Ice and water coexist in frozen soil, and their respective contents (ice content, theta i; unfrozen water content, theta u) are critical factors influencing the mechanical properties of frozen soil. Currently, these two parameters are measured separately. Existing measurement methods require specialized equipment, are time-consuming. To improve measurement efficiency, this paper proposes an inverse analysis surrogate model, which can simultaneously predict both theta i and theta u within one minute. The method process is as follows: 1. A three-dimensional numerical model is established to simulate the transient heat conduction in frozen soil under heat pulse. 2. Six parameters (theta i, theta u, rho s, lambda s, Cs, Gs) need to be determined for each simulation. Through Monte Carlo sampling of six parameters, thousands of numerical simulations are performed. Then, a dataset comprising thermal response curves (TRC) labeled with (theta i, theta u, rho s, lambda s, Cs, Gs) is established. 3. A machine learning algorithm is used, where TRC and soil property parameters serve as inputs, and (theta i, theta u) as outputs. 4. In the laboratory, soil property parameters are measured, and in the field, TRC within one minute of frozen soil is measured in real-time. By inputting soil property parameters and TRC into the machine learning model, (theta i, theta u) can be obtained in real-time.The method was validated through laboratory experiments. The results show that with TRC and rho s, lambda s, Cs as inputs, mean absolute errors (MAE) for theta i and theta u were 2.3 % and 3.1 %, respectively. The proposed method significantly improves measurement efficiency, allowing for the simultaneous measurement of theta i and theta u within one minute.