Inverse modelling of electrical submersible pump subject to multiphase flow using physics-informed neural networks

• Grant number: 22/00934-2
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: December 01, 2022
• End date: February 28, 2023
• Field of knowledge: Engineering - Mechanical Engineering
• Principal Investigator: Alberto Luiz Serpa
• Grantee: Felipe de Castro Teixeira Carvalho
• Supervisor: George Em Karniadakis
• Host Institution: Faculdade de Engenharia Mecânica (FEM). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Institution abroad: Brown University, United States
• Associated to the scholarship: 19/14597-5 - Modelling the dynamical behavior in Electrical Submersible Pump (ESP) for liquid-liquid flow, BP.DR

Abstract

In many circumstances, the oil wells cannot provide enough energy to produce fluids to the surface at cost-effective rates. In these cases, artificial lift equipment enhances output rates by injecting more energy into the system. The ESP (electrical submersible pump) has been the second most extensively employed artificial lift technology. The pumped fluids often contain hydrocarbons in both liquid and gaseous phases, water, and sediments. Liquid-liquid two-phase flows due to the chemical characteristic can form colloidal dispersions such as emulsions. The emulsion effective viscosity can considerably exceed single-phase viscosities and is subjected to phase inversion. It can occur inside ESP due to the shear rate and temperature and cause instabilities. A mathematical model of a real-world system is simply an incomplete depiction of its behavior. It can never account for everything, and there will always be some source of error that will affect the accuracy of the model's prediction. Furthermore, even in relatively low complexity and parsimonious model, accuracy loss may occur due to idealizations of a given aspect of the system and inaccurate parameter estimation. The inverse problems are used when it is desired to estimate specific unknown properties of interest based on measurements that are only indirectly related to these attributes. The inverse problems are, in general, ill-posed, and solve them is regularly prohibitively expensive computationally, requiring complicated formulations, novel algorithms, and complex computer codes. Also, the standard techniques are still incapable of solving real-life physical problems involving missing, gappy, or noisy boundary conditions. In this context, the neural network, particularly the physics-informed neural networks, has recently gained attention on inverse problems due to its ability to handle complex and ill-posed problems. The PINNs are a subset of deep learning algorithms that combine data with abstract mathematical operators, such as PDEs with or without missing physics. Finally, obtaining an accurate dynamic model that considers the liquid-liquid two-phase flow within the ESP is challenging as it depends on several unknown system parameters and idealizations necessary to reduce system complexity. Therefore, the physics-informed neural networks are a promising technique that can handle ill-posed inverse problems, such as finding unknown parameters, such as dispersed and continuous phase viscosities and specific mass, fluid surface tension, emulsion droplet size, unknown valves constants, and unknown minor losses of ESP systems. (AU)