Relative suitability of surrogate-modeling paradigms under input-limited conditions

Determine the relative suitability of deep surrogate modeling paradigms for reconstructing knee joint contact mechanics under practically relevant input-limited conditions in which inputs such as joint posture and joint reaction forces may be corrupted or partially unavailable.

Background

Finite element analysis is the gold standard for studying knee joint contact mechanics but is computationally expensive. Deep surrogate models promise rapid prediction of internal stress fields, yet most prior evaluations emphasize ideal, complete inputs.

Practical deployments often involve degraded or missing inputs (e.g., posture measurement errors, uncertain loads, or absent load channels). The paper frames an open question about which modeling paradigms are most appropriate under such input-limited conditions and proceeds to compare five representative paradigms—local diffusion (MeshGraphNet), history-context enhancement, hierarchical multi-scale propagation, explicit global interaction, and a local-global hybrid model—across several input-limited scenarios.