Abstract: The mesosphere, which is located approximately 50–90 km above the Earth's surface, is a crucial part of the Earth's atmosphere. Data assimilation in the mesosphere is essential for accurately simulating and forecasting its state. However, the lack of sufficient observations results in this field being relatively underdeveloped. In this study, we conducted an intelligent optimization particle filtering algorithm to optimize the uncertainty parameters in a physics-based model, which was used to simulate the terrestrial mesosphere. This algorithm was employed to improve the accuracy of mesospheric state simulation via the injection of sparse observations. The mesospheric temperature, density, and pressure profiles, measured by the SABER (Sounding of the Atmosphere using Broadband Emission Radiometry) onboard the TIMED (Thermosphere Ionosphere Mesosphere Energetics and Dynamics) satellite, were injected into the data assimilation model. The comparison results demonstrated that the statistical error in the mesospheric temperature simulation from the data assimilation model is comparable to that from the theoretical model. However, owing to the improved accuracy in simulating individual temperature profile, the assimilation model significantly improved the accuracy of the mesospheric pressure and density estimation. Notably, our model also improved the simulation accuracy of the lower thermosphere, where none of the measurements were injected. Moreover, the results indicated that fine-tuning the uncertainty parameters in the physics-based model can contribute to maintaining the level of forecasting accuracy for the mesosphere over several days' lead time, which is essential for long-term mesospheric prediction capabilities. This study highlights the effectiveness of intelligent optimization of the uncertainty parameters in a theoretical model in improving model accuracy and extending forecast reliability within the mesosphere.