From Prediction to Causal Inference: Multi-Model Double Machine Learning for Life-Cycle Carbon Optimization of RAP Mixtures

Efficient recycling of reclaimed asphalt pavement (RAP) is essential to the low-carbon transformation of transportation infrastructure. Traditional carbon accounting methods based on four factors—RAP content, aging level, rejuvenator dosage, and mixing temperature—struggle to capture emission patterns under multi-parameter coupling or isolate confounders to identify true causal effects. The study considers a single-lane asphalt pavement functional unit (1 km length, 3.75 m width, and 0.18 m thickness), and to address this problem, this paper integrates life cycle assessment (LCA), a thermodynamic mechanism model, multi-model machine learning, and double/debiased machine learning (DML) to construct a carbon emission prediction and causal analysis framework. First, reclaimed asphalt pavement is used as the LCA model to generate 10,206 carbon emissions and cost datasets covering the four factors, with embedded nonlinear thermodynamic calculations. Then, 3000 semi-synthetic plant-day observations are constructed by introducing observable engineering conditions, batch variation, and Gaussian white noise. Finally, four prediction models—BP neural network, random forest, XGBoost, and SVR-RBF—were compared using grouped five-fold cross-validation. The BP model achieved the highest predictive accuracy (R2 = 0.9638) and was selected as the auxiliary learner for cross-fitted DML. The estimated model-based intervention effects for a 10-percentage-point increase in RAP content, a 0.1 increase in aging coefficient, a 1-percentage-point increase in rejuvenator dosage, and a 10 °C increase in mixing temperature were 12.492, 7.077, 13.189, and 8.632 t CO2-eq, respectively. These results show that material circularity and carbon reduction do not automatically guarantee economic sustainability. The framework links resource circularity, causal process understanding, and low-carbon process optimization to support sustainability-oriented RAP decisions. This study offers a reference for low-carbon design and process optimization of asphalt pavement recycling.