The high-fidelity model is not easy to analyze and optimize because of the high computational cost. When the classical design of experiments is applied to construct the metamodel to replace such a computationally intensive model for analysis or optimization, it usually needs more samples compared with hybrid adaptive sampling to ensure the reliability of the metamodel due to ignoring the system information. In this study, considering the general feature of the chemical model, a new method of hybrid adaptive sampling named hybrid adaptive sampling algorithm based on scores (HASAS) is proposed on the basis of k-nearest neighbor for exploration and nearest neighbor expected improvement for exploitation to enhance the global quality of the metamodel. Furthermore, a weight coefficient is introduced to balance exploration and exploitation during sample placement. Sixteen benchmark cases are utilized to evaluate the performance of the HASAS and four other hybrid adaptive methods. Results show that with the same number of samples, HASAS can perform well in terms of global accuracy on most of them. The effect of the number of initial sample points on HASAS is also discussed. Finally, it is applied for the construction of the Kriging for the hydropurification process of a typical chemical (terephthalic acid) and sensitivity of each variable is done with the metamodel.