Sequential Bayesian optimization for accelerating the design of sodium metal battery nucleation layers

Abstract

Complex interactions between battery materials make design optimization difficult. In this work, we quickly and efficiently optimize three interacting design variables of a carbon nucleation layer used in sodium metal batteries. Unlike existing materials optimization approaches focusing on maximizing favorable material properties, we execute a multi-parameter materials optimization scheme that uses cell-level aging data as the target objective. We employ a Bayesian optimization algorithm that intelligently selects the nucleation layer designs to test in sequence, quickly finding the design that yields the highest lifetime. Results from simulation studies conducted after the fact indicate that a well-tuned Bayesian optimization algorithm can optimize the nucleation layer properties roughly five times faster than a random sampling approach and roughly two times faster than a poorly-tuned Bayesian algorithm. We then examine correlations on all 177 cells with different nucleation layers and use electrochemical impedance spectroscopy and imaging to propose the primary mechanism for nucleation layer performance, considering dead sodium as the dominant mode of capacity loss. The algorithms, hyperparameter tuning strategies, and post-optimization mechanism analysis used in this work broadly apply to other battery chemistries and electrode designs and are essential for quickly bringing metal battery performance on par with existing battery technologies.