It is well recognized that the model predictability is more or less hampered by the imperfect representations of atmospheric state and model physics. Therefore, it is a common problem for any numerical models to exhibit some sorts of biases in the prediction. In this study, the emphasis is focused on the cold bias of surface temperature forecast in Naval Research Laboratory's three-dimensional mesoscale model, COAMPS (Coupled Ocean/Atmosphere Mesoscale Prediction System). Based on the comparison with the ground station data, there were two types of ground temperature cold biases identified in LLNL (Lawrence Livermore National Laboratory) operational forecasts of COAMPS over the California and Nevada regions during the 1999 winter and the 2000 spring. The first type of cold bias appears at high elevation regions covered by snow, and its magnitude can be as large as 30 - 40 *F lower than observed. The second type of cold bias mainly exists in the snow-free clear-sky regions, where the surface temperature is above the freezing point, and its a magnitude can be up to 5 - 10 *F lower than observed. These cold biases can affect the low-level stratification, and even the diurnal variation of winds at the mountain regions, and therefore impact the atmospheric dispersion forecast. The main objective of this study is to explore the causes of such cold bias, and to further the improvement of the forecast performance in COAMPS. A modified version of COAMPS with improved model physics is used in this study. The modifications related to the cold bias of surface temperature forecast include the radiation transfer calculation, the albedo and the heat capacity of snow surface, the change of nudging coefficient in the soil heat flux, and a subjective correction of climotological value of snow depth. A California winter precipitation case, occurring on January 21, 2000, is used for this study. A series of experiments are performed to gauge the sensitivity of the model forecast due to the physics changes and large-scale data with different horizontal and vertical resolutions. These simulations act to determine an optimum setting of physics options for the improvement of COAMPS surface temperature forecast. Results also suggest additional uncertainty in the data quality of initial conditions for the high elevation regions, where the observations are very limited.