Geophysical inversion involves determining the subsurface properties of the earth by analyzing geophysical data. Conventionally, geophysical inversions have encountered several challenges including non-unique solutions, nonlinearity, low-resolution and noisy data analyses, mandatory constraints and simplifications, computational costs, and subjective interpretations. Addressing these challenges necessitates the development of advanced inversion algorithms to enable a more comprehensive and robust analysis of subsurface properties. Implementing deep neural networks, this thesis conducts nonlinear inversions of gravity and magnetic data for subsurface modeling by learning complex patterns and relationships in large training datasets. Nevertheless, a key challenge lies in the scarcity of large-scale training datasets required for the intelligent inversion problem. To address this issue, a novel technique has been developed to simulate geopotential datasets that represent the characteristics of real-world subsurface properties and their corresponding geopotential data. The technique's adaptability to diverse subsurface complexities allows for more comprehensive and accurate nonlinear inversion of geopotential data. The dataset simulation technique adopts forward modeling to visualize the subsurface into crustal layers and incorporates physic-based constraints into the process. To ensure comprehensive coverage of geological complexities in the forward models, the technique incorporates multiple structural parameters. This results in randomized changes in the topography and depth of the subsurface layers. The forward model simulations are followed by calculating their synthetic gravity and magnetic anomalies. The final training dataset is created by putting together the calculated gravity and magnetic anomalies of the forward models as input features and the topography of the subsurface layers as labels. The application of the proposed technique is practiced on airborne gravity and aeromagnetic anomalies offshore Abu Dhabi, United Arab Emirates. Using simulated datasets, several deep neural network models are trained to implement inversion of gravity anomalies, inversion of magnetic anomalies, and joint inversion of gravity and magnetic anomalies. The performance of the models is evaluated on actual and noise-added synthetic gravity and magnetic anomalies. By leveraging the trained models, the salts and basement structures are investigated, providing valuable insights into the geological structures of hydrocarbon reservoirs in this region.