Journal de géologie et géophysique
Libre accès

Abstrait

Basement Relief of Sedimentary Basin Using Gravity Data from Deep Neural Network Approach

Chandra Prakash Dubey*, Madhusree Majhi, Laxmi Pandey

Subsidence of the Earth’s crust has allowed sediments to accumulate on the top of a basement of igneous and metamorphic rocks in form of sedimentary basin. These sediments and associated fluids are chemically and mechanically transformed through the several physical events like compaction and heating to a course of time. Consequently, it becomes the reservoir of the energy resources of petroleum, natural gas, coal, geothermal energy, and uranium etc. Their generation, development and disappearance are directly related to plate tectonic movements and other important geological events to understand the evolution history. Therefore, it is very crucial to evaluate the thickness of the sediments in terms of basement relief to highlight the depositional settings and basin formation factors. Here, we have developed a MATLAB based Artificial Neural Network approach to obtain the depth of a sedimentary basin considering the density variation with depth. In this work, a synthetic model is created initially by using 2D rectangular prism and later the model is perturbed with a 5% Gaussian White Noise. A supervised learning process is used to train the neural network and backpropagation with stochastic gradient descent technique is used to optimize the network output. The prism model is then used to create synthetic sedimentary basin to determine the depth profile with known density contrast using computed gravity datasets. After checking this optimization for various synthetic model, the technique is used on real data taken from Sayula Basin, Mexico and the results are compared with previous basement depths to validate its efficacy. The novelty of the proposed neural network approach is fast and efficient computation without any initial model assumptions that can map complex input output relations very efficiently, where other optimization process lack in this segment.