<p>Rail collisions pose significant safety challenges, particularly due to vehicle override phenomena. Anti-climbing energy absorber systems (EASs) mitigate this risk through controlled deformation. This study develops a hybrid machine learning framework integrating an artificial neural network surrogate model with the NSGA-II genetic algorithm to optimize Anti-climbing EAS designs. The approach addresses three conflicting crashworthiness objectives: maximizing energy absorption, minimizing mass, and reducing peak crushing force. A high-fidelity finite element model was developed and validated against experimental data. Subsequently, 4000 geometric variants were systematically generated via Hammersley sampling. The hybrid ANN model, trained on eight geometric parameters, achieved exceptional predictive accuracy (<i>R</i>² &gt; 0.98 for all outputs) and was benchmarked against traditional response surface methodology, demonstrating superior performance in predicting safety-critical metrics. Global sensitivity analysis quantified the influence of all design parameters, revealing that wall thickness and tube radius are the dominant factors. Furthermore, the anti-climber profile parameters (<i>y</i>1–<i>y</i>4) were identified as significant secondary levers for performance tuning, while the horizontal parameter (×3) provides a unique control for mass reduction. Optimization results revealed clear trade-offs: the maximum EA design absorbed 156.04 kJ but incurred a 62% mass penalty. Multi-criteria decision-making analysis showed that TOPSIS and SAW provided balanced solutions, while VIKOR favored high energy absorption. FE validation confirmed the framework’s accuracy, with the TOPSIS-optimized design exhibiting exceptional precision. This research establishes a computationally efficient, data-driven design methodology that replaces trial-and-error approaches, advancing rail safety.</p>