¹ Al-Bayan University, Iraq;

² College of Mechanical Engineering, University of Technology-Iraq, Iraq.

* Correspondence: luttfi.a.alhaddad@uotechnology.edu.iq

Deep space missions expose spacecraft and crew to elevated fluxes of galactic cosmic rays (GCRs), making radiation shielding a critical materials selection problem. High-fidelity radiation transport tools (e.g., OLTARIS) provide accurate shielding estimates but are computationally expensive when screening large materials sets or evaluating design trade-offs. To accelerate this process, this study explores a machine learning–based surrogate approach for predicting the whole-body effective dose equivalent of aerospace materials across a wide shielding spectrum. Dose–shielding data were obtained for 59 aerospace materials across 18 shielding thickness levels (0.01–1000 g/cm²) using OLTARIS v3.5 under GCR boundary conditions for both the MAX and FAX adult voxel phantoms. A Support Vector Regression (SVR) model with optimized hyperparameters was trained to forecast effective dose as a function of material and thickness. Model performance was assessed using MAE, RMSE, MAPE, MaxAE, CVRMSE, and R² metrics, and residual analyses were conducted to evaluate model bias and robustness. The SVR model achieved high predictive fidelity across all materials, with R² > 0.97 and MAPE < 3% for the ten representative materials analyzed. Hydrogen-rich fuel and hydride materials (e.g., Liquid Hydrogen, Lithium Hydride) exhibited the lowest errors (R² = 0.996–0.997, MAPE = 0.863–0.984%), followed by composites (R² = 0.990–0.994) and polymers (R² = 0.986–0.992), while structural metals produced modestly higher errors (R² = 0.973–0.985). Residual distributions centered near zero with median spreads below ±0.01 Sv, confirming minimal systematic over- or under-prediction bias. Comparisons between MAX and FAX phantoms indicated small anatomical effects on prediction consistency (≈5–10% variation in residual spread). Machine learning–based surrogate models can reliably approximate radiation shielding performance and significantly reduce evaluation time relative to repeated transport simulations. The findings support nonlinear regression frameworks to assist early-stage material screening and spacecraft design studies for future integration with automated materials selection workflows and multi-physics mission optimization pipelines.