Abstract: Complex geological conditions, long drilling cycles, and difficult wellbore pressure control are common problems during oil and gas resource development in deep formations. Intelligent well killing methods, combined with multi-source real-time information feedback, can predict and update gas-liquid distribution and pressure change law in the wellbore in real time. However, the combination of different correction coefficients may derive the same pressure calculation result, which leads to the problem of multiple solutions of the model. By analyzing the evolution law of the spatial morphology of the solution at different historical time nodes, it was revealed that the essence of the multi-solution of the model came from the imperfection of the model training under the constraint of sparse data. The global model training optimization method based on real-time information sequence and the dynamic random population training optimization method were established correspondingly, and their search ability and applicable conditions for the global optimal solutions of the model were tested. The results show that the global training optimization method can achieve accurate control in the early stages of well killing, but the calculation time is long. The dynamic random population training optimization method is slightly different from the expected value in the early stage of well killing, but the calculation is rapid. According to available computing resources, a suitable training optimization method can be selected to achieve deep learning of the gas-liquid flow law in the wellbore under the constraints of multi-source real-time data.