Multiphase flow models are crucial in various engineering applications, often involving the coupling of a transport solver with a local physical model that calculates thermodynamic equilibrium equations within each grid cell to determine the fluid's local thermodynamic properties. To accelerate the frequently significant computational cost of these physical models, surrogate models have become a standard approach. However, challenges persist, particularly when physical models depend on numerous parameters, complicating the training process and diminishing the accuracy of surrogate models across different simulations. Vari-ous strategies are typically employed to address these issues, including active learning and online learning . The ISAT methodology, originally applied in the combustion field has been successfully applied in various fields with problems where the parameter space is large and highly nonlinear. In our paper, we introduce a novel, flexible, and adaptive strategy for training, validating, and inferring surrogate models. Our approach is based on the ISAT methodology associated with dynamic learning strategies. The dynamic design of the training dataset allows for the continuous updating and enhancement of the model during simulations by integrating new data. This not only improves the accuracy of the models but also adapts to changing condi-tions in real-time. Moreover, we have developed a flexible client-server architecture to facili-tate efficient on-demand training and inference. This system includes online prediction error estimation, ensuring the surrogate model's quality and accuracy throughout its operation. Our methodology has been tested using synthetic test cases and further validated through its application in a realistic reservoir simulation involving a single injection well. The results demonstrate the effectiveness of our active and adaptive learning method in significantly enhancing the performance of multiphase flow models, thereby offering a robust solution to the challenges posed by high computational costs and the need for high accuracy in dynamic simulation environments.