To appreciate why this matters, it helps to understand what formation actually is. When a lithium-ion cell is assembled, it is not yet a functional battery. It must first be charged and discharged in carefully controlled cycles, a process called formation, during which the solid electrolyte interphase, a critical protective layer on the anode, is established. This stage is among the most expensive and time-consuming steps in battery manufacturing, and manufacturers around the world use different formation protocols, varying the charging currents, voltages, and cycle patterns, in pursuit of faster production or better performance. The trouble is that these protocol differences leave distinct fingerprints on the cell’s long-term degradation behavior. A prediction model trained on batteries formed one way may fail badly when applied to batteries formed another way, which is precisely the problem the research team set out to solve.

Traditional approaches to remaining useful life prediction, the researchers note, often suffer from poor accuracy and weak robustness when batteries have undergone different formation protocols. The core difficulty is that formation-induced differences create systematic biases in the early-cycle data that prediction models rely on. Features extracted from charge-discharge curves may look deceptively similar across cells while hiding fundamentally different aging trajectories. The new method confronts this head-on with a two-pronged strategy: an unscented transformation tailored for variable formation protocols, and a hybrid framework combining an Unscented Particle Filter with a Gated Recurrent Unit neural network, complete with a protocol-sensitivity compensation module designed to mitigate the bias that protocol differences inject into predictions.

The first component of the framework is an exponential empirical degradation model, fitted using an Unscented Particle Filter, that captures the underlying aging mechanism of the battery. Empirical degradation models describe how capacity fades over cycles using mathematical functions, and the particle filter is a Bayesian inference technique that continuously updates the model’s parameters as new measurements arrive. The unscented variant is particularly well suited to this task because it propagates uncertainty through nonlinear transformations without requiring the model to be linearized, a step that can introduce serious errors in battery systems where degradation is anything but linear. By anchoring the prediction in a physically motivated degradation model, the framework avoids the worst failure mode of purely data-driven approaches: extrapolating confidently into regions where the training data offers no guidance.

The second component is where the deep learning enters. The researchers constructed a Gated Recurrent Unit neural network, a streamlined relative of the long short-term memory architecture that has become a workhorse of time-series prediction. The network takes key features from constant-current and constant-voltage charge-discharge cycles as its inputs, treats the battery’s state of health as a hidden state that evolves over time, and outputs the remaining useful life under varying formation protocols. Gated recurrent units are prized for their ability to capture temporal dependencies in sequential data while remaining computationally lighter than full LSTM networks, making them attractive for applications where predictions must be updated continuously as new cycling data streams in.

What makes the architecture genuinely novel is the way these two components collaborate. Rather than letting the neural network operate alone, the framework integrates empirical modeling with data-driven prediction, allowing the particle-filter-based degradation model to refine the network’s RUL estimates while the protocol-sensitivity compensation module adapts dynamically to variations in formation conditions. In effect, the empirical model provides a physically grounded skeleton, the neural network fleshes out the complex, protocol-dependent details, and the compensation module corrects for the systematic distortions that different formation histories introduce. This model-data fusion philosophy reflects a broader trend in battery prognostics, where researchers increasingly recognize that neither pure physics nor pure machine learning is sufficient on its own.

Validation was carried out on public data from the University of Michigan Battery Lab, specifically the fast formation study dataset published through the university’s Deep Blue Data repository. This dataset is a valuable testing ground precisely because it contains batteries cycled under different formation protocols, allowing the researchers to systematically quantify and model the impact of protocol differences on long-term degradation and remaining useful life prediction. The results were striking: the proposed method achieved an R-squared value of 99.41 percent, meaning it explained nearly all of the variance in the observed degradation trajectories, and a root mean square error of 2.23 percent, indicating tight agreement between predicted and actual values. The team reports that the method significantly outperformed comparison methods in both state of health and remaining useful life prediction accuracy, demonstrating superior cross-protocol generalization and robustness compared with existing hybrid approaches.

The practical implications of this work extend well beyond the laboratory. Accurate remaining useful life prediction is essential for determining appropriate battery replacement times, a decision with enormous consequences for safety, cost, and sustainability. In electric vehicles, overestimating a battery’s life risks stranding drivers with a depleted pack, while underestimating it wastes money by replacing cells that still have serviceable capacity. In aerospace and space applications, where some of the co-authors’ institutions operate, the stakes are even higher, since battery failure can compromise entire missions. A prediction framework that remains accurate regardless of how the cells were formed could allow manufacturers to mix batteries from different production lines or formation batches into the same fleet management system without sacrificing prognostic reliability, something current methods struggle to guarantee.

