introduced: Net-I trains in a depth-weighted, weighted-density space to boost depth resolution, and Net-II is fine-tuned with sparse density well-logging constraints and 3-D total variation to enforce geological plausibility. Across synthetic benchmarks and a Bishop Model, the approach yields higher MA and DA than unconstrained DL and offers competitive advantages over focusing inversion methods, with stronger performance in field data from the San Nicolas area. The combination of depth weighting, well-logging masking, and a two-stage training scheme provides a robust, transferable workflow for integrating multi-physics constraints into gravity inversion, enhancing both accuracy and interpretability.

Abstract

Gravity exploration has become an important geophysical method due to its low cost and high efficiency. With the rise of artificial intelligence, data-driven gravity inversion methods based on deep learning (DL) possess physical property recovery capabilities that conventional regularization methods lack. However, existing DL methods suffer from insufficient prior information constraints, which leads to inversion models with large data fitting errors and unreliable results. Moreover, the inversion results lack constraints and matching from other exploration methods, leading to results that may contradict known geological conditions. In this study, we propose a novel approach that integrates prior density well logging information to address the above issues. First, we introduce a depth weighting function to the neural network (NN) and train it in the weighted density parameter domain. The NN, under the constraint of the weighted forward operator, demonstrates improved inversion performance, with the resulting inversion model exhibiting smaller data fitting errors. Next, we divide the entire network training into two phases: first training a large pre-trained network Net-I, and then using the density logging information as the constraint to get the optimized fine-tuning network Net-II. Through testing and comparison in synthetic models and Bishop Model, the inversion quality of our method has significantly improved compared to the unconstrained data-driven DL inversion method. Additionally, we also conduct a comparison and discussion between our method and both the conventional focusing inversion (FI) method and its well logging constrained variant. Finally, we apply this method to the measured data from the San Nicolas mining area in Mexico, comparing and analyzing it with two recent gravity inversion methods based on DL.
Paper Structure (12 sections, 17 equations, 18 figures, 9 tables)

This paper contains 12 sections, 17 equations, 18 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Testing Model Inversion Results
  4. Test-I: Syncline Structure Model
  5. Test-III: Bishop Model
  6. Test-IV: Comparison with FI Methods
  7. Inversion of Real Data
  8. Discussion
  9. Conclusion
  10. Data Availability Statement
  11. Appendix A
  12. Well Logging Constrained Focusing Inversion

Figures (18)

  • Figure 1: Visualization of the sensitivity matrix (the example grid $\Delta$ = 100m). During forward modeling, the weight distribution corresponding to a surface observation point at different depths is as follows: (a) ${\bf A}$ without weighting, (b) ${\bf A}_w$ with depth weighting.
  • Figure 2: The selection methods for the well logging space: well logging (a) located within the grid, (b) located on the measurement line, (c) densely distributed at the intersection of measurement lines, and (d) sparsely distributed at the intersection of measurement lines. (e) The 3-D well logging space. (f) FG-Mask in well logging space ${\bf m}_L$.
  • Figure 3: (a) The flowchart of the dual-phase training approach, (b) the ECA block, and (c) the detailed architecture of the ResU-Net network.
  • Figure 4: The randomly generated models used in (a) Test-I, (b) Test-II, and (c) Test-III for NN training.
  • Figure 5: Schematic diagram illustrating the growth laws of (a) GPU and (b) RAM computational costs with increasing model resolution. Blue and orange icons represent the data under the purely data-driven and depth weighting scenarios, respectively. Circular and triangular icons denote the empirically measured data and the theoretically extrapolated data, respectively.
  • ...and 13 more figures