Advances in sensing technologies allow to design safer vehicles by tailoring restraint deployment to crash and occupant conditions. Machine Learning (ML)-based controllers can enhance passive safety by tuning the restraint systems on the field to specific crash and occupant variability. Production vehicles rely on discrete calibration maps that link a limited set of crash severities and occupant classes to fixed airbag and belt strategies. As more variables are added, the number of map entries grows combinatorially and becomes difficult to calibrate and maintain. This discrete strategy cannot cope with the variability of real-world crashes and occupants, inhibiting real time adaptation to the real-world scenarios. This project develops and benchmarks an ML-based workflow for intelligent adaptive restraint design. Human Body Models (HBM) are used in finite element (FE) simulations to evaluate occupant injury across various restraint configurations and crash scenarios. A Bayesian Experiment Design (BED) strategy focuses simulations on the most informative regions of the design space. The results are employed to train Gaussian Process Regression (GPR) surrogate models. The GPR surrogate models are leveraged to collect data and train a Neural Network (NN) controller that reacts to variables like occupant characteristics, seating position, and crash speed, fine-tuning the activation of airbags with adaptive vents and belt systems with switchable load limiters. The proposed optimization framework yields deployment strategies that minimize selected injury metrics and can, in principle, be embedded into real-time decision logic. A direct comparison with a production vehicle calibration map shows that the learned controller achieves comparable occupant protection in regulatory crash scenarios, while offering greater flexibility and robustness in off-nominal conditions. The proposed workflow demonstrates how ML-based surrogate modelling, adaptive sampling, and NN controllers can be embedded in passive-safety development to accelerate calibration, broaden design-space coverage, and enable scalable restraint design.