Device-aware machine learning for cohesive-soil erodibility under measurement heterogeneity

Critical shear stress ( τ c ) and erodibility coefficient ( k d ) translate hydraulic loading into cohesive-soil erosion response, but their data-driven prediction is confounded by systematic device effects when records from different erosion-testing protocols are pooled. We present a device-aware machine-learning framework, built on 715 cohesive-soil records from six erosion-testing devices, that predicts both parameters from routine geotechnical properties and states the validity of every prediction. Analysis of covariance identifies the testing device as a first-order source of variation, with the two internal-flow devices reporting τ c one to more than two orders of magnitude above the surface-loading group. These offsets must be estimated before records are pooled and cannot be reused as conversion factors between devices. Under validation that successively withholds sample groups, devices, sources, and projects, the models predict accurately within the represented domain with statistically calibrated intervals; beyond it, accuracy is lost without the device offset. Interval width serves as an explicit decision rule between using a prediction and testing directly. A new instrument requires a short anchoring test on reference soils, and applications beyond the covered domain require direct erosion testing; in both cases the missing element is a measurement that no model refinement can supply. The learned density– k d relationship reproduces a controlled compaction experiment in direction and ranking, and a laboratory dam-breach case carries the chain from index properties to breach hydrograph, matching the measured peak discharge within 7%. The framework converts heterogeneous erosion-test records into erodibility inputs with stated validity limits for preliminary scour assessment and embankment-breach screening.