Abstract: The flow pattern of gas-liquid two-phase flow is an important parameter in the production process of chemical industry, petroleum, thermal power, nuclear power generation and other industries. The accurate prediction of flow pattern is of great significance for production and application. Nowadays, the flow pattern recognition mainly depends on the images obtained by visualization technology and the hydraulic characteristics obtained by conductivity probe and resistance void meter. The two-phase flow patterns obtained by these experimental methods are empirical flow patterns. The liquid-phase apparent velocity and gas-phase apparent velocity are generally in the low velocity region (0.4 m/s), which is unable to meet the practical production application. In order to predict the flow regime under high velocity flow conditions and develop the corresponding flow regime map, a machine learning model of random forest (RF) based on artificial fish swarm algorithm (AFSA) optimization is proposed in the present study. Based on the optimal and simplified parameters, the model can take the four basic characteristics of gas-phase apparent velocity, liquid-phase apparent velocity, Reynolds number (RE) and Schmidt number (SC) and the processing characteristics obtained through simple calculation and processing as the feature input, take the flow pattern label as the network output, and carry out the intelligent identification of flow pattern after training. This model prevents over fitting and under fitting behavior by setting parameter target criteria, and ensures the rationality and accuracy of prediction. The model also obtains a reasonable combination of parameters through artificial fish swarm acceleration, and obtains higher training accuracy and test accuracy than other optimization models. In this study, the two-phase flow pattern data in the vertical-downward under different conditions were obtained through experiments. Then the AFSA-RF model, ordinary RF model and other mainstream classifiers were used to identify these data, and then the identified data were generated into the corresponding flow pattern diagram and its transformation boundary. After repeated tests, the proposed model was then successfully applied to the identification of flow regimes in the vertical-downward two-phase flow. With the calculation of experimental data by different classification models and optimization methods, it is found that the accuracy of the flow regime identification under high velocity flow conditions reaches 90.91%. The accuracy and stability of flow regime identification of the proposed model are better than that of the RF model and other similar models, which further verifies the effectiveness of the prediction model. With the proposed model, the application range of the flow regime map is extended to the high-velocity flow conditions, which can meet the needs requirement of practical engineering application.