Rock grain size plays a significant role in the analysis of hydraulic conditions and the identification of depositional environments. Traditional methods for grain size measurement, for instance sieve analysis and laser diffraction, are time-consuming, costly, and suffer from discontinuity in depth due to limited core recovery during drilling. Although the combination of well log curves and machine learning methods can compensate for the limitations of rock physics experimental techniques, existing studies mainly focus on one-dimensional characteristic values of grain size, lacking a comprehensive representation of the two-dimensional grain size distribution. In this study, we propose a machine learning approach that combines the automatic hyperparameter optimization framework (Optuna) with gradient boosting algorithms (LightGBM and XGBoost) to address the challenge of predicting two-dimensional grain size distribution in reservoirs. Based on well log data and grain size distribution experimental data from a certain block in the Chengdao oilfield, we compare eight different machine learning methods, including linear regression, Support Vector Regression (SVR), k-Nearest Neighbors (k-NN), random forest, Gradient Boosting Decision Tree (GBDT), XGBoost, LightGBM, and Convolutional Neural Network (CNN). By optimizing the machine learning parameters, we identify the most appropriate method for predicting reservoir grain size distribution. The research results demonstrate significant differences in the accuracy of grain size distribution prediction among the ten machine learning methods. When using nine well log parameters, including natural potential, sonic, wellbore diameter, compensated neutron, natural gamma, formation resistivity, deep lateral resistivity, micro lateral resistivity, and shallow lateral resistivity, as inputs, the proposed method achieves the highest accuracy in predicting the two-dimensional grain size distribution in reservoirs, with R² coefficients approaching 0.7 and smaller errors. Furthermore, linear regression, SVR, as well as GBDT attain lower accuracy in predicting reservoir grain size distribution, which are not eligible for grain size prediction in reservoirs.