Formation drillability characteristics are critical for optimizing drilling performance and designing bottom-hole assemblies. Accurate estimation of drillability parameters, including uniaxial compressive strength (σc), drillability grade (Kd), and abrasiveness index (Ia), is fundamental for safe and efficient drilling operations. To overcome the generalization constraints of conventional single models and their inability to capture multi-scale feature responses, this study proposes a novel data-driven framework, termed the multi-scale convolutional neural representation with stacking ensemble (MSCNR-SE), for accurately estimating σc, Kd, and Ia. The framework uses a three-stage strategy: (1) a multi-scale convolutional module extracts features across different spatial scales; (2) an ensemble of diverse base regressors (e.g., random forest (RF), gradient boosted decision tree (GBDT)) enhances representational capacity; and (3) extreme gradient boosting (XGBoost) is employed as a meta-learner to integrate the outputs of base regressors, thereby mitigating the risk of model-specific bias. This hierarchical deep-learning ensemble effectively captures complex nonlinear relationships intrinsic to formation properties. Experimental results show that MSCNR-SE accurately models depth-aligned drillability profiles, with strong agreement between predictions and measurements (the coefficient of determination, R2 > 0.93 on the blind testing sequence). The MSCNR-SE model achieved superior predictive performance compared with other models across all three parameters. Specifically, MSCNR-SE reduced the mean absolute error (MAE) by approximately 3%–50%, the root mean squared error (RMSE) by 4%–64%, the mean absolute percentage error (MAPE) by 3%–59%, and increased the R2 by 1%–32%. These results highlight the superior predictive accuracy and robustness of MSCNR-SE, underscoring its potential as a practical, highly efficient data-driven solution for real-time parameter estimation, monitoring, and control in complex drilling.