Abstract
Reliable identification of pavement structural and material properties from surface deflection measurements remains challenging, even as modern monitoring technologies provide increasingly dense data. In practice, data-driven surrogate models are widely adopted to enable efficient inverse identification, yet high forward prediction accuracy does not necessarily translate into stable or physically meaningful inverse solutions. This study investigates why surrogate-based inverse identification can remain unreliable and how data characteristics, model structure, and physics-guided constraints jointly influence inverse performance. Using a representative pavement monitoring scenario, we develop an integrated diagnostic framework that examines inverse behavior across three levels: the information structure of the data, the forward mapping characteristics of surrogate models, and the resulting inverse optimization performance. Conceptual analysis and numerical experiments reveal that uneven parameter sensitivity and limited observability in the data impose intrinsic constraints on inverse identifiability that cannot be resolved by surrogate modeling alone. Furthermore, surrogate models with excellent forward accuracy may still induce unfavorable inverse landscapes, leading to boundary-dominated solutions, instability, or non-unique parameter estimates. Under the tested constraint design and model families, monotonicity-guided constraints modify inverse behavior in an architecture-dependent manner, with beneficial effects observed mainly when coupled with the conditionally parameterized surrogate. Overall, the results demonstrate that inverse reliability depends not only on surrogate accuracy, but fundamentally on how data information structure and surrogate parameterization jointly shape the geometry and conditioning of the inverse problem.