تخمین عمق سنگ‌بستر از داده مغناطیس هوابرد در حوزه رسوبی ریگ جن با استفاده از شبکه‌های عصبی عمیق

Determining the geometry and depth of sedimentary basins is a fundamental objective in geophysical exploration, particularly in regions with hydrocarbon potential. Magnetic methods are commonly applied for this purpose when sufficient magnetic contrast exists between the basement rocks and the overlying sedimentary sequences. The inversion of magnetic data, however, is inherently nonlinear and ill-posed, and conventional inversion approaches generally rely on iterative optimization schemes combined with regularization strategies to obtain stable solutions. While these methods have been successfully applied in numerous studies, their performance may depend on factors such as the choice of initial models, regularization parameters, and computational cost, especially in regional-scale applications. In recent years, data-driven techniques based on machine learning have been increasingly explored as complementary tools for addressing inverse problems in geophysics. Within this context, the present study examines the applicability of deep neural networks (DNNs) for the inversion of airborne magnetic data, with the aim of estimating the depth of the magnetic basement in the Rig-e Jen sedimentary basin, Central Iran.

Due to the limited availability of dense, labeled field data suitable for supervised learning, a large and diverse synthetic dataset was generated to train the neural network. The subsurface was represented by a simplified two-layer model consisting of a non-magnetic sedimentary cover overlying a magnetic basement. The basement surface geometry was constructed using ensembles of rectangular prisms with laterally varying depths designed to reflect geologically plausible scenarios. The lower boundary of the model was fixed at a depth of 25 km, while the average basement depth was randomly varied between 3 and 15 km. Magnetic responses were calculated along profiles 80 km in length with a sampling interval of 1 km, using the analytical formulation of Rao and Babu (1991). A constant magnetic susceptibility value of 0.034 SI was assumed for the basement, along with fixed geomagnetic field parameters. In total, 100,000 paired synthetic models and corresponding magnetic anomalies were generated efficiently to form the training dataset.

A fully connected multilayer perceptron (MLP) architecture was employed to approximate the nonlinear mapping between magnetic anomalies and basement depth. The network consisted of three hidden layers with 215 neurons each and utilized the GELU activation function. He normal initialization was applied to the network weights, and dropout layers were included to mitigate overfitting. Training was carried out using the AdamW optimization algorithm with a mean squared error loss function. The dataset was divided into training, validation, and independent test subsets to enable an objective evaluation of model performance. Tests conducted on synthetic data indicate that the trained network is capable of reproducing the general characteristics of basement topography. Additional experiments incorporating Gaussian noise with an amplitude of 10% demonstrate that the model maintains a reasonable level of stability under noisy conditions.

Following validation on synthetic datasets, the trained DNN was applied to airborne magnetic data from the Rig-e Jen region. A network of eight intersecting profiles, oriented along NE–SW and NW–SE directions, was designed to facilitate consistency checks at profile intersections. The estimated basement depths range approximately from 5 to 14 km, with greater depths predominantly observed in the northwestern part of the basin. Differences in predicted depths at profile intersection points average about 400 m, which is small relative to the overall basin thickness and suggests an acceptable level of internal consistency. The reconstructed magnetic anomalies show a satisfactory correspondence with the observed data, with maximum residuals on the order of 12 nT.

To evaluate uncertainty associated with physical assumptions, a sensitivity analysis was performed by varying the basement magnetic susceptibility within a realistic range (0.028–0.040 SI) while keeping the basement geometry fixed. The resulting variations in predicted basement depth were quantified using the standard deviation, providing an estimate of uncertainty linked to susceptibility heterogeneity. The results indicate that susceptibility variations influence depth estimates, although the predicted basement geometry remains relatively stable over most parts of the study area.

In summary, this study demonstrates that deep neural networks can serve as a complementary, data-driven approach for the inversion of airborne magnetic data in sedimentary basins characterized by thick sedimentary cover. The proposed methodology offers a computationally efficient framework for estimating basement depth once the model is trained, while producing results that are consistent with regional geological constraints. The approach is particularly suitable for regional-scale investigations in data-limited settings and may be further refined through the incorporation of additional physical parameters or integration with independent geophysical datasets.