In this study, the multidirectional behavior of secondary neutron fields generated by the interaction of 50 MeV protons with a copper target was comprehensively investigated in a multilayer Fe-concrete shielding geometry using a hybrid framework combining Monte Carlo (MC) simulations and machine learning (ML). The proton accelerator tunnel was modeled as a three-dimensional structure consisting of symmetric 2 m Fe + 4 m concrete side shields, a concrete roof, a concrete floor, and the surrounding soil environment. FLUKA MC simulations were performed to analyze secondary particle production, neutron energy spectra, and spatial ambient dose equivalent H*(10) distributions. Neutron energy spectra were evaluated using the USRTRACK scoring estimator, providing insight into the energy-dependent transport, moderation, and attenuation mechanisms within the multilayer shielding system. The results demonstrated that the iron layer effectively suppresses the fast-neutron component through inelastic scattering, whereas the surrounding concrete layers provide efficient neutron moderation and absorption, producing a progressive spectral softening together with significant reductions in H*(10). Using the FLUKA-generated H*(10) datasets, in which the one-dimensional spatial position along each axis (x or y) was used as the single input feature and the corresponding H*(10) value as the output target, Linear Regression (LR), Random Forest (RF), XGBoost, Support Vector Regression (SVR), and LightGBM models were developed and compared. Following log(1+y) transformation, min-max normalization, an 80%/20% train-test split, 5-fold cross-validation, and GridSearchCV-based hyperparameter optimization, the Random Forest model achieved the highest predictive accuracy (X-axis test R² = 0.9994; Y-axis test R² = 0.9991), while XGBoost and LightGBM exhibited comparable performance. Computational efficiency analysis showed that high-statistics FLUKA simulations require several hours on the TRUBA/TR-Grid HPC infrastructure, whereas the trained ML models generate predictions within milliseconds, corresponding to computational speedups of approximately 10⁶–10⁹ during the prediction stage. The combined analyses of neutron spectra, multidirectional dose distributions, and surrogate modeling demonstrate that the proposed hybrid MC-ML framework provides both a physically interpretable and computationally efficient methodology for rapid radiation shielding optimization and accelerator shielding design.