Precise measurements of the nuclear composition and energy spectrum of primary cosmic rays around the knee are essential to understand their origin, acceleration, and propagation. The GRAPES-3 experiment in Ooty, India, recently reported a spectral hardening in the proton spectrum at $\sim 166 \mathrm{TeV}$ using Gold’s unfolding method based on muon multiplicity distributions. To enhance composition sensitivity by incorporating additional observables, we implement a Deep Neural Network (DNN) using both muon multiplicity and shower age, along with other high-level reconstructed shower parameters. This work presents the strategy, performance, and reliability of the DNN-based approach for mass composition studies at GRAPES-3.