AI-driven symbolic regression for preliminary material estimation in pre-stressed concrete (PSC) bridges using multi-expression programming

Early-stage estimation of material quantities in prestressed concrete (PSC) bridges is often uncertain because design information is limited, leading to reliance on assumptions and potential inaccuracies. This study proposes a predictive modeling approach to estimate steel weight (B s ) and concrete volume (V c ) in PSC bridges during the conceptual stage, aiming to generate estimates that closely align with actual project quantities. The study collected a dataset of 172 completed bridges from the Highway Department in Pakistan. Eleven input parameters and two output parameters (B s and V c ) were compiled from international sources and evaluated against each dataset. Predictive performance was compared between Multiple linear regression (MLR) and Multi-expression programming (MEP) models. The generated models were represented as explicit symbolic mathematical expressions and evaluated using MAE, RMSE, NSE, R 2 , RSE, and RRMSE. Sensitivity and parametric analyses were also performed to examine the relative influence of the input parameters on the predicted outputs. Under the adopted 70/30 train–test split, the MEP models outperformed the MLR models, achieving average R 2 values of 0.965 for B S and 0.910 for V c . Within the investigated input ranges, the developed MEP models show potential to support preliminary material estimation during the conceptual design of PSC bridges. Further validation using independent datasets and alternative resampling strategies is required before their broader application.