Can machine learning create a useful reference-electrode signal without putting a reference electrode in the battery?

The electrochemical problem

Battery-management systems normally see cell voltage, current, and temperature, but dangerous processes can begin at one electrode before they create an obvious whole-cell signature. Lithium plating on graphite is a key example: negative-electrode potential would reveal the risk directly, yet a physical reference electrode is impractical in most commercial cells.

The paper asks whether that hidden electrode potential can be reconstructed from signals already available in a conventional two-electrode cell.

What the researchers did

The authors generated training data using three-electrode cells, where negative-electrode potential can be measured directly. A machine-learning model then learned to estimate that internal potential using only variables available from a two-electrode cell.

The predicted potential was compared with the measured reference signal and with microscopy evidence for the transition from lithium intercalation to metallic lithium plating. The virtual signal was then used in an adaptive charging protocol.

What they found

The virtual reference electrode predicted negative-electrode potential with a root-mean-square error of 0.023 V and a mean absolute error of 0.018 V. Electron microscopy supported the predicted transition near the thermodynamic plating threshold.

When the estimate was used to reduce charging current as plating risk increased, cycle life was extended by 8.25 times relative to a constant-current/constant-voltage baseline under the tested low negative-to-positive capacity-ratio condition.

Why this matters

The predicted quantity is physically interpretable. Rather than asking the model for an opaque health score, the system reconstructs an electrode potential that electrochemists already know how to use.

That creates a direct path from machine-learning inference to control: predict the hidden state, detect risk, and change the applied current before the failure mode accelerates.

One important caution

The 8.25× lifetime improvement is specific to the tested low N/P-ratio cells and charging comparison. It should not be reported as a universal lifetime gain for all lithium-ion batteries.

The model also depends on training data from instrumented three-electrode cells. Transfer to different chemistries, formats, ages, and operating conditions requires validation of whether the learned relationship between ordinary cell signals and negative-electrode potential remains accurate.