The nonuniform freezing of the sample has a considerable influence on the result of the freezing point test. It is associated with the size of samples and called the “size effect.” A novel apparatus based on solid-state cooling technology was developed to solve this problem. A linear freezing mode called the “ramped freezing” (RF) was proposed to replace the “constant freezing” (CF) and provide a uniform freezing condition for the soil sample. The single neuron network was used to dynamically optimize the parameters of the Proportional-Integral-Derivative algorithm and improve the performance of variable temperature control. Freezing tests of water solutions were designed to verify the applicability of this apparatus. Key parameters, such as the freezing point (Tf) and the supercooling point (Ts), could be reflected in the cooling curves. Then, freezing point tests of soils with different freezing modes were carried out. Under the CF condition, the soil sample partly freezes from the surface to the central point. Because the temperature monitored by the sensor is the average value of the sensor-soil interface, the “apparent freezing point” obtained from the sample with CF mode is higher than the actual freezing point. For the sample with RF mode, the latent heat cannot raise the sample temperature to the equilibrium freezing point if the pore water is “hypercooled.” The freezing point will be underestimated in this condition. Two measures, applying an external disturbance on the test box and turning up the set temperature after the release of latent heat, can be adopted to avoid the influence of hypercooling.