Carbon dioxide (CO2) is one of the major greenhouse gases, and changes in CO2 concentration in the air is affected by many factors. At present, there are still few stations that can continuously observe ground air CO2 concentration, but conventional meteorological data are relatively easy to obtain. There is still a lack of research on how to use conventional meteorological data to simulate ground air CO2 concentration at a specific location. Based on the measured meteorological data and CO2 concentration data from the Hejiashan small watershed in Chunhua County, Shaanxi Province, the MLP, LSTM, Bi-LSTM and RF models were driven by ground air CO2 concentration or the combination of environmental factors at different locations, respectively. The potential of machine learning methods to simulate daily-scale air CO2 concentration on the ground of typical point in the gully region of the Loess Plateau was evaluated. The results show that the four machine learning methods can simulate the overall change process of ground air CO2 concentration, whether using the same type of observed data as input or combination of environmental factors as input, and can be used for data interpolating. When the measured CO2 concentration at other points is used as input, the simulation accuracy of the CO2 concentration at C2 by MLP is the highest, and the error in the test set is only 3.8 %. When the environmental factors are used as input, the simulation accuracy of the CO2 concentration at C2 by RF is the highest, and the error in the test set is 6.3 %. However, when environmental factors are used as input, it is impossible to simulate the dramatic daily variation of CO2 concentration. For data interpolation, it is better to use the same type of observed data as input, which can better simulate the daily process of ground air CO2 concentration.