EntroLnn : Entropy-Guided Liquid Neural Networks for Operando Refinement of Battery Capacity Fade Trajectories Conference: The 41st ACM/SIGAPP Symposium on Applied Computing; March 23–27, 2026; Thessaloniki, Greece The 41st ACM/SIGAPP Symposium on Applied Computing (SAC ’26), March 23–27, 2026, Thes

Wei Li, Wei Zhang, Qingyu Yan · 2026

Read EntroLnn : Entropy-Guided Liquid Neural Networks for Operando Refinement of Battery Capacity Fade Trajectories Conference: The 41st ACM/SIGAPP Symposium on Applied Computing; March 23–27, 2026; Thessaloniki, Greece The 41st ACM/SIGAPP Symposium on Applied Computing (SAC ’26), March 23–27, 2026, Thes by Wei Li, Wei Zhang, Qingyu Yan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06195.txt.

Introduction

Battery health tools usually estimate one number at a time. Wei Li and 2 co-authors instead refine the whole capacity fade trajectory in a 2026 paper, tracking how a battery's capacity declines over its life during actual use. That fuller picture supports better management decisions.

Their EntroLnn framework uses entropy-guided liquid neural networks, a flexible model family suited to time-varying behavior. New entropy-based features come from online temperature fields, applied to battery analytics for the first time here. The system adapts both its fixed structure and its moment-to-moment response. Across different batteries and conditions, trajectory errors stayed very small and end-of-life predictions landed close, with lightweight computation suited to practical battery management systems.

Catalog record

CREATORWei Li, Wei Zhang, Qingyu Yan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-06195

