Lab study uses physics-informed AI to speed thermal storage design

A wax-based thermal storage system sits at the center of a laboratory-scale experiment in South Korea. When the material melts, it absorbs heat; when it hardens, it releases it. Researchers led by Joo Hyun Moon at Hanbat National University have now combined that system's physical equations with an AI model to search for better designs, using 15 high-fidelity simulations as training data.

Latent heat thermal energy storage can hold a large amount of energy and release it at a nearly constant temperature, a useful property for managing heating and cooling. But its internal heat transfer and fluid flow are difficult to simulate together. Computational fluid dynamics can describe those processes, yet the calculations are slow and costly when engineers want to compare many geometries.

The team first built a laboratory-scale storage setup and a corresponding CFD model. Measurements were used to validate the simulations, which showed excellent agreement with only minor deviations. The researchers then trained a physics-informed neural network, or PINN, a model taught to respect physical laws rather than rely only on large datasets. It learned case-specific heat-transfer behavior, while a response-surface model estimated how geometry and flow conditions would affect unseen designs.

The resulting digital twin was connected to NSGA-II, a genetic algorithm for balancing several goals at once. It searched for systems that maximize discharged heat and average discharge power while minimizing the pumping power needed to move fluid. In numerical experiments, the optimized design matched the performance of the strongest baseline while significantly reducing pumping power. Flatter pipes emerged as the more favorable geometry.

So what changes in practice? Engineers could move from manually testing a limited set of designs to rapidly exploring a much wider design space before building hardware. The immediate result is still a laboratory-scale and numerical study, not a commercial system or field deployment. But the same method could be applied to thermal management in buildings, electric-vehicle batteries, data centers, cold-chain logistics and solar thermal systems, where faster design iteration could help reduce energy use and emissions.