Battery lifetime depends on chemistry, temperature and the way a cell is charged and discharged. Fully cycling every design can take a long time, but models trained under one narrow set of conditions often perform poorly when those conditions change. ^[1]

BatLiNet was designed to learn both from a cell's own early-cycle changes and from differences between that cell and reference cells whose lifetimes are already known. ^[1]

The researchers combined public datasets spanning different cycling protocols, temperatures, pack structures and cathode chemistries. They evaluated five setups, including predictions from the first 100 cycles and a harder mixed-data task using the first 20 cycles. ^[1]

Across the five evaluation sets, BatLiNet reduced root-mean-square error against the best comparison model by amounts ranging from 6.8% to 40.1%. Compared with a similar neural network using only within-cell changes, its average percentage error fell by up to 40%. ^[1]

In transfer tests, the model used 275 lithium-iron-phosphate cells as a well-supplied reference and tested learning for smaller groups of other chemistries. Comparing cells helped in most low-data settings, although the benefit varied with the target chemistry and number of examples. ^[1]

The mixed datasets broaden the test beyond a single laboratory, but they still represent a limited set of cells and protocols. Batteries that reached end of life during the early observation window were excluded from two mixed-data tasks, and reference-cell choice can affect error. ^[1]

The authors say physical understanding remains essential. Prospective tests on new cells, calendar ageing and field conditions are needed before treating the model as a dependable battery-management tool. ^[1]