Data‐Driven Design Optimization of Solar Salt Latent Heat Storage Systems Using Computational Fluid Dynamics and Machine Learning

Design optimization of latent heat storage (LHS) systems is essential for improving their charging performance and facilitating efficient integration with renewable energy systems. However, repeated computational fluid dynamics (CFD) simulations for large‐scale geometric optimization are computationally intensive and time‐consuming due to the nonlinear melting behavior governed by coupled conduction and buoyancy‐induced natural convection. To address this limitation, the present study deploys a coupled CFD and machine learning (ML) framework to optimize tube eccentricity ( L ₁, L 2 ) and angular orientation ( θ ) in a shell and multi‐tube LHS system. ML models are utilized with CFD‐generated datasets for rapid prediction of melting time and efficient design optimization. The results reveal that melting dynamics are strongly governed by the coupled effects of eccentricity and tube angular orientation, which alter the strength of buoyancy‐driven convection. Angular tube orientation has a strong effect, as changing θ from 35° to 60° (with L ₁ and L 2 nearly constant) increases melting time from 6000 s to 15 000 s. Moreover, at a fixed angular orientation ( θ = 50°), reducing the tube eccentricities (from case B to case C) reduces the melting time by 14.9%. The optimal configuration of θ = 35°, L ₁ = 34.5 mm, and L 2 = 30 mm (Case K) reduces the total melting time to 6000 s, achieving a 65.7% reduction compared to the least favorable Case E, L ₁ = 40 mm, L 2 = 20 mm, and θ = 70°. Four supervised ML models, including Polynomial Regression (PR), Support Vector Regression (SVR), Random Forest Regression (RFR), and XGBoost, are trained and tested using CFD‐generated data to predict melting time. The predictive capability of the models is estimated using the mean squared error (MSE) and coefficient of determination ( R 2 ) for both the training and test datasets. Among them, SVR showed the best predictive performance ( R 2 = 0.97, MSE = 59.86) while XGBoost showed comparatively lower generalization capability ( R 2 = 0.93, MSE = 201.83). The proposed CFD–ML framework provides a computationally efficient and scalable approach for rapidly optimizing LHS systems.