Evaluating the modulus of the existing asphalt concrete (AC) layer is a critical procedure in Mechanical-Empirical (ME) rehabilitation analysis. Generally, the modulus could be back-calculated by the Falling Weight Deflectometer (FWD) test. However, the raw FWD data of each pavement section is not always readily prepared for local highway agencies. To address this issue, the main objective of this study is to establish a reliable model by machine learning (ML) methods to predict AC layer modulus for the existing flexible pavement with data readily available from the local pavement management system, which could be an auxiliary tool for network-level sections with no FWD tests. The long-term pavement performance (LTPP) database was used to collect the original data for model training and testing, including pavement structures, service age, climate records, and pavement distresses. After preliminary data processing, matrix correlation analysis, and feature selection, the prepared dataset (total data points = 6477) with 14 predictors was fed into three regression models, including Ordinary Least-Squares regression (OLS), Random Forest regression (RF), and Gradient Boosting regression Method (GBM). The related key hyperparameters were optimized by grid search and 5-folds cross-validation. By comparison, the GBM model was finally selected due to its considerably higher prediction accuracy (R2 = 0.7921) than RF model (R2 = 0.7525) and OLS model (R2 = 0.4371) in the test set. According to the variable importance given by GBM model, surface temperature and AC layer thickness are more dominant variables in modulus estimation. In addition, a case study with predicted AC layer moduli in ME rehabilitation design was provided to verify the model application. In summary, the trained GBM model can be utilized to predict AC layer modulus for pavement evaluation and then ME rehabilitation when FWD data is not available.