Clogging risk mapping of cemented paste pipeline transport using a knowledge guided transfer learning framework

Pipeline blockage is a critical reliability constraint in cemented paste backfill (CPB) transport. Existing machine learning studies mainly predict rheology, pressure drop, deposition velocity, or blockage diagnosis from relatively consistent input sets, but CPB clogging risk mapping often has to use sparse records, heterogeneous labels, and missing operating variables. This study proposes a knowledge guided transfer learning framework for screening transport failure related to clogging. The framework converts available slump and yield stress measurements into a unified flowability index (UFI), combines UFI with solids mass concentration to build an interpretable chart prior, and represents unreported inlet velocity and pipe diameter through latent process priors and Monte Carlo marginalization. Bootstrap samples with jitter provide a pseudo source dataset for pretraining, followed by fine tuning and evaluation on a compiled dataset of 52 records with six positives under the main label definition. The evidence is evaluated through repeated splits containing both classes, LOPO case level analysis, label sensitivity definitions, prior learning ablations, paired baseline comparisons, UFI uncertainty propagation, and process prior sensitivity tests. The resulting risk map integrates sparse field records with explicit prior information and shows plausible increases in screening score at higher solids concentration and poorer flowability, which provides a transparent tool for design screening, diagnostic scenario comparison, and selection of cases requiring closer hydraulic assessment.