Seawater electrolyte-based metal-air batteries (SMABs) have emerged as attractive energy storage and conversion systems for deep-sea exploration and offshore engineering applications, owing to their inherent safety, eco-friendliness, and low cost. The oxygen reduction reaction (ORR) is the core reaction of SMABs, and its catalyst performance directly affects the device’s discharge efficiency and operating time. However, ORR catalysts are extremely sensitive to the complex ionic environment of seawater, including high concentrations of Cl⁻, Mg²⁺, and Ca²⁺. Even subtle differences in catalyst composition or structure can lead to drastically distinct discharge curves and performance decay patterns. This complex reaction behavior has led to the absence of a unified quantitative criterion for accurately evaluating the end-of-discharge (EOD) point and operating duration of SMABs. Such uncertainty not only undermines the stability of energy supply for marine equipment but also elevates the risks associated with engineering operation and system management. Therefore, establishing a reliable quantitative method to predict the discharge duration and EOD conditions of SMABs is a crucial prerequisite for promoting practical engineering applications.
Currently, the performance evaluation of ORR catalysts primarily relies on two classic electrochemical characterization techniques: linear sweep voltammetry (LSV) and chronoamperometry (CA). LSV is widely used to assess the initial electrochemical activity of catalysts, with key indicators including onset potential, half-wave potential (E_1/2), and limiting current density, directly reflecting ORR reaction kinetics and mass transport limitations. CA is employed to characterize current retention and decay behavior under constant-potential conditions, thereby effectively capturing the deactivation trends of active sites and carbon supports in the complex seawater environment. Although both LSV and CA measurements have clear physical correlations with the sustained discharge capability of SMABs, a unified quantitative framework that can reliably map the LSV and CA performance data to the actual discharge duration or EOD point remains lacking. Traditional statistical regression methods rely on manually engineered features and linear/weakly nonlinear assumptions, which are prone to introducing feature extraction bias and model fitting errors under multi-factor coupling conditions (e.g., ion interference, interface reactions, catalyst degradation) and time-varying noise that is inherent in SMABs. These limitations severely restrict the engineering reliability and practical applicability of traditional methods for SMABs’ performance prediction.
Neural network-based time series regression offers a feasible and effective solution to the aforementioned challenges. Deep learning models can automatically extract multi-scale features directly from raw LSV and CA curves without manual intervention, enabling efficient learning of the nonlinear coupling relationships between these electrochemical features and battery discharge performance. Furthermore, the deep time series regression framework facilitates the quantitative prediction of key parameters (e.g., discharge duration, EOD point) by modeling the correlation between input electrochemical performance sequences and external target variables. This capability lays a solid foundation for transforming standard electrochemical performance curves into quantitative tools for engineering-oriented catalyst screening and evaluation.
The research team led by Professor Xinlong Tian from Hainan University constructed a dual-branch time-series regression model incorporating an InceptionTime backbone with prior-biased attention pooling (PBAP) to learn the nonlinear relationship between electrochemical performance sequences and the discharge duration of SMABs. The main branch takes CA sequences as input and extracts multi-scale temporal features reflecting the long-term stability of catalysts via the InceptionTime backbone, while the auxiliary branch takes LSV curves as input and constructs an attention mechanism combined with electrochemical physical priors. This attention mechanism is designed to enhance feature weighting of critical potential regions (e.g., E_1/2 and limiting current density plateaus), which are closely related to ORR activity and thus battery discharge performance. The proposed model was validated on a small-sample test set consisting of multiple different air cathode catalysts, achieving a coefficient of determination (R²) exceeding 0.90 across different aggregation strategies, indicating favourable prediction accuracy within the current dataset. This work not only achieves competitive performance metrics at the interdisciplinary interface of electrochemistry and artificial intelligence, but also provides a preliminary data-driven approach for catalyst evaluation and discharge-time prediction in seawater metal-air batteries.
The work entitled “Neural Network Driven by Electrochemical Performance Data for Predicting the Discharge Termination Time of Seawater Electrolyte-Based Metal-Air Batteries” was published in Journal of Electrochemistry (published on Aug. 28, 2026).
DOI：10.61558/2993-074X.3616