Enhancing efficiency and accuracy of structural feature recognition in battery cathode materials by collaborating Kolmogorov-Arnold Networks（KAN）with self-attention mechanism

In the past few decades, the application of machine learning algorithms, especially deep learning, has experienced explosive growth. There have been mature and large-scale applications of ML in many engineering fields, such as autonomous driving, facial recognition, etc. Besides these achievements, AI attracts very much attention in modern scientific research of Chemistry and materials. Due to the complexity of battery materials, how to reasonably select features and use appropriate machine learning methods to predict their chemical properties (e.g., volumetric, discharge average voltage, etc.) has become a topic of ongoing discussion among scholars. In this work, a new way which directly approaches the problem from a crystallographic perspective. A hybrid model based on Kolmogorov-Arnold Networks (KAN) and self-attention mechanisms was constructed to predict A and B in battery materials. The feasibility of the model is also validated using real experimental data. This fully demonstrates that machine learning methods can efficiently predict the chemical properties of materials, aiding in the development and application of new battery materials.