Parameter estimation for thermal runaway of Li-ion cells: a Gaussian Process approach

Lithium ion (Li-ion) cells are the most prominent electrochemical energy storage device in todays world as they are utilised in many applications across many scales. However, Li-ion cells can suffer from a severe safety issue known as thermal runaway (TR). This process is due to exothermic chemical decomposition of a cells components. Being able to understand and model this process is essential for the development of safer batteries. Lithium iron phosphate (LFP) cells are known to have the safest chemistry of Li-ion cells. However, TR models developed for LFP cells have not been validated and when compared to new experimental work are shown to be inaccurate. Hence, the development of an accurate and validated LFP TR model is the focus of present research. Classical TR modelling of Li-ion cells utilises four Arrhenius equations to predict the reaction rate of these reactions and in turn the heat generated. However, the development of a TR model (via parameter estimation) is difficult due to 1) the large range some of the parameters of the Arrhenius equation can take i.e. several 10s order of magnitude between literature values for a given reaction, 2) the complex interaction between different parameters within an Arrhenius equation, and 3) the effect of the heat generated by each reaction on the others. Given this, direct minimisation of the root mean squared error (RMSE) between simulated and experimental results has proven to be difficult, with optimized parameters unable to represent important experimental features. As such, an approach is developed in which the viable parameter space is defined and sampled using a pseudo-random sampling. The error of the heuristic fit (RMSE) is then emulated by a Gaussian Process, further refined by Global Sensitivity Analysis. The result is an emulator providing predictions (with attached uncertainties a part of the process) as a function of a reduced number of combined parameters. This is a reduction in model order achieved by a novel, optimal rotation of input basis.