Engineering properties of peat: a data-driven approach

Peat is a highly variable organic soil that presents major challenges for geotechnical engineering. In Finland, where peatlands cover one-third of the land area, infrastructure often intersects deep deposits that are difficult to characterise and costly to improve. Conventional practice relies on conservative assumptions or large-scale replacement, which can be expensive and carbon intensive. In this study, we compiled a harmonised database of over 250 datapoints from Finnish, Nordic, and Western European sources, focusing on compressibility, yield stress, and undrained shear strength. Regression analyses were benchmarked against Random Forest machine learning models. The results confirmed that the compression index is well predicted from water content, whereas yield stress and undrained shear strength display high variability under regression. Random Forest models provided modest improvements over conventional regression for strength-related parameters, while compressibility remained well captured by empirical correlations. Cross-parameter estimators offer additional tools where direct strength testing is unavailable. The findings underline the complementary roles of regression and machine learning in peat characterisation. Moreover, a hybrid workflow is proposed: regressions for early screening and conservative design, and machine learning for refined, site-specific assessment, supporting more sustainable infrastructure development on peatlands.