<p>The physics-based Doyle-Fuller-Newman (DFN) model, widely adopted for its precise electrochemical modeling, stands out among various simulation models of lithium-ion batteries (LIBs). Although the DFN model is powerful in forward predictive analysis, efficient methods for the uncertainty quantification (UQ) of the key parameters over the performance of LIBs remain a long-standing challenge in the field. Numerous physical parameters are needed for resolving the nonlinear, time-dependent, and multiscale DFN model, yet many of them are difficult to be determined precisely via experimental measurement, hence hindering the practical use of this model. To tackle this challenge, we propose the extended dual order reduced Gaussian process (ExDORGP), a novel high-efficiency data-driven method for accurate UQ of LIBs. The method is designed for the accurate prediction of time-varying battery information, the data volumes of which can vary significantly across different operating conditions. For certain special and extreme conditions, the data can be highly limited. To achieve this, ExDORGP employs an intelligent tree-like framework: a root model maps input parameters to information under a root operating condition. Then, extended models within this framework map information from the root condition to other operating conditions, thereby enabling effective model extrapolation. We use LiCoO₂ battery as our benchmarks to test the accuracy of our proposed method, and results show that the method can achieve high predictive accuracy even with limited data. We further use this method to assess the influences over the performances of LIBs due to the uncertainties of key physical parameters, including the reaction rate constant, the Bruggeman constants, the initial lithium concentration, etc. These results suggest that ExDORGP is a versatile and powerful method for the UQs of LIBs, and thereby advancing the development and health monitoring of high-performance LIBs under complex working conditions.</p>