Development of predictive model for mild steel in Benin city using random forest

Abstract

Buried steel infrastructure in Nigeria suffers premature failure due to soil-induced corrosion, yet existing degradation models rely on idealized laboratory simulations that ignore real-world soil heterogeneity and welding parameter interactions. This study addresses this critical gap by investigating mechanical property decay and corrosion behavior of mild steel weldments through longitudinal field exposure in Benin with the aim of developing robust, field-validated predictive models to enhance infrastructure durability. Field exposure tests were conducted over twelve months on 20 weldment specimens, with tensile strength, impact energy, and hardness evaluated post-exposure alongside weight loss-based corrosion rates. Random Forest (RF) was trained on welding parameters (current, voltage, gas flow) using 16 samples, with rigorous validation on 4 unseen field runs. Hyperparameter optimization, residual analysis, and error distribution diagnostics ensured model robustness assessment under actual environmental stressors. Random Forest demonstrated exceptional predictive capability across all sites and responses, achieving test-set R² values of 0.798 (tensile) 0.752 (impact), and 0.814  (hardness), with corrosion rate predictions exceeding R² = 0.65 even in aggressive soils.