Inverse-designed neural networks for rheological prediction of asphalt binders using SARA properties

Asphalt binders, critical for flexible pavement performance, owe their viscoelastic properties to their complex chemical composition, characterised by Saturates, Aromatics, Resins and Asphaltenes (SARA) fractions. This study introduces a bidirectional neural network (BidiNN) framework to predict and optimise the rheological behaviour of asphalt binders based on their SARA composition. The forward model maps chemical inputs, including SARA fractions and environmental conditions, to rheological outputs such as complex modulus (G∗) and phase angle (δ), achieving high accuracy (R2>0.965). The inverse model infers optimal SARA fractions to achieve target viscoelastic properties, addressing the non-linear interplay between chemical constituents and mechanical performance. Cycle-consistency losses ensure robust bidirectional predictions, while an ensemble approach captures the non-uniqueness of the inverse problem, generating multiple chemically viable SARA compositions. Chemical characterization separates the binder into SARA fractions. Asphaltenes are first precipitated and removed by filtration, after which the resulting maltene fraction is separated by Solid Phase Extraction (SPE) into saturates, aromatics, and resins.. Aging processes, monitored by Fourier‑Transform Infrared (FTIR) spectroscopy, shift SARA distributions toward higher asphaltene content, impacting binder durability. The BidiNN framework enables precise tailoring of chemical compositions for enhanced pavement performance and sustainability.