Customized Federated Kernel Regression learning for predicting natural gas hydrate equilibrium with thermodynamic inhibitors: A comprehensive study

Gas hydrates, crystalline structures formed under specific conditions, have drawn substantial research interest in both flow assurance risks and promising applications. Accurately predicting their equilibrium conditions in the presence of thermodynamic inhibitors is crucial for safety and economic risks, especially in flow assurance and deep-water drilling operations. Traditional machine learning (ML) methods assume independent and identically distributed (IID) data collected from various experiments, which is often violated due to distinct conditions, chemicals, and protocols used in different gas hydrate formation experiments. This results in non-IID data with varying local distributions, making conventional ML models less effective and leading to performance degradation. This paper introduces an innovative Customized Federated Kernel Regression (CFKR) learning framework to address these challenges. CFKR customizes individual models on each experiment’s local data and establishes a novel collaboration strategy to capture similarities among various hydrate formation experiments’ local data. A comprehensive literature review collected data from 40 experiments on the impact of four categories of thermodynamic hydrate inhibitors (THIs): salts, alcohols, ionic liquids, and amino acids. The proposed CFKR framework demonstrated superior performance in predicting natural gas hydrate formation temperature and pressure compared to twelve ML algorithms across five evaluation metrics. For hydrate formation temperature prediction, CFKR achieved a coefficient of determination of 0.9849 and a root mean squared error of 0.2367. For hydrate formation pressure prediction, CFKR obtained 0.9359 and 0.9512 for the coefficient of determination and root mean squared error, respectively. •Predicting natural gas hydrate equilibrium with thermodynamic hydrate inhibitors.•Customized Federated Kernel Learning for predicting hydrate formation conditions.•Addressing non-IID data across diverse gas hydrate formation experiments.•Customizing individual nonlinear models for each experiment’s local data.