Core Concepts
The author proposes learning on the correct class as a strategy to address the ill-posed inverse gravimetry problem, ensuring reliable solutions within specific constraints.
Abstract
In the study of inverse gravimetry, end-to-end learning approaches are explored to determine Earth's mass distribution. The reliability of these methods is questioned due to ill-posedness, leading to the proposal of learning on the correct class. By employing well-posedness theorems and unique solution constraints, a neural network architecture is designed for domain inverse problems of gravimetry. The approach involves mimicking level-set formulations and utilizing density-contrast functions to recover non-constant mass models efficiently.
The content delves into the challenges posed by ill-posed inverse gravimetry problems and introduces a novel strategy of learning on the correct class. By restricting solutions within specific classes based on uniqueness theorems, reliable outcomes can be achieved. The proposed method involves designing neural networks that adhere to certain geometric constraints and utilize a priori information for accurate mass model recovery. Through numerical examples and theoretical frameworks, the efficacy and promise of this approach in geophysical explorations using gravity data are highlighted.
Stats
Given some a priori information of f and imposing certain geometric constraints on D, the domain inverse problem of gravimetry admits a unique solution.
We build pairs of mass models and gravity data for training neural networks.
The loss function for training includes terms measuring discrepancy between true model µs and neural-network output.
Table 1 lists average values of PSNR and SSIM for reconstructions on test samples and salt models.