In the present study, conducted with the aim of modeling groundwater quality, a method for calculating the spatial structure of data in the modeling process is proposed. This method considers the distance between observation points and the estimation point as one of the inputs to the model. The GBR, GPR, KNN, MLP, SVM, and RF models were utilized, and the models were trained and tested using groundwater quality data obtained from the Qazvin province in northwestern Iran. Specifically, data sets included 3,058 wells for 4 nearby observation wells, 2,724 wells for 5 nearby observation wells, 2,080 wells for 6 nearby observation wells, 1,364 wells for 7 nearby observation wells, and finally 631 wells for 8 nearby observation wells. Two separate data sets, comprising information from the first six months and the second six months of the year, were used. The average error (MAE), R-squared, Pearson correlation coefficient, and RMSE for models with four, five, six, seven, and eight neighboring wells indicated satisfactory performance of the Random Forest model. This model demonstrated very good performance in both the training and testing phases, exhibiting the lowest error and highest correlation. Although the complexity and execution time of this model may be higher, its high accuracy compensates for these drawbacks.