Nonlinear 1D Time-Domain Electromagnetic Inversion using Adaptive Differential Evolution: Application to Volcanic-Hydrothermal System

Technical Summary: Nonlinear 1D Time-Domain Electromagnetic Inversion using Adaptive Differential Evolution

Problem Statement
Time-Domain Electromagnetic (TDEM) subsurface resistivity inversion is characterized as a highly nonlinear, ill-posed optimization problem. The process is hindered by significant challenges including parameter coupling, layer equivalence (where different combinations of resistivity and thickness yield indistinguishable responses), and the prevalence of local minima. Conventional deterministic methods (e.g., Gauss–Newton, Levenberg–Marquardt) often fail in these scenarios due to their dependence on initial model estimates and susceptibility to getting trapped in local minima. While global stochastic optimization algorithms like Differential Evolution (DE) offer a solution, standard DE implementations rely on static control parameters (mutation scaling factor $F$ and crossover rate $Cr$). These fixed settings lack the flexibility to adapt to the dynamic search requirements of different evolutionary phases, often leading to premature convergence or stagnation, particularly in complex, multimodal landscapes typical of volcanic-hydrothermal systems.

Methodology
This study proposes an advanced stochastic inversion framework based on Improved Linear Population Size Reduction Success-History Based Adaptive Differential Evolution (iL-SHADE). The methodology integrates several adaptive mechanisms to balance global exploration and local exploitation without manual parameter tuning:

1. Adaptive Parameter Control: Unlike canonical DE, iL-SHADE utilizes historical memory arrays ($M_F$ and $M_{Cr}$) to store successful mutation and crossover parameters from previous generations. These memories are updated using weighted means of successful trials, allowing the algorithm to self-regulate its search behavior based on past performance.
2. Linear Population Size Reduction (LPSR): The algorithm dynamically reduces the population size ($N_p$) linearly from an initial maximum to a minimum threshold as the optimization progresses. This strategy maintains high diversity during early exploration phases and focuses computational resources on promising regions during later exploitation phases.
3. Mutation Strategy: The framework employs a current-to-pbest/1/bin mutation strategy, which selects target vectors from the top $p\%$ of the population, further balancing exploration and exploitation.
4. Forward Modeling: The TDEM forward problem is solved using a grounded horizontal electric dipole (HED) source configuration. The solution involves transforming Maxwell's equations into the Laplace domain, solving for the magnetic vector potential in layered media, and applying the Gaver–Stehfest algorithm for numerical inverse Laplace transformation to obtain the time-domain magnetic flux density response ($B_z(t)$).
5. Validation Framework: The algorithm was evaluated using:
   • Synthetic Benchmarks: Four distinct 5-layer earth models (Ascending, Bell, Bowl, and Descending resistivity profiles) to test performance across varying geological complexities.
   • Statistical Rigor: 30 independent stochastic runs per model, analyzed using non-parametric Wilcoxon rank-sum hypothesis testing ($\alpha = 0.01$) to confirm statistical significance.
   • Robustness Tests: Assessment under Gaussian noise levels up to 20% and layer mismatch scenarios (inverting 3-layer data with a 5-layer model).
   • Field Application: Validation using real TDEM data from Station UZ01 on Unzen Volcano, Japan.

Key Contributions

• Algorithmic Adaptation: This work presents the first systematic application of iL-SHADE to TDEM inversion, demonstrating that adaptive parameter control and population management significantly outperform standard DE and SHADE in geophysical contexts.
• Empirical Guidelines: The study establishes specific baseline configurations for TDEM inversion, recommending an initial population size of $N_p = 21 \times D$ (where $D$ is the model dimension) and a $p$-best range of $[0.10, 0.20]$, eliminating the need for costly trial-and-error tuning.
• Statistical Verification: By conducting 30 independent runs and applying Wilcoxon rank-sum tests, the study provides rigorous statistical proof (p-values $< 0.01$) that the performance improvements are not due to chance but are inherent to the algorithm's mechanics.
• Handling Non-Uniqueness: The framework demonstrates the ability to navigate "equivalent" model spaces (specifically in the challenging "Bowl" model) more effectively than competing algorithms, achieving lower misfits and higher stability.

Results

• Synthetic Performance: Across all four synthetic models, iL-SHADE achieved median misfits in the range of $10^{-8}$ to $10^{-9}$. This represents an improvement of 2 to 4 orders of magnitude over standard DE and SHADE. In the most challenging "Bowl" model (characterized by a low-resistivity intermediate zone), iL-SHADE achieved a minimum misfit of $2.555 \times 10^{-9}$ and a median of $5.979 \times 10^{-9}$, significantly outperforming competitors.
• Statistical Significance: Pairwise Wilcoxon tests confirmed that iL-SHADE's superiority is statistically significant ($p \approx 3.020 \times 10^{-11}$) across all model profiles.
• Noise Robustness: Under Gaussian noise levels up to 20%, the algorithm maintained consistent convergence to noise-floor-bounded minima, with misfit values increasing linearly and predictably (from ~0.8066 to ~0.8094), indicating high stability.
• Field Data Application: Applied to Unzen Volcano data (1.4 km transmitter, 2.0 km offset), iL-SHADE successfully resolved a 5-layer subsurface structure extending to 5.6 km depth with an RMS misfit of 24.388. The model identified:
  • A surficial resistive layer (unaltered volcanic debris).
  • A moderate resistivity layer (weathered pyroclastics).
  • An ultra-conductive layer ($\rho \approx 1.48 \, \Omega\cdot\text{m}$) at 2.0–2.5 km depth, interpreted as a hydrothermal clay cap (smectite/illite).
  • A conductive geothermal reservoir ($\rho \approx 3.32 \, \Omega\cdot\text{m}$) extending to 5.6 km.
  • A deep resistive basement ($\rho \approx 977.7 \, \Omega\cdot\text{m}$) representing the magmatic heat source.

Significance and Claims
The paper claims that iL-SHADE provides a highly robust, parameter-adaptive optimization framework that effectively overcomes the severe non-uniqueness and parameter coupling inherent in TDEM inverse problems. By integrating historical memory and linear population reduction, the algorithm achieves a superior balance between exploration and exploitation, resulting in more reliable and accurate resistivity imaging in complex volcanic-hydrothermal systems. The study positions this approach as a practical, automated solution that reduces the reliance on manual hyperparameter tuning, thereby enhancing the reliability of geological interpretations derived from TDEM surveys. The authors note that while the current implementation is for 1D inversion, the framework's success suggests potential for future extension to 2D/3D problems, potentially aided by high-performance computing.