Prediction of the Bond Work Index (BWI) and Hardgrove Grindability Index (HGI) from routinely measured rock mechanical and index parameters is critical for energy-efficient comminution design, yet remains challenging due to complex, nonlinear relationships among strength, abrasivity and energy consumption. This study proposes a deep neural network (DNN)-based machine-learning framework to estimate BWI and HGI using mechanical and index test results obtained from 24 coal surrounding rock samples (sandstone and siltstone) from the Zonguldak Basin. The input space is constructed from seven routinely performed laboratory tests describing rock strength, drillability, abrasivity and hardness. Correlation analysis and Variance Inflation Factor (VIF)-based screening are employed to identify and remove redundant or weakly contributing parameters, leading to more compact DNN architectures. The resulting models achieve high predictive performance, with R2 values of about 0.8-0.86 for BWI and up to about 0.95-0.96 for HGI on the available dataset and additional stratified train-test splits and cross-validation analyses confirm the potential of the approach, particularly for HGI prediction. Overall, the proposed methodology demonstrates that DNNs can capture multi-parameter nonlinear interactions more effectively than traditional empirical formulations and offers a practical tool for rapid grindability assessment, mining energy optimization and the development of data-driven rock characterization workflows.