The application of learning biases in machine learning, as we denominate the constraints that guide training to physically meaningful structures from data, improves generalization and accuracy, also in data-limited circumstances. Digital twins drove by these systems bridge reality and machines for prediction and decision-making. However, some phenomena are hard to be characterized with ordinary sensors or machine vision. That is the case with fluid dynamics, where we usually do not have access to the internal state of fluids. We present a digital twin that forecasts the behavior of real liquids from the evaluation of the free surface with computer vision techniques. We leverage deep learning structures, and specifically recurrent neural networks, to exploit temporal information of fluid motion to recover the internal state of the liquid to perform thermodynamically informed forecasting with a neural network based on the GENERIC formalism. Nevertheless, machine reasoning about fluid dynamics may encounter additional difficulties. Due to the impossibility to learn a global fluid model or small specific approximations for each type of liquid, we must rely on active learning for training on one single model and adjusting to new dynamics perceived. The proposed method adapts itself to previously unseen liquids with the information on the motion of the free surface captured with computer vision. This strategy has been evaluated in computational and real scenarios, exhibiting the capacity of GENERIC to learn sufficiently general features of the dynamics of one liquid to evolve to the prediction of new behaviors.