Advanced Hybrid AI Models for Predicting Softening Point and Penetration of Modified Bitumen

Abstract

Ensuring bitumen quality is essential for the durability of asphalt pavements. Conventional laboratory tests for softening point (SP) and penetration (P) are reliable but labor-intensive and time-consuming, particularly for modified binders incorporating recycled additives. This study proposes an advanced hybrid machine learning (ML) framework integrating Neural Architecture Search (NAS) with Artificial Neural Networks (ANN) to predict the SP and P of modified bitumen containing crumb rubber (CR), coffee grounds (CG), and olive pomace (OP). A dataset of 49 laboratory-prepared samples was used to train and validate three models – Random Forest (RF), ANN, and NAS-ANN – evaluated through statistical indicators (MSE, RMSE, R², R). The NAS-ANN achieved the best performance, with R = 0.9665/0.9567 for SP and 0.9726/0.8711 for P (training/validation). Sensitivity analysis indicated that OP had the most significant influence on SP, followed by CR and CG. The optimized NAS-ANN model was implemented in a user-friendly interface named SmartBit-AI, designed to facilitate rapid prediction and mix-design preselection. Overall, the proposed NAS-ANN framework complements conventional laboratory testing, offering a practical and efficient tool for accelerating sustainable bitumen evaluation and design.