تخمین عمق گنبدهای نمکی با استفاده از داده‌های گرانی از طریق شبکۀ عصبی رگرسیون تعمیم‌یافته، مطالعۀ موردی: میدان مورس، دانمارک

در این مقاله تخمین عمق گنبدهای نمکی با استفاده از روش شبکۀ عصبی رگرسیون تعمیم‌یافتهgrnn، از طریق داده‌های گرانی‌سنجی بررسی شده است. بدین منظور یک شبکۀ عصبی grnn به وسیلۀ داده‌های گرانی که از روش پیشرو، مدل گنبد نمکی را به دست می‌آورد، به ازای اعماق مختلف به‌دست‌آمده آموزش داده شد و با محاسبۀ خطای شبکه، شبکه مرتب اصلاح شد تا معماری شبکه با خطای پذیرفتنی به دست آید. سپس به‌منظور تست شبکه از داده‌های مصنوعی با 5 درصد و10 درصد نویز استفاده شد که دقت خوبی (خطای نسبی تخمین عمق در حضور 5 درصد نویز برابر با 3/8 درصد و در حضور 10 درصد نویز برابر با 5/43 درصد) را نشان می‌دهد. همچنین به‌منظور آزمون شبکه برای داده‌های واقعی، مشخصه‌های لازم از داده‌های گرانی مربوط به گنبد نمکی مورس در دانمارک، استخراج و به‌عنوان ورودی به شبکه اعمال شد و نتایج تخمین عمق تحلیل و بررسی گردید. نتایج نشان داد که تخمین عمق به‌دست‌آمده تا حدود زیادی به مقدار واقعی نزدیک و قابل‌قبول است.

Depth estimation of Salt Domes using gravity data through General Regression Neural Networks, case study: Mors Salt dome Denmark

In this paper an intelligent method through General Regression Neural Networks (GRNN) is presented to estimate the depth of salt domes from gravity data. Neural networks are as a good tool for automatic interpretation of geophysical data especially for depth estimation of gravity anomalies. The gravity signal is a nonlinear function of depth and density and the geometrical parameters of the buried body. One of the common modern tools for nonlinear systems identifications is neural networks. The parallel processing and the ability of the network to learn from training data is a good motivation to use them for interpretation of gravity data. Salt domes are as a target of the gravity explorations in oil exploration because in the most cases in Middle East, America and some parts of the Europe like Denmark they are as a good locations for oil traps and diapers. The nonsimple structure of the salt domes is noticeable. Almost in most of the available methods of salt dome modeling for depth estimation they are considered simply to simple geometrical bodies like sphere or cylinder.These simplifications cause to no adaption to the real nature of salt domes. The salt domes modeling in this paper is not followed these simplifications and the near to real shape of salt dome bodies is modeled through Grav2dc software. Different possibilities for the salt dome model are considered: salt dome with oil, salt dome with oil and salt water, salt dome with gas and oil, salt dome with none of the gas oil or salt water. For all the mentioned salt dome models both the Grav2dc software and surfer are used to calculate the gravity effect of the body and then the related feature are extracted. To train the general regression neural network the range of the salt dome depth(s) are selected regard to available geological prior information. For exle if the possible range of the salt dome is regard to the geological properties and/or well log data between 2 to 4 kilometers the GRNN is trained with models of salt dome with depths from 1 to 4 kilometer. In this way, first the gravity effect of several salt dome models with different depths were calculated via forward modeling and the GRNN was trained with this set of data. The GRNN architecture was modified regard to Root Mean Square Error of the GRNN network and modifications were followed and repeated until achieving the network with acceptable Root Mean Square Error (RMSE) for the training process. To test the GRNN the synthetic gravity data of salt dome with two different level of noise 5% as low noise, and 10% as high noise were applied to the designed GRNN and the related depth was estimated. Totally the results showed good ability of GRNN for depth estimation of salt domes. Finally, to test the GRNN for real data the GRNN was tested for gravity data over Mors Salt dome in Denmark. Mors salt dome is a gravity field for oil exploration and is also an interesting case study for a lot of the geophysics researchers and geoscientists. The results for real data also proved the ability of the general regression neural network for estimating the depth of salt domes with low root mean square error.