Researchers Use Physics-Informed AI to Accelerate Thermal Energy Storage Design

A team of engineering researchers in South Korea has developed a machine learning methodology aimed at accelerating the design of thermal energy storage units, an essential technology for lowering building emissions. Led by Joo Hyun Moon, an assistant professor of building systems engineering at Hanbat National University, the researchers designed a physics-informed neural network (PINN) that automates the evaluation of heat storage configurations.

The initiative focuses on latent heat thermal energy storage (LHTES) architecture, which relies on phase change materials to store and discharge thermal energy. As these wax-like substances melt and solidify, they release or absorb heat at nearly static temperatures, functioning like rechargeable thermal batteries. Past studies indicate that integrated LHTES hardware can lower a structure's heating and cooling power consumption by as much as 45 percent.

Despite their efficiency, LHTES systems remain challenging to bring to market because simulating the dynamic fluid mechanics and heat transfer inside them requires immense computing power. Traditional computational fluid dynamics (CFD) modeling delivers high structural accuracy across operational scales, but running these complex simulations for every iteration makes extensive design optimization impractically slow. Conversely, physical lab experiments are constrained to small, fixed hardware setups.

To solve this computational bottleneck, the research team built a hybrid framework that blends machine learning with fundamental physics principles, as first reported by TechXplore. They began by establishing a laboratory-scale LHTES baseline and a corresponding CFD model, using physical testing to validate the computer simulations. After verifying the physics model, the team executed just 15 high-fidelity CFD simulations to serve as a sparse training dataset for their machine learning architecture.

The resulting neural network incorporates physical conservation laws directly into its loss functions, allowing the AI to maintain high accuracy without overfitting despite the limited dataset. By combining a zero-dimensional physical model with a response surface model (RSM), the system acts as a high-speed digital twin capable of estimating fluid flow dynamics and heat transfer rates across unseen pipe shapes and operating conditions.

The digital twin was subsequently coupled with a Non-dominated Sorting Genetic Algorithm II (NSGA-II) to run multi-objective optimization routines. The integrated pipeline evaluates thousands of potential structural iterations in parallel, balancing three targets: maximizing total heat discharged, maximizing average output power, and minimizing the fluid pumping power required to run the unit.

In numerical trials, the AI platform accurately mirrored results from traditional CFD tools while completing structural searches autonomously. The optimization routine delivered a high-efficiency configuration that matched top-tier baseline thermal performance while substantially lowering pumping power requirements. Notably, the modeling run indicated that flatter pipe geometries consistently perform better than traditional round designs.

The study, published July 30, 2026, in the Journal of Energy Storage by Gwangwoo Han, Moon, and their collaborators, points to applications well beyond structural HVAC installations. The same physics-informed framework could be adapted to optimize cooling channels in electric vehicle battery packs, hyper-scale data center cooling racks, cold-chain logistics equipment, and solar thermal collectors. The team noted that intelligent control of latent heat systems can yield electricity cost savings exceeding 70 percent in operational environments.