Abstract: Accurate temperature monitoring is crucial for ensuring the thermal safety and performance of lithium-ion batteries (LIBs), which are extensively used in electric vehicles, robotics, and energy storage systems. The optimal operating temperature for LIBs is strictly limited between 20 oC and 40 oC. Temperatures outside this range can lead to performance degradation, capacity loss, accelerated aging, or even thermal runaway. Conventional methods, such as sensor-based measurements and electrothermal modeling, face challenges in providing rapid and comprehensive temperature distribution data for large-format batteries due to spatial and cost limitations, which hinder sensor deployment. To address these challenges, this study introduces a novel temperature field modeling and reconstruction approach for ternary lithium batteries using Physics-Informed Neural Networks (PINNs). This method integrates battery thermal modeling with deep learning techniques, enabling real-time, sensor-free monitoring of spatiotemporal temperature distributions within the battery system. The primary aim of this research is to develop a predictive model capable of estimating the temperature field of large-format batteries under various operational conditions without requiring extensive sensor networks. The proposed PINN-based method constructs a physical model of the battery, incorporating experimental data and coupling the heat generation rate equation—derived from an enhanced Bernardi model accounting for Joule heating and reversible reaction heat—with a data-driven nonlinear mapping. This integration improves prediction accuracy and allows for the online identification of unknown battery parameters, such as internal resistance, hereby enhancing the model's ability to adapt to diverse scenarios. The PINN framework incorporates the heat transfer partial differential equation as a constraint within the loss function of the neural network, which includes data loss, physics loss, boundary conditions, and initial conditions. The network architecture consists of an input layer featuring variables such as charging time, current, equivalent resistance, and temperature, followed by six hidden layers (each containing 64 neurons) and an output layer that predicts surface temperature. Experimental validation was performed using a 50 Ah prismatic ternary LIB under constant-current charging/discharging (ranging from 0.5C to 3C) and dynamic conditions, including the Urban Dynamometer Driving Schedule (UDDS). Temperature data were collected from 12 thermocouples placed on the battery surface, with environmental temperatures ranging from 1 oC to 40 oC. The results demonstrate that the PINN model achieves a maximum root mean square error (RMSE) and mean absolute error (MAE) of less than 0.6 oC under constant-current conditions, outperforming traditional multilayer perceptron (MLP) and long short-term memory (LSTM) models, particularly in sensor-free scenarios. In dynamic UDDS conditions, the PINN model maintained high accuracy (RMSE < 0.9 oC, MAE < 0.8 oC), even when temperatures exceeded the optimal range (e.g., reaching 50 oC at an ambient temperature of 30 oC under UDDS), highlighting its superior generalization and robustness due to the incorporation of physics-informed constraints. Beyond temperature prediction, the model reconstructs the temperature field of the battery by inferring spatial temperature distributions across a finite element grid, with validation against thermal imaging data. Although minor deviations were observed in high-temperature regions—likely due to grid resolution or boundary condition simplifications—the reconstructed temperature field closely aligns with experimental observations. Compared to conventional data-driven approaches, the PINN model significantly enhances both accuracy and interpretability with limited training data, providing a high-precision temperature distribution for battery management systems. This advancement contributes to improved thermal safety and performance optimization strategies, with future work focused on extending the methodology to module-level temperature field predictions.