Physics-informed neural networks for multi-objective design optimization of latent heat thermal energy storage systems

Designing latent heat thermal energy storage (LHTES) systems is computationally expensive due to the reliance on slow computational fluid dynamics (CFD) simulations. This study overcomes this bottleneck by developing a hybrid physics-informed neural network (PINN) framework. This PINN, governed by a 0D lumped-capacitance physical model, was trained on a sparse dataset of only 15 validated conjugated heat transfer (CFD) simulations. The resulting digital twin demonstrated exceptional fidelity, achieving a coefficient of determination ( R 2 ) greater than 0.999 against the ground truth data. This validated, instantaneous surrogate model was then coupled with a non-dominated sorting genetic algorithm II (NSGA-II) to perform a comprehensive multi-objective design optimization (MODO). The optimization autonomously navigated the fundamental thermo-hydraulic trade-off by simultaneously maximizing total discharged heat ( Q tot ) and average power ( P avg ) while minimizing pumping power ( W pump ). The balanced optimal designs on the global Pareto front matched the thermal performance of the best baseline (Flat-22), while reducing pumping power. This study demonstrates a powerful PINN-driven framework that transforms the LHTES design process from slow, manual evaluation to a rapid, autonomous exploration of the entire continuous design space, enabling the discovery of holistically optimized solutions. • Hybrid PINN framework accurately models LHTES using experiment and CFD results • Physics-regularized learning ensures consistency in sparse data regimes • 0D lumped model enables instantaneous prediction of transient dynamics • Multi-objective optimization reveals global thermo-hydraulic trade-offs • Optimized configurations balance thermal power and pumping energy costs