Estimating the State of Health of Lithium-Ion Batteries: Comparing the Computational Cost and Precision of Kolmogorov-Arnold Networks and Multi-Layer Perceptron Artificial Neural Networks

• Published in: Meeting abstracts (Electrochemical Society) Vol. MA2025-01; no. 5; p. 609
• Main Authors: Bueno, Reginaldo, Bianchi, Reinaldo, Giacomini, Renato
• Format: Journal Article
• Language: English
• Published: The Electrochemical Society, Inc 11.07.2025
• ISSN: 2151-2043; 2151-2035
• DOI: 10.1149/MA2025-015609mtgabs

Lithium-ion batteries are currently the primary energy storage for various devices, from smartphones to electric vehicles. Although this technology has numerous advantages over its predecessors, lithium batteries still have some points of concern. In an electric vehicle, the battery price accounts for a large portion of the total vehicle cost, and its production requires the metallic element lithium, whose deposits are concentrated in a few countries worldwide. These points lead to the need to accurately estimate the battery's State of Health (SoH) to utilize its full potential before discarding or recycling it, aiming to reduce costs, minimize the environmental impacts of its disposal, and reduce the production of new batteries. Besides the economic and ecological effects, there is also a safety concern. When the battery's state of health falls below a specific range (usually when its capacity drops below 70% of the initial off-the-factory capacity), it becomes more prone to failure, posing a risk of overheating and explosions. Another aggravating factor is that the SoH of lithium batteries is challenging to estimate quickly and without relying on laboratory tests due to the complexity of the degradation phenomena in the batteries and the need for continual monitoring. Despite being introduced commercially in the early 1990s, the number of papers concerning new methods for estimating their State of Health (SoH) is growing yearly, revealing that this subject is still a fertile field for research. Academia has witnessed a race of papers aiming to reach the lowest prediction error using various methods, from electrochemical modeling to deep neural networks. Nonetheless, few papers inform or discuss the computational cost their methods require for prediction, leaving a gap concerning the applicability of such methods in low-end and real-life hardware. The paper presenting the concept of the Kolmogorov-Arnold Networks (KAN) (Liu et al. 2024) introduced an innovative approach compared to the well-established Multi-Layer Perceptron Neural Artificial Networks (MLP). While MLPs have fixed nonlinear activation functions on nodes and learnable linear weights on synapses, KANs have no linear weights but learnable 1D functions on their nodes. The authors of that paper claim that the KANs are more interpretable and precise than MLPs, citing the lack of interpretability of MLPs as one of their main drawbacks. When it is possible to analyze a Machine Learning (ML) model and understand why it provided a specific response given a particular input and how its response changes when the inputs change, the ML model is considered interpretable; such understanding is not possible with MLPs, but as KANs provide a mathematical formula that describes the output based on the inputs, it is possible to understand the contribution of each input for the output. Here, the interest is not in the interpretability but in the formula the KAN provides to solve the problem. Calculating the result of a formula may be way faster than running a deep neural network regression algorithm. The purpose of this study is to compare the training time, regression time and precision of an MLP algorithm based on Keras + TensorFlow with a KAN model, based on the well-known NASA Ames Prognostics Center of Excellence (PCoE) battery dataset in both cases, on the same hardware (PC with i5 processor, GPU not used). Two features based on Incremental Capacity Analysis (ICA) and one dV/dt feature are extracted from the battery charging curve during the constant current phase. A Gaussian Filter (GF) is used to filter out the high-frequency noises from the ICA curves. The MLP and KAN models are then trained on the features obtained. The main results are: while the Root Mean Square Error (RMSE) is 14.2% lower with the MLP approach (RMSE = 0.928 and 1.082 for MLP and KAN, respectively), the regression time with the KAN approach is four orders of magnitude lower (regression time = 62.5 ms and 3.095 µs for MLP and KAN, respectively). The mean training time of the KAN is also shorter: 17.7 s against 67.7 s of the MLP. The study concludes that if the computational cost is a concern for a specific application/hardware and a higher root mean square error is tolerable, using a KAN model is not just a solution to consider but a practical and viable option for real-life hardware, given its significantly lower training and regression time. Figure 1