In my talk, I will show that a combination of battery `biomarkers’ derived from electrochemical impedance signals with probabilistic machine learning enables us to accurately forecast future battery performance amid uneven usage, even when cell history is completely unknown. More broadly, our results bring into question the concept of a scalar State of Health, and instead suggest that battery state is better quantified by a multidimensional vector. I will discuss how this insight lays the foundations for novel algorithms for battery prognostics and control. I will further outline our application of such algorithms in optimal protocol design for intelligent battery charging, and low-data inference via transfer learning across cell chemistries and usage patterns.