⚡ electrical engineering

Endpoint definitions and censoring in battery lifetime prediction across sources

This paper analyzes how endpoint definitions and censoring strategies significantly impact battery lifetime prediction accuracy across different data sources, advocating for a standardized engineering evaluation procedure that explicitly declares references, separates observed and censored outcomes, and reports cohort-level error bounds to ensure reliable comparisons.

Original authors: Jufeng Li, Wenjie Li, Tao Li

Published 2026-10-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jufeng Li, Wenjie Li, Tao Li

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern technology, the battery is the silent engine that powers everything from smartphones to electric vehicles. For engineers and scientists, a critical challenge is predicting how long a battery will last before it can no longer hold a sufficient charge. This is not a simple matter of counting days; it involves watching a battery slowly lose its ability to store energy over hundreds of charging cycles. To make these predictions, researchers often set a specific "retirement" point, such as the moment a battery's capacity drops to 80 percent of its original strength. However, a major problem arises when scientists try to compare results from different laboratories. One lab might define that 80 percent threshold based on the battery's factory rating, while another might base it on the very first measurement they took. Furthermore, many batteries in a study do not run until they die; their testing stops early, leaving their final lifespan unknown. If researchers ignore these unfinished stories or mix up the rules for when a battery "retires," they risk drawing false conclusions about which prediction methods actually work.

A team of researchers from the China University of Petroleum set out to audit how these hidden rules and incomplete data affect the reliability of battery lifetime predictions. They did not invent a new computer algorithm or a faster way to predict the future. Instead, they acted as accountants for the field, examining how different definitions and data-handling choices change the final numbers. They gathered data from seven different sources, covering 220 physical battery cells and over 300,000 recorded cycles. Their investigation focused on three specific decisions that researchers make: how to define the retirement capacity, how to treat batteries that stop testing before they retire, and how to fairly compare new models against old ones using a limited number of completed examples.

The researchers found that the choice of definition alone can completely alter the outcome of a study. When they applied two different ways of defining the retirement point to the same set of batteries, the results were startlingly different. For one group of batteries from the HUST source, changing the definition meant the number of batteries that had actually "retired" jumped from zero to 77. For another group from the SNL source, the change caused 15 out of 61 batteries to switch from being considered "finished" to "still running." This means that two studies looking at the exact same batteries could report entirely different success rates simply because they used different rulers to measure the end of life. The study emphasizes that without explicitly stating which ruler was used, a reported prediction number is meaningless because it might be answering a different question than the reader thinks.

The team also tackled the issue of incomplete data, or batteries that were taken out of the test before they died. In many studies, scientists only calculate errors for the batteries that finished their full life, ignoring the ones that stopped early. The researchers showed that this practice hides a significant amount of error. For the SNL data, a model that looked perfect when judging only the finished batteries actually had a much higher error rate when the unfinished ones were included. By accounting for the fact that these unfinished batteries lasted longer than their last recorded cycle, the researchers calculated a realistic error range for the entire group, which was significantly worse than the score reported by looking only at the finished cells. This revealed that the prediction model was systematically underestimating the life of long-lasting batteries.

When the researchers tested methods to improve predictions using a small number of completed examples, the results depended heavily on which battery group was being studied. For the HUST data, using just three completed battery lifetimes to adjust the model reduced the average error dramatically, from over 75 percent down to about 16 percent. This improvement was consistent across thousands of different ways of choosing those three batteries. However, when they applied the exact same adjustment method to the SNL data, the results were mixed. While some adjustments helped, others made the predictions much worse, with errors spiking to over 200 percent in certain cases. This demonstrated that a method that works beautifully for one type of battery chemistry or testing protocol can be harmful for another. The researchers concluded that simply reporting a single average number is dangerous because it can hide these extreme failures.

The study concludes with a call for a new standard in how battery research is reported. The authors propose a checklist for scientists to ensure their findings are clear and comparable. This includes declaring exactly how the retirement point was defined, separating finished batteries from those that were stopped early, and reporting the error range for the entire group rather than just the finished ones. They also insist that any new model must be compared against simple, constant baselines using the exact same set of data to prove it offers real value. By following these steps, the scientific community can stop comparing apples to oranges and build a more honest, reliable understanding of how long batteries will truly last. The work does not solve the problem of predicting battery life, but it provides the necessary tools to ensure that when a prediction is made, everyone knows exactly what it means and how much trust to place in it.

Technical Summary: Endpoint Definitions and Censoring in Battery Lifetime Prediction Across Sources

1. Problem Statement

The field of lithium-ion battery lifetime prediction has expanded from individual experimental campaigns to large, heterogeneous collections of public datasets (e.g., BatteryLife). However, reliable cross-source comparison of lifetime predictors is hindered by three critical, often unreported, evaluation decisions:

Endpoint Definition: The reference capacity used to define a "percentage" threshold (e.g., 80% End of Life) varies significantly. It may be based on nominal nameplate capacity, the first observed discharge capacity, or nominal capacity adjusted for State-of-Charge (SOC) window width. These choices define different absolute capacities and, consequently, different retirement events.
Treatment of Censoring: Many cells in a cohort do not reach the defined threshold within the available data. Benchmarks often report accuracy only on cells that reached the endpoint (observed events), ignoring the "right-censored" cells whose lifetimes are known only to be greater than the last observed cycle. This biases evaluation toward the full cohort a deployment would face.
Label Budget and Calibration: When adapting a transferred model to a new source using a small number of completed target lifetimes (e.g., three cells), evaluations often fail to exclude these specific cells from the test set or fail to compare against simple constants derived from the same labels. This can conflate access to target scale information with genuine predictive skill.

