Many linear(ized) geophysical inverse problems cannot be solved without regularization. Finding the regularization parameter that best balances the model complexity and data misfit is often a key step in the inversion problem. Traditionally, this is done by first plotting the measure of model complexity versus data misfit for different values of regularization parameter, which manifests as an L-shaped curve, and then choosing the regularization parameter corresponding to the corner point on the L-curve. For this approach, the difference in units between model complexity and data misfit must be considered, otherwise the result will be strongly affected by the scaling between these two quantities. Inspired by the machine learning literature, we here propose an extension to the traditional L-curve method. We first split the raw dataset into training and validation sets, obtain a solution by performing inversion on the training set only, and calculate data misfits on the validation set. We demonstrate the efficacy of this approach with a toy example and with two synthetic datasets. In realistic global surface-wave tomography studies where sampling of the Earth is nonuniform, we devise a procedure to generate a validation dataset with sampling as uniform as possible. We then show that the regularization parameter can be determined using this validation set, and this determination is apparently robust to the ratio of data split between training and validation sets. For both synthetic tests and realistic inversions, we find that our procedure can produce a minimal point that can be easily identified on the misfit curves calculated on the validation sets, and avoids the nuances encountered in the traditional L-curve analysis.
17 July 2025 in Journal of Geophysical Research: Machine Learning and Computation
Precise determination of Moho depth is essential for understanding lithospheric deformation, lower crustal rheology, and crust‐mantle interactions. This study employs a Random Forest Regressor (RFR) model to predict Moho depth across the contiguous United States using a variety of geophysical data, including gravity anomalies, topography, heat flow, magnetic anomalies, sediment thickness, and electrical conductivity. The RFR outperforms the Airy isostatic model and other Machine Learning (ML) approaches, achieving a Coefficient of determination (R2) of ∼0.75, Root Mean Square Error of ∼3.3 km, and a standard deviation of ±10 km for ∼80% of predictions. Permutation importance identifies elevation and Bouguer anomaly as primary predictors, with their independent contributions confirmed despite correlation. Sensitivity analyses validate robustness, though the model is not transferable to oceanic crust. An innovative error estimation methodology integrates correlated and uncorrelated noise with variable standard deviations to capture geophysical data uncertainties and enhance model reliability. This study highlights the potential of accessible ML methods for first‐order Moho depth estimation in data‐scarce regions.
We determine the mantle attenuation (1/Q) structure beneath 70 Myr seafloor in the central Pacific. We use long-period (33-100 sec) Rayleigh waves recorded by the NoMelt array of broadband ocean-bottom seismometers. After the removal of tilt and compliance noise, we are able to measure Rayleigh wave phase and amplitude for 125 earthquakes. The compliance correction for ocean wave pressure on the seafloor is particularly important for improving signal-to-noise at periods longer than 55 sec. Attenuation and azimuthally anisotropic phase velocity in the study area are determined by approximating the wavefield as the interference of two plane waves. We find that the amplitude decay of Rayleigh waves across the NoMelt array can be adequately explained using a two-layer model: in the shallow layer, in the deeper layer, and a transition depth at 70 km, although the sharpness of the transition is not well resolved by the Rayleigh wave data. Notably, observed in the NoMelt lithosphere is significantly higher than values in this area from global attenuation models. When compared with lithospheric measured at higher frequency (~3 Hz), the frequency dependence of attenuation is very slight, revising previous interpretations. The effect of anelasticity on shear velocity (VS) is estimated from the ratio of observed velocity to the predicted anharmonic value. We use laboratory-based parameters to predict attenuation and velocity-dispersion spectra that result from the superposition of a weakly frequency dependent high-temperature background and an absorption peak. We test a large range of frequencies for the position of the absorption peak (fe) and determine, at each depth, which values of fe predict and VS that can fit the NoMelt and VS values simultaneously. We show that between depths of 60 and 80 km the seismic models require an increase in fe by at least 3-4 orders of magnitude. Under the assumption that the absorption peak is caused by elastically accommodated grain-boundary sliding, this increase in fe reflects a decrease in grain-boundary viscosity of 3-4 orders of magnitude. A likely explanation is an increase in the water content of the mantle, with the base of the dehydrated lid located at ~70-km depth.