There is great exploration potential for ultra-deep reservoirs in Junggar Basin， and reservoir prediction with seismic data is the main technical means of oil and gas exploration at present. However， due to the low signal-to-noise ratio of seismic data of ultra-deep reservoirs， unclear reservoir-seismic data correspondence， and few actual drilling wells， it is difficult to establish an effective initial model for seismic inversion， which restricts the accuracy and reliability of ultra-deep reservoir prediction. Gravity inversion，as an important quantitative interpretation method， can obtain the characteristics of underground density distribution and provide support for geological interpretation. According to the density model， a relatively reliable low-frequency model can be established for seismic inversion， which can overcome the difficulty of applying seismic data in ultra-deep reservoirs to a certain extent.Meanwhile， the acquisition of gravity data is economical and convenient compared with that of seismic data， and it is easier to be applied in practice. Therefore， a new technique applying gravity inversion to seismic data-based reservoir prediction was developed in this article. Firstly， a quasi-neural network gravity inversion technique based on Gaussian radial basis functions was proposed to solve the gravity inversion problem and improve the resolution and reliability of gravity inversion. Then， the density body obtained by gravity inversion was used as training data， and a neural network was trained together with seismic and logging data to establish the initial model of seismic inversion. Finally， seismic inversion was carried out with the initial model constraints. This technique broke through the application bottleneck of single seismic data in ultra-deep reservoir prediction and overcame the limitations of logging constraints， providing a reliable initial model for seismic inversion. The application of this technique to the prediction of ultra-deep clastic rock reservoirs in Junggar Basin was consistent with the existing geological knowledge， indicating that this method had high practical value and application potential for ultra-deep reservoir prediction and could provide technical support for ultra-deep reservoir exploration.