Fig. 8: Comparative analysis of M(P)Na and M(B)Na on Qdis prediction for Na-ion batteries.
From: Attention towards chemistry agnostic and explainable battery lifetime prediction
Prediciton analysis for M(P)Na (Panel I) and M(B)Na (Panel II) for Qdis prediction of Na-ion batteries. The scatter plots a illustrate the models' alignment with actual measurements. Density plots b compare the distributions of predicted and actual values, demonstrating the models' accuracy in estimating Qdis. Calibration plots in c depict how well the predicted probabilities match the observed outcomes against the benchmark line, with the discrete points representing the observed proportions of actual values that fall within three quantile intervals. Both models demonstrate a pattern of marginal overconfidence below the 70th percentile and a slight underconfidence above this percentile, as evidenced by the calibration points positions beneath and above the diagonal line, respectively. M(P)Na shows a larger area of divergence, A = 0.06, while M(B)Na presents a closer fit with a miscalibration of 0.053, highlighting both models’ well-calibrated prediction capabilities across different chemistries. Boxplots d visualize the spread and consistency of prediction intervals across predicted cycles. Histograms in e represent the distribution of the quantile intervals of the models’ prediction, highlighting uncertainty; these distributions indicate where, within the prediction range, the models’ confidence is concentrated, with sharpness values of 1.7 × 10−5 for M(P)Na and 2.0 × 10−5 forM(B)Na, demonstrating a precise estimation of uncertainty.
