Allosteric modulators are important regulation elements that bind the allosteric site beyond the active site, leading to the changes in dynamic and/or thermodynamic properties of the protein. Allosteric modulators have been a considerable interest as potential drugs with high selectivity and safety. However, current experimental methods have limitations to identify allosteric sites. Therefore, molecular dynamics simulation based on empirical force field becomes an important complement of experimental methods. Moreover, the precision and efficiency of current force fields need improvement. Deep learning and reweighting methods were used to train allosteric protein-specific precise force field (named APSF). Multiple allosteric proteins were used to evaluate the performance of APSF. The results indicate that APSF can capture different types of allosteric pockets and sample multiple energy-minimum reference conformations of allosteric proteins. At the same time, the efficiency of conformation sampling for APSF is higher than that for ff14SB. These findings confirm that the newly developed force field APSF can be effectively used to identify the allosteric pocket that can be further used to screen potential allosteric drugs based on these pockets.