There is also a manufacturing angle that could prove commercially significant. Formation is one of the most energy-intensive and capital-heavy stages of cell production, and the industry has been actively exploring faster formation protocols to cut costs and increase throughput. One reason manufacturers hesitate to adopt aggressive fast-formation schemes is uncertainty about how those choices will affect long-term life and whether existing quality-control and prognostic tools will still work. A prediction method that explicitly compensates for protocol sensitivity could lower that barrier, giving manufacturers the confidence to experiment with formation strategies that shorten production times without losing the ability to forecast field performance. The research team, which includes authors affiliated with CATL, the world’s largest battery manufacturer, is well positioned to understand these industrial pressures.

The study, led by Gaige Chen and Haitao Meng, who contributed equally, with Yahong Yang, Xiaoshen Zhang, Wei Deng, and Jinsong Liu, was supported by the National Natural Science Foundation of China and builds on the group’s earlier work on transfer learning for state of health prediction under cross-formation protocols. As batteries proliferate through every corner of the energy economy, tools like this hybrid UPF-GRU framework point toward a future in which the lifespan of every cell can be forecast accurately from its earliest cycles, regardless of the manufacturing choices that shaped it. That kind of predictive certainty could reshape how batteries are designed, produced, warranted, and retired, turning one of the industry’s most persistent blind spots into a solved problem.

Subject of Research: Machine learning-based remaining useful life prediction of lithium-ion batteries across different formation protocols

Article Title: Remaining useful life prediction of lithium-ion batteries across formation protocols

Article References: Chen, G., Meng, H., Yang, Y., Zhang, X., Deng, W., & Liu, J. (2026). Remaining useful life prediction of lithium-ion batteries across formation protocols. Ionics. https://doi.org/10.1007/s11581-026-07472-5

Image Credits: AI Generated

DOI: 10.1007/s11581-026-07472-5

Keywords: lithium-ion batteries, remaining useful life, formation protocols, unscented particle filter, gated recurrent unit, state of health, battery degradation, machine learning, battery manufacturing, prognostics, CATL, University of Michigan Battery Lab

The research, led by Hima Soni of Gujarat Technological University together with colleagues at Vishwakarma Government Engineering College, Tata Consultancy Services, MultiQoS and MDI Gurgaon, introduces a framework that combines variational autoencoders with actor-critic deep reinforcement learning to estimate remaining useful life sequentially, refining its estimate step by step rather than producing a single one-shot prediction. The result, according to the authors, is a method that outperforms conventional supervised sequence-learning baselines, including long short-term memory networks and multilayer perceptrons, by 26.7 to 40.3 percent in root mean square error, with the best-performing variant reaching an RMSE of 0.1262.

To understand why this is notable, it helps to unpack the machinery underneath. Remaining useful life estimation is fundamentally a problem of reading degradation. Industrial machines are fitted with sensors that stream measurements of temperature, vibration, pressure and dozens of other signals. As components wear, these signals drift in characteristic ways, and the art of prognostics lies in mapping those drifts onto a timeline of future failure. Classical approaches feed sensor histories into supervised models that learn the mapping directly from labeled run-to-failure data. The trouble is that such data is scarce, noisy and expensive to collect, and the models often struggle to generalize across different degradation patterns.

The new framework attacks the problem in two stages. First, a variational autoencoder compresses the raw sensor streams into a compact latent representation. A variational autoencoder is a generative neural network that learns to encode data into a probabilistic latent space and then decode it back, forcing the latent variables to capture the essential structure of the data. In this application, the latent space preserves the underlying degradation dynamics while discarding redundant or irrelevant sensor noise. Crucially, the trained autoencoder then serves as a surrogate environment, a simulated world in which a reinforcement learning agent can act without touching real machinery or requiring vast amounts of fresh field data.

