This technology pre-trains AI models, such as the DeepONet architecture, using virtual datasets generated by a multi-physics computational model consisting of state estimation, thermodynamics, chemical reaction, and pressure estimation models. The trained model receives real-time battery temperature and cooling performance data to predict internal states, such as SEI concentration and electrode concentration, as well as the likelihood of thermal runaway.
Predicting thermal runaway in secondary batteries has historically been difficult in real-time due to the complex chemical and physical calculations required. Furthermore, training models with actual experimental data is time-consuming, costly, and limited by the availability of data.
This technology generates virtual datasets by inputting thermal curves derived from virtual scenarios into a multi-physics computational model, which are then used to pre-train AI algorithms. During operation, it quickly infers thermal runaway status and probability by inputting measured battery surface temperature, coolant temperature, flow rate, and convection coefficients. Applicable to EV battery packs, large-scale ESS containers, and battery safety monitoring platforms, it enables early thermal runaway warnings and proactive cooling system activation without the need for expensive thermal abuse testing.
This invention was developed with support from the Ministry of Science and ICT for the development of impedance management and lifespan enhancement technology for electric vehicle lithium-ion batteries using low- and high-frequency synthetic surface pressure excitation.
US2026-0141140A1
