Prior-guided online estimation of peak proximity from partial shear-curve geometry in multi-stage direct shear tests

Reliable peak-oriented stage termination remains a key challenge in multi-stage direct shear testing, particularly in automated workflows where stopping decisions must be made online from the evolving response. Fixed checkpoints such as AK3 and AK4 provide useful operational references, but their relation to the true peak condition varies significantly across soil types and states. This paper proposes a two-layer framework for supervisory peak-oriented monitoring, evaluated on a dataset of 297 laboratory direct shear tests compiled from multiple experimental series. A descriptor-based prior model provides an initial preference between AK3 and AK4 based on pre-shear material and state variables. During shearing, a geometry-based estimator predicts peak proximity from the evolving partial shear-stress–displacement curve. The prior model indicates that fines content dominates checkpoint preference at the dataset level, while within mixed-grained soils a combined representation including median grain size d 50 improves consistency. For online estimation, geometry-based inputs already yield strong predictive performance ( R 2 ≈ 0 . 90 ), while a combined formulation integrating checkpoint-state and geometric descriptors achieves R 2 ≈ 0 . 93 . Temporal analysis demonstrates that reliable prediction is possible before the full AK3 checkpoint is reached, with R 2 ≈ 0 . 91 at 60% of the AK3 displacement. The results show that fixed checkpoints alone do not provide a consistent mechanical reference for peak proximity. Instead, combining checkpoint-state information with partial-curve geometry enables robust online estimation. The predicted mobilization ratio further allows accurate reconstruction of peak shear stress from pre-peak data, reaching R 2 = 0 . 969 , a mean absolute percentage error of approximately 6.0%, and an empirical 90% prediction-interval coverage of 89.8%.