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

EntroLnn : Entropy-Guided Liquid Neural Networks for Operando Refinement of Battery Capacity Fade Trajectories Conference: The 41st ACM/SIGAPP Symposium on Applied Computing; March 23–27, 2026; Thessaloniki, Greece The 41st ACM/SIGAPP Symposium on Applied Computing (SAC ’26), March 23–27, 2026, Thessaloniki, Greece DOI: 10.1145/3748522.3779909 ISBN: 979-8-4007-2294-3/2026/03 CCS: Computing methodologies Machine learning algorithms CCS: Applied computing Engineering CCS: Computing methodologies Modeling methodologies Wei Li, Wei Zhang, Qingyu Yan Abstract . Battery capacity degradation prediction has long been a central topic in battery health analytics, and most studies focus on state of health (SoH) estimation and end of life (EoL) prediction. This study extends the scope to online refinement of the entire capacity fade trajectory (CFT) through EntroLnn , a framework based on entro py-guided transformable liquid neural networks ( LNN s). EntroLnn treats CFT refinement as an integrated process rather than two independent tasks for pointwise SoH and EoL. We introduce entropy-based features derived from online temperature fields, applied for the first time in battery analytics, and combine them with customized LNNs that model temporal battery dynamics effectively. The framework enhances both static and dynamic adaptability of LNNs and achieves robust and generalizable CFT refinement across different batteries and operating conditions. The approach provides a high fidelity battery health model with lightweight computation, achieving mean absolute errors of only 0.004577 for CFT and 18 cycles for EoL prediction. This work establishes a foundation for entropy-informed learning in battery analytics and enables self-adaptive, lightweight, and interpretable battery health prediction in practical battery management systems. † † cc-license: by 1. Introduction Capacity degradation is an inevitable process in lithium-ion batteries (LIBs) and has long been a central topic in their application and development 19 . Recent studies have shown advances in battery degradation prediction in terms of the estimations of state of health (SoH), remaining useful life (RUL), and end of life (EoL) 16 ; 4 . In general, the methodological focus of battery health modelling has gradually shifted from traditional data-driven methods, such as Gaussian process regression, to more advanced machine-learning (ML) algorithms that can model nonlinear and high-dimensional behaviors 17 ; 6 . More recently, purely data-driven models have evolved into physics-informed ML frameworks, which integrate physical knowledge and ML to enhance modelling accuracy, robustness, and interpretability 13 . These advancements show a transition from empirical parameter fitting to ML models capable of capturing complex electrochemical degradation patterns. However, significant challenges remain in developing generalizable, interpretable, and sensor-efficient models for battery degradation. In practice, batteries operate under diverse operating conditions that may differ from those in controlled laboratory environments. Most existing models rely on complete measurements of current, voltage, and cycle data, and often assume homogeneous operating conditions; consequently, they tend to lose accuracy and reliability in real systems. To support practical battery analytics, degradation models must be both data-efficient and physically meaningful, with operando adaptability to track degradation under real-world conditions 10 . Several recent approaches have considered the above challenges by improving data efficiency and incorporating physical constraints into learning models, and have achieved progresses in SoH and EoL prediction accuracy 20 . In 12 , one-dimensional convolutional neural networks (1D-CNNs) and Transformer models are applied to model temporal dependencies within batteries and estimate SoH from partial cycle data. Bidirectional long short-term memory (Bi-LSTM) networks are employed in 15 to reconstruct capacity-fade trajectories (CFTs) to provide a more continuous representation of degradation compared with single-point estimation. In parallel, liquid neural networks (LNNs) 8 have emerged as a promising ML algorithm for modeling dynamic processes through continuous-time state evolution. While good performance has been reported for these models, their scalability remains limited with insufficient physical grounding, and the modeled degradation patterns are specialized to certain datasets without sufficient generalizability 9 . To enhance generalization, recent models have been developed to integrate electrochemical insights with ML 14 ; 5 . These models combine the flexibility of data-driven learning with physical modeling to achieve CFT refinement without being dominated by specific data distributions. However, these models still rely on large datasets that span diverse operating conditions, since their embedded electrochemical insights are themselves parameterized from data. As such, their learned physical constraints often fail to generalize under unseen conditions, and the resulting representations remain largely empirical. This reflects a deep conceptual limitation, that is the absence of universal and physically meaningful features that can quantify degradation consistently across diverse batteries and operating conditions. Figure 1. The holistic framework of EntroLnn . The model integrates multiple features of LIBs, including temperature curves, entropy features derived from temperature profiles, and additional attributes extracted from entropy curves. EntroLnn adapts to real-time CFT refinement across different batteries and maintains high generalizability and stable performance. To address the absence, we introduce entropy as a set of statistical features of internal degradation of batteries. Our proposed entropy is inspired by information entropy, which quantifies the uncertainty within a distribution. We extend this principle to describe the spatial and temporal irregularity of the battery temperature field during degradation, where the irregularity tends to increase as degradation progresses. In this way, entropy provides a physically inspired and thermally interpretable descriptor to reflect the growing temperature heterogeneity during battery aging. Unlike previous models that use data-parameterized electrochemical equations, our proposed entropy is derived directly from measured temperature. The entropy is insensitive to short-term fluctuations like noises and highlights consistent thermal patterns associated with degradation. As such, entropy in this study serves as an interpretable and operando-computable indicator that offers a unified representation connecting the evolving temperature field and degradation for real-time CFT refinement. Based on the proposed entropy formulation, we develop EntroLnn , the entro py-guided transformable LNN s designed to refine battery CFTs in real-time. The overall workflow consists of two main components, including constructing entropy-based degradation indicators and training transformable LNNs. Ambient temperature and current measurements are first used to reconstruct the temperature field of a battery. The field is then converted into entropy features that quantify the dispersion of heat generation across space and time. These features, together with voltage and temperature data, form an entropy-based SoH indicator that captures the joint evolution of thermal variability and battery degradation. The indicator establishes an interpretable link between entropy and measurable battery capacity loss and serves as the physical input to our LNN models. In the first stage, a static LNN is trained using the data of a single long-lifespan reference battery, which has stable early-cycle degradation dynamics and can be the foundation for CFT representation. The static LNN cannot perfectly generalize to entire CFTs. So, we train a dynamic LNN using capacity-fade data from several batteries with different degradation profiles to learn the temporal evolution of degradation until the EoL of each battery. During EntroLnn deployment, the two LNNs operate jointly. Early-cycle data, e.g., the first 100 cycles of SoH and entropy-based features, are used to fine-tune the static LNN for generating CFT, which is primarily optimized for the early degradation stage. When the SoH decreases beyond, for example, 98%, the dynamic LNN adapts to operating conditions and extends the refined trajectory until EoL. We have conducted extensive experimental studies using a public dataset, and the proposed EntroLnn demonstrates strong performance in both accuracy and adaptability across different batteries. With

Read the complete text