Inside that surrogate environment, the researchers deployed two actor-critic deep reinforcement learning algorithms: Twin Delayed Deep Deterministic Policy Gradient, known as TD3, and Proximal Policy Optimization, known as PPO. Both algorithms learn a policy, a rule for choosing actions, while simultaneously learning a value function that estimates how good those actions are. The agent’s task is to iteratively refine degradation trajectories within the learned latent space, adjusting its estimates of remaining useful life in a sequence of decisions rather than a single leap. This sequential formulation turns prognostics into a control problem, allowing the model to correct course as new evidence accumulates.

One of the study’s most interesting findings concerns reward shaping, the practice of designing the feedback signal that guides reinforcement learning. The researchers compared multiple reward formulations, including a weighted cosine similarity reward that measures how closely the agent’s predicted trajectory aligns with the true degradation path. Their comparative evaluation showed that the choice of optimization strategy meaningfully shapes both learning stability and downstream prediction accuracy. The on-policy approach, PPO, yielded more consistent convergence in 11 of 12 experimental configurations, a result that speaks to the well-known stability advantages of on-policy methods, which evaluate and update the same policy they use to collect experience. Meanwhile, the weighted cosine similarity reward produced comparable prediction accuracy, with large Cohen’s d effect sizes exceeding 0.8 in 10 configurations, indicating that the reward design produced differences large enough to matter in practice, not merely statistical noise.

The performance numbers are striking. An RMSE of 0.1262 on remaining useful life estimation, achieved while beating LSTM and MLP baselines by margins approaching 40 percent, suggests that the latent-space policy optimization is doing genuine work rather than simply repackaging existing sequence-learning tricks. The authors attribute this to the synergy between generative representation learning and sequential policy optimization: the autoencoder gives the agent a clean, information-dense state space, while the reinforcement learning loop gives the model a mechanism for iterative refinement that supervised models lack. Because the agent learns within a surrogate environment, the architecture is also computationally scalable and data efficient, qualities that matter enormously in industrial settings where labeled failure data is the bottleneck.

The broader context makes the contribution clearer. Predictive maintenance has accumulated a rich toolkit over the past decade, from deep convolutional neural networks for prognostics to Bayesian neural networks that quantify uncertainty, from regularized LSTM models for rolling bearings to ensemble deep learning with multi-objective optimization for rotating machinery. Reinforcement learning has already appeared in maintenance research, notably in scheduling and optimization roles, such as deep reinforcement learning for predictive aircraft maintenance using probabilistic prognostics. What distinguishes the new work is its modular marriage of a generative surrogate environment with policy optimization directly in latent space, a design the authors argue is adaptable to a variety of degradation patterns and therefore suited to challenging industrial scenarios where one-size-fits-all models falter.

There are, of course, caveats worth keeping in mind. The study reports that no datasets were generated or analysed during the current study in its data availability statement, and the evaluation rests on the configurations the authors describe, so practitioners will want to validate the approach on their own fleets before deploying it. Reinforcement learning remains notoriously sensitive to hyperparameters and reward design, and although the comparative evaluation across 12 configurations is a step toward taming that sensitivity, the finding that PPO converges more consistently than TD3 in most settings is itself a reminder that algorithm choice is not a detail but a first-order design decision. The authors also emphasize that the framework is modular, which means each component, the autoencoder, the policy optimizer and the reward function, can in principle be swapped as better versions emerge.

Still, the direction is compelling. If machines could reliably tell operators how much life they have left, maintenance could shift from reactive scrambling and wasteful over-servicing toward precisely timed intervention, saving money and preventing failures before they happen. By reframing remaining useful life estimation as sequential policy optimization in a learned latent space, this work offers a template for how generative models and reinforcement learning might together make that vision practical. The authors thank Gujarat Technological University and Bacancy DataProphets Private Limited for supporting the research, and their results suggest that the intersection of representation learning, reward design and control may be one of the more productive frontiers in reliability engineering for years to come.

Subject of Research: A VAE-based deep reinforcement learning framework with reward shaping for sequential estimation of machinery remaining useful life.

Article Title: Sequential Estimation of Remaining Useful Life Through VAE-Based Deep Reinforcement Learning With Reward Shaping

Article References: Sequential Estimation of Remaining Useful Life Through VAE-Based Deep Reinforcement Learning With Reward Shaping. (n.d.). https://doi.org/10.1007/s10994-026-07152-5

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07152-5

Keywords: predictive maintenance, remaining useful life, deep reinforcement learning, variational autoencoder, reward shaping, TD3, PPO, prognostics, latent space, reliability engineering, machine learning, sensor data