This study does not propose a new dataset, network architecture, or learning method. Instead, it conducts an empirical audit of how these reporting decisions affect the interpretation of lifetime numbers across different experimental sources.

2. Methodology

The authors analyzed a registry of seven sources (CALB, CALCE, HNEI, MICH_EXP, UL_PUR, SNL, HUST) containing 220 physical cells and 318,707 cycle records. The study utilized two specific historical, source-held-out records as case studies:

SNL: Evaluated at an 80% threshold ( $\theta=0.80$ ) with 50 observed events and 11 right-censored cells.
HUST: Evaluated at an 85% threshold ( $\theta=0.85$ ) where all 77 cells were observed (the 80% threshold yielded no observed events for this source).

Prediction Protocol:

Model: A simple ridge regression predicting log-lifetime from three prefix features (relative capacity at cycle 100, slope, and curvature) plus the threshold. The model was "frozen" (no retraining or hyperparameter tuning) to isolate the impact of reporting decisions.
Calibration: A "few-label" adaptation was tested using three completed target lifetimes. Methods included:
- Zero-shot (no correction).
- Variance-informed shrunk median correction (Empirical-Bayes style).
- Ordinary unshrunk log-median correction.
- Target-only constants (arithmetic mean, median, geometric mean, and MAPE-minimizing constant).
Exhaustive Allocation: To assess sensitivity, the authors performed an exhaustive analysis of all possible 3-cell calibration sets: 73,150 allocations for HUST and 19,600 allocations for SNL.
Censoring Handling: For right-censored cells, the study did not impute lifetimes. Instead, it calculated an identification range for the Mean Absolute Percentage Error (MAPE) based on the bounds of the censored survival data ( $T > L$ ).

3. Key Results

3.1 Impact of Endpoint Definitions

Changing the reference capacity definition significantly alters the observable dataset:

HUST: Switching from a nominal reference to a first-observation reference changed the number of observed events at the 80% threshold from 0 to 77.
SNL: The same change altered the status of 15 out of 61 cells, reducing observed events from 50 to 35.
Conclusion: A benchmark's sample size and event set are functions of the endpoint contract. Studies reporting different event counts for the same cells are answering different questions, not measuring different campaigns.

3.2 Full-Cohort vs. Event-Only Evaluation (SNL)

Event-Only MAPE: 36.31% on 50 observed cells.
Full-Cohort MAPE: When including the 11 right-censored cells, the identification range for the full cohort MAPE is 45.37%–47.80%. The event-only score understates the cohort error by ~9 percentage points because the predictor systematically underestimates long-lived cells (predictions fall below the observed survival bound).
Coverage: The prediction interval covered 60.0% of observed events but only 49.18% of the full 61-cell cohort, as the upper bound of the interval failed to cover the survival bound for all censored cells.

3.3 Calibration Sensitivity and Allocation Variability

HUST Results: The shrunk correction improved performance in all 73,150 allocations compared to zero-shot prediction, with a mean paired difference of –59.39 percentage points (indicating the corrected error was lower). It also outperformed matched target-only constants in 80.30% of allocations. However, a specific fixed allocation (historically used) yielded 12.77% error, while the allocation average was 16.25%, demonstrating that reporting a single fixed allocation can hide sensitivity to label choice.
SNL Results: The same correction strategy behaved differently. While the matched zero-shot score was improved in 58.78% of allocations, the median error actually increased slightly (from 36.31% to 36.11% is incorrect; the paper states the median is 36.11% which is slightly below the zero-shot value of 36.31%, but the mean error increased significantly from 36.31% to 48.50%). This increase in the mean is driven by a "harmful tail" of large errors in a minority of allocations (97.5th percentile: 215.98%).
Conclusion: A single reported number (mean or median) is insufficient to describe label-based adaptation on heterogeneous cohorts. The distribution of allocation errors must be reported.

4. Significance and Claims

The paper explicitly claims modest contributions, focusing on evaluation protocol and reporting standards rather than new predictive algorithms. Its primary significance lies in:

Defining an "Endpoint Contract": Establishing that battery lifetime numbers are only interpretable if the reference capacity, threshold, and censoring treatment are explicitly declared.
Exposing Evaluation Bias: Demonstrating that ignoring censored data and using fixed calibration sets can lead to misleadingly optimistic or pessimistic performance metrics.
Proposing a Reporting Checklist: The authors derive a six-item protocol for cross-source studies:
- Declare the reference capacity and SOC convention.
- Separate observed, censored, and extrapolated outcomes.
- Report both event-only scores and full-cohort identification ranges.
- Report interval width alongside coverage.
- Match every calibration comparator (same labels, same excluded cells).
- Show allocation sensitivity (mean, median, and tail) rather than relying on a single fixed allocation.

The study concludes that these checks are inexpensive to implement using existing data tables and are essential for distinguishing genuine transferable predictive skill from artifacts of data definition and label selection. The authors caution that their findings are retrospective case studies and do not isolate specific causes (chemistry vs. protocol) for the observed differences in calibration behavior between SNL and HUST.