17:15 〜 19:15
[SEM15-P16] A Study on Physics-Guided Machine Learning for Near-Surface Quality Factor Q Estimation
キーワード：Q-value, Physics-Guided, Machine Learning, subsurface
Quality factor (Q) is a key physical parameter for quantifying the inelastic properties of subsurface media, aiding in the characterization of attenuation intensity within geological strata. Accurate estimation of Q is essential for improving imaging quality and ensuring reliable inversion of physical parameters. Previous research has shown that viscoelastic attenuation significantly affects dispersion-based spectral imaging and the matching of theoretical dispersion curves, especially in near-surface layers characterized by stronger absorption and lower Q-values.
Conventional Q-value inversion methods often adopt the constant-Q assumption, implying that Q is independent of frequency. Such simplification can introduce inaccuracies due to an incomplete depiction of underlying physical phenomena. This issue becomes particularly pronounced in highly absorptive near-surface regions, where inversion results exhibit large uncertainties and can vary significantly depending on the selected data and frequency bands. Although state-of-the-art machine learning models can sometimes outperform purely physics-based models—especially when abundant training data are available—they may yield solutions that are physically inconsistent.
In this paper, we propose a physics-guided neural network model that combines artificial intelligence with physics-based modeling, capitalizing on their complementary strengths to enhance the modeling of physical processes. This framework opens new avenues for high-precision Q-value inversion in near-surface applications. A data-driven model guided by physical principles requires a substantial number of paired training samples. Given fixed layer thicknesses, we randomly generate numerous subsurface models within a constrained parameter range. We incorporate an empirical relationship between the Q-value and frequency f, derived from laboratory rock-physics experiments: Q-1f=Q0-1(f/f0)γ, thus producing a large set of synthetic sample pairs. We then employ a Transformer model for inverting the formation’s absorption factor Q. The input features include frequency, velocity, time, amplitude, and the Q-value estimated via the spectral-ratio method. These features are standardized and fed into the Transformer, whose multi-head self-attention mechanism and feedforward layers capture the complex relationships among inputs. The network architecture comprises multiple encoder layers, and positional encoding is added to preserve sequential information.
Conventional Q-value inversion methods often adopt the constant-Q assumption, implying that Q is independent of frequency. Such simplification can introduce inaccuracies due to an incomplete depiction of underlying physical phenomena. This issue becomes particularly pronounced in highly absorptive near-surface regions, where inversion results exhibit large uncertainties and can vary significantly depending on the selected data and frequency bands. Although state-of-the-art machine learning models can sometimes outperform purely physics-based models—especially when abundant training data are available—they may yield solutions that are physically inconsistent.
In this paper, we propose a physics-guided neural network model that combines artificial intelligence with physics-based modeling, capitalizing on their complementary strengths to enhance the modeling of physical processes. This framework opens new avenues for high-precision Q-value inversion in near-surface applications. A data-driven model guided by physical principles requires a substantial number of paired training samples. Given fixed layer thicknesses, we randomly generate numerous subsurface models within a constrained parameter range. We incorporate an empirical relationship between the Q-value and frequency f, derived from laboratory rock-physics experiments: Q-1f=Q0-1(f/f0)γ, thus producing a large set of synthetic sample pairs. We then employ a Transformer model for inverting the formation’s absorption factor Q. The input features include frequency, velocity, time, amplitude, and the Q-value estimated via the spectral-ratio method. These features are standardized and fed into the Transformer, whose multi-head self-attention mechanism and feedforward layers capture the complex relationships among inputs. The network architecture comprises multiple encoder layers, and positional encoding is added to preserve sequential information.