Abstract
Slug flow in multiphase production systems is a critical challenge in the oil and gas industry, characterized by complex and oscillatory dynamics, e.g., limit cycles. First-principle (FP) models often employ physics simplification, such as a virtual valve for the slug formation. To capture complex physics poorly modeled by FP models, hybrid models combine data-driven techniques and physical knowledge, such as the architecture known as universal differential equation (UDE). This work aims to employ a hybrid model based on neural networks to enhance the modeling of multiphase oil production systems. In the UDE model, a neural network is embedded within the structure of the FP differential equations. To demonstrate the feasibility of the methodology, the model was trained based on synthetic data, employing parameters estimated from OLGA simulations. Since the system faces oscillatory behavior, we trained the UDE in two stages: the first one employs smooth collocation on data to obtain an estimation of the derivatives. The results show that the two-stage strategy was successful in addressing the resolution of an oscillatory system. In the second stage, the algorithm reached a scaled MSE of 0.009 in 502 iterations. For the testing phase, using 20%, 40%, 60%, and 80% valve openings, the scaled MSE reported was 0.088. For the 100% valve opening, outside of the training part, the MSE reported was 0.039, showing a good extrapolation capacity.