Heat and weight optimization methodology of thermal batteries by using deep learning method with multi-physics simulation

• Published in: Energy conversion and management 2021-05, Vol.236, p.114033, Article 114033
• Main Authors: Park, Tae-Ryong, Park, Hyunseong, Kim, Kiyoul, Im, Chae-Nam, Cho, Jang-Hyeon
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Batteries; Deep learning; Design; Design optimization; Flux density; Heat; Methodology; Military applications; Modelling; Multi-physics simulation; Neural networks; Optimization; Pain; Physics; Primary batteries; System effectiveness; Temperature; Thermal batteries; Thermal simulation; Weight
• ISSN: 0196-8904; 1879-2227
• DOI: 10.1016/j.enconman.2021.114033

•A novel methodology is presented for optimizing a thermal battery.•The design method of the thermal battery was modeled.•DNN model provides accurate and time-efficient temperature prediction results.•A case study is presented to demonstrate the methodology. Thermal batteries are primary batteries, used in military applications, and must be designed to meet mission-specific requirements. This mission-specific nature makes the design of thermal batteries a pain-stacking task as no specific guideline or concrete modeling technology is yet provided on optimizing heat, weight, energy density, and other important criteria when the requirements are given. In this study, we present an optimization methodology that can facilitate the weight-optimized design of a thermal battery by modeling the design processes and optimizing accordingly. Using this methodology, a 25 cell Li-Fe thermal battery was optimized with specific performance requirements as a case study. The performance of the thermal battery is highly dependent on heat and thus it is important to accurately predict the temperature during the optimization process. The multi-physics model can obtain accurate temperature prediction, but it takes a very long time to calculate and is difficult to apply to the optimization problem. Therefore, a deep neural network (DNN) model trained from a small number of multi-physics simulation data is adopted to overcome this issue. As a result, the DNN model showed an accurate prediction result with a maximum error of 1.5 °C, which is less than 1 % error. The DNN model significantly improved the time-efficiency by reducing the calculation time per solution of the multi-physics model, from 10 min to 0.01 s. The results indicate that our methodology can effectively guide the optimal design of complex thermal battery systems.