Abstarct: Different ways are used in geophysics to detect underground structures and quantitative and qualitative interpretation of anomalies in consequence of such structures. Detection of surface, depth expansion and the depth of embedment of bodies, masses and geophysics phenomena are of geophysics aims. Among geophysics methods, the gravity one is applied to detect geology anomalies in varied scales. To interpret the qualitative and quantitative anomalies, identification of their depth and shape with modeling is used in different ways. To do this, simple models are used such as sphere and cylinder and compound model. One of the used ways about qualitative and quantitative interpretation is artificial neural network. In this study the aim of using of artificial neural network is the estimation of underground structure shape, the estimation of depth, radius and the density of anomalies made by them. So by providing artificial models and accounting their gravity effect, the artificial networks have been designed. The MLP has been the used as a neural network in this study that has been trained by some artificial models, a part for training and some for validating and assessing. In order to estimate the shape, depth and the density of anomalies, four distinct neural networks have been provided and made. At first anomalies shape has been identified and baxsed on that the estimation of depth and density has been gotten possible. To do this to coordinate the modeling results along with earth fact, sphere simple models and cylinder model and vertical cylinder-sphere compound artificial model (salt dome) with different dimensions and parameters have been used. Considering the anomaly shape of gravity is dependent on dimensions and depth and shape of underground structures, so these dependent features of the diagram of the gravity effect of these phenomena has been used as artificial network input. To train the network, firstly some of results of different models were used to train the network and the other models for evaluating and validating of the network. A network with different laxyers and neuron numbers has been considered to produce an optimum neural network and the performance diagram of every network in different states has been provided and calculated to gain the optimum model. Thus by making some networks, the estimation of shape, depth and radius and density of underground anomalies have been dealt with. Finally the optimized network has been used to interpret the real gravity data, of salt dome of Hormoz and the salt dome of Humble. The study results showed that the shape of Humble salt dome up to higher extent is sphere and its depth is 2.5. This method also was used to estimate the shape and the depth of anomaly of Hormoz salt in Hormozgan province and the result of this study shows that the shape of this salt structure is compound meaning the chimney state along with spheroid skullcap.