Probabilistic machine-learning approach for asphalt binder formulation: case study on antioxidant dosage optimization

A key challenge in asphalt binder modification is determining the optimal dosages of additives, commonly known as asphalt binder formulation. Complex interactions between the base binder and additives make theoretical models difficult to establish, leaving current practice heavily reliant on trial and error. This study proposes a methodology to improve efficiency and reduce subjectivity in asphalt binder formulation through a feedback loop coupling probabilistic machine learning with experimental validation. Unlike conventional machine learning approaches that typically require large datasets, the proposed probabilistic approach leverages uncertainty quantification to learn from limited experimental data. The proposed approach adopts a Bayesian optimization framework, with a Gaussian process model predicting material properties and associated uncertainties. Based on these probabilistic predictions, trial designs are selected, experimentally evaluated, and incorporated into the dataset to iteratively refine the model and guide subsequent trials toward the optimal formulation. The predictive uncertainty enables the balance between exploration and exploitation during trial design, thereby improving material design efficiency. The methodology is applied to optimize the dosage of an antioxidant additive, ethylenediaminetetraacetic acid. The case study demonstrates improved binder formulation efficiency and reduced subjectivity. The proposed methodology contributes to the pursuit of intelligent material design and improved pavement material performance.