Soil Stratification from Sparse Boreholes Based on Transition Probability–Informed Machine Learning

Reliable soil stratification from sparse borehole data poses significant challenges in geotechnical engineering. Among the limited information extractable from boreholes, transition probability between soil types serves as a crucial statistical feature. Machine learning approaches have been increasingly adopted for soil stratification due to their superior predictive capabilities, but fail to incorporate this valuable transition probability information. This study presents a novel transition probability–informed machine learning (TPIML) framework to bridge this gap. The method first generates Markov chain realizations based on the transition probabilities. A key innovation lies in the proposed mapping strategy, which distinguishes soil subtypes based on their neighboring layers in the Markov chain realizations and then maps the discrete soil types in original boreholes into continuous latent values accordingly. These latent values at the sampling locations are then extended into spatially continuous fields through interpolation methods. The final probabilistic soil stratification is generated by machine learning algorithms, which utilize both the latent field values and spatial coordinates of each soil unit as input features. The effectiveness of TPIML is validated through both synthetic and real-world case studies. Results demonstrated that by incorporating the continuous latent fields transformed from transition probabilities, the method reproduces naturally interbedded soil layer patterns in the layer boundaries, rather than the oversimplified clear boundaries typically seen in conventional methods. The probabilistic predictions, obtained through ensemble voting from multiple realizations, demonstrated superior accuracy compared with four mainstream methods while providing comprehensive uncertainty quantification in a data-driven manner.