Fig. 7: Evaluation of M(B)Na’s predictive performance and input sensitivity on our in-house Na-ion data.
From: Attention towards chemistry agnostic and explainable battery lifetime prediction
Plot a shows the discharge current profile, while plots b–d depict the predictions for voltage drop, CE, and Qdis against the ground truth. The color bar here shows the aggregated attention weights across the input data. Plots e–g provide a detailed logarithmic sensitivity analysis per predictive cycle for each input parameter, and plots h–j aggregate these sensitivities, highlighting the model’s focus on different input cycles, especially the most recent ones, reflecting M(B)Na’s protocol adaptability and robust response to experimental noise.
