A multimodal machine learning framework for assessing moisture and thermal damage in asphalt binder aggregate systems

Abstract Moisture and thermal degradation of asphalt binder and aggregate interfacial bonds is a major precursor to stripping, adhesion loss, and premature pavement deterioration, but conventional mixture level indicators often fail to capture early interface damage. This study develops a multimodal learning framework for post-conditioning diagnostic assessment of durability loss by integrating binder rheology, aggregate surface characteristics, conditioning variables, contact angle measurements, and image derived damage descriptors. A 720-record experimental dataset was assembled using eight asphalt binders, six aggregate sources, five conditioning states, and three replicates. Target outputs included pull off adhesion strength, retained adhesion index, moisture damage probability, failure mode, interfacial durability score, and critical conditioning class. Mean pull off adhesion strength decreased from 722.0 kPa to 453.1 kPa after saline thermal aging, while retained adhesion declined from 100.0% to 61.7%. Contact angle increased from 77.4° to 101.9° and debonded area ratio increased from 5.0% to 31.8%, confirming progressive wetting related and morphological damage. For post test diagnostic inference, the Multimodal Transformer and CatBoost fusion model achieved R² values of 0.984 and 0.968 for adhesion strength and retained adhesion, moisture damage classification accuracy of 0.96, and failure mode macro F1 of 0.90. The reported multimodal performance therefore characterizes already conditioned interfaces and should not be interpreted as prospective prediction before environmental exposure or mechanical testing.