YSNO: A Yield-Source Neural Operator for Inverse Modeling of Slope Mechanics

Frequent landslide disasters underscore the need for monitoring-driven models that can assess slope risk before catastrophic failure. A key capability underlying such models is to infer the spatially varying strength of a slope from observed deformation, which in turn requires a differentiable forward model capable of resolving both localized yielding and its nonlocal mechanical consequences. However, existing neural operators and their benchmarks have been developed predominantly for fluid dynamics and generic PDEs, leaving heterogeneous elastoplastic slope mechanics and inversion largely underexplored. We introduce the Yield-Source Neural Operator (YSNO), an inverse-oriented neural operator for spatial material-field inference in slopes, together with SlopeBench, a paired finite-element benchmark for elastoplastic field prediction and material inversion. YSNO combines a differentiable finite-element physics lift, a plasticity-gated local operator for localized nonlinear corrections, and a finite-rank global integral operator that propagates yield-related source information across the domain. The resulting fully differentiable surrogate can be frozen and directly embedded into gradient-based inversion, establishing an end-to-end computational pathway from displacement observations to spatially distributed strength estimates. On SlopeBench, YSNO reduces full-setting displacement relative-$L_2$ error by approximately 44.9% relative to the strongest evaluated baseline and yields substantially more accurate material-field reconstruction in inverse tasks. Cross-resolution and out-of-distribution experiments further demonstrate useful generalization beyond the standard training regime. By providing a benchmark for jointly evaluating forward accuracy and inverse suitability in elastoplastic slope mechanics, SlopeBench fills an underexplored regime in scientific operator learning. More broadly, YSNO provides a computational foundation for transforming deformation monitoring into interpretable estimates of internal slope strength, with the potential to support more reliable landslide risk assessment, early warning, and disaster mitigation for vulnerable communities worldwide.