Abstarct: Abstract Increasing the safety and performance of structures during an earthquake is one of the most important and challenging issues in civil engineering. Meanwhile, soil is a factor that can significantly affect the performance of the structures. Generally, the dynamic analysis of structures is performed assuming that the foundation is rigid, and the effect of soil interaction under the foundation is ignored. But in order to achieve the real answer and correctly predict the behavior of the structure, soil-structure interaction studies have become more important especially since the 1970s and received the attention of many researches. Therefore, nowadays, many studies are conducted on the effect of soil on the performance of the structure, and in fact, the studies of the interaction between the soil and the structure show the difference between the real response of the structure and the response of the structure in the fixed baxse state. The common methods for solving soil-structure interaction problems are time-consuming and expensive methods, and therefore limit the consideration of soil-structure interaction for structures. In recent years, with the increasing progress of technology and the increase of information and data, it forces us to enter the field of artificial intelligence and use it. In the meantime, one of the methods of artificial intelligence known as neural network can be a solution to the complex problems of soil and structure interaction in a short period of time and very accurately. In this research, the seismic behavior of steel bending frxame structures under the effect of soil and structure interaction with different heights was investigated by conventional and classical methods. For this purpose, three steel structures 4, 8 and 12 story were modeled with the help of OpenSees software with fixed baxse states and interaction with soft soil in accordance with the 2800 standard. Each of the modeled structures was subjected to incremental dynamic analysis (IDA) by six records of the near field and six records of the far field, which had the same components but different stations, and their seismic behavior included maximum floor drift and maximum baxse shear by The IDA curve was drawn. Then, the neural network model was built using a multilxayer perceptron artificial neural network with one input laxyer, three hidden laxyers, and one output laxyer, and a comparison was made between its results and the results of modeling in OpenSees software. The results of this research showed that due to the high correlation coefficient values of 0.9 , very low MSE and MAE function errors, as well as the proper matching of forecast and real data, the neural network method has been successful and has good accuracy for solving complex soil-structure interaction problems and we can hope to replace them with usual and classical methods.