Abstract: Thermal stratification hinders natural circulation and affects passive heat removal. In addition, unstable thermal stratification causes low-frequency temperature oscillations, which not only reduce the stability of the neutron and thermal-hydraulic systems but may also damage the reactor vessel and internal components. Therefore, accurately studying thermal stratification is crucial for the safety analysis of liquid metal reactors. A promising approach for modeling thermal stratification is the multiscale method that couples CFD with system codes. However, this approach must address the challenges of large spatiotemporal scale differences and high computational resource consumption in CFD simulations. To improve the efficiency and accuracy of transient analysis for reactor thermal stratification, a transient data-driven coarse-grid turbulence model was developed. Taking the flow of sodium in the SUPERCAVNA facility as the computational object, the standard k-ε model was used on the CFD platform OpenFOAM to generate fine-grid flow field data. The kNN algorithm was then applied to convert the fine-grid data into coarse-grid data, creating a training database. Using the machine learning platform TensorFlow, a hybrid architecture combining a self-attention mechanism and a bidirectional long short-term memory (BiLSTM) network was introduced. The self-attention mechanism extracted dependencies between different elements in the input sequence, effectively capturing the distribution of physical quantities. BiLSTM processed time-series data through bidirectional information flow, allowing the model to utilize both past and future time-step information, thus enhancing its ability to learn complex temporal patterns. Based on this architecture, a transient surrogate model was built to predict turbulent eddy viscosity at the next time step using local flow field features from the current time step under coarse-grid settings. The surrogate model was embedded into the OpenFOAM solver to develop a data-driven turbulence model. The computational accuracy, generalization ability, and acceleration capability of the model were evaluated by simulating the transient flow of sodium in the SUPERCAVNA facility. The results show that in the machine learning model, the mean squared error (MSE) on the validation set is 3.38×10⁻⁹, the R² is 0.990 511, and the symmetric mean absolute percentage error (SMAPE) is 5.258 4%. This indicates that the model achieves parameterized modeling from local flow field features to turbulent eddy viscosity. The transient data-driven coarse-grid turbulence model has high accuracy in capturing the thermal stratification effect. Outside the Reynolds number range covered by the training dataset, the data-driven model reproduces the thermal mixing phenomenon, showing a certain level of generalization ability. The computational time of the transient data-driven coarse-grid turbulence model is 1/13 of that of the fine-grid CFD simulation, indicating high computational efficiency.