Geotechnical Properties Inference Using MWD Data – A Probabilistic Machine Learning Approach-Paper

Paper

Pdf Summary

The study presents a probabilistic method for inferring subsurface properties from microtunneling measurements, addressing the uncertainty and noise that complicate interpretation of measurement-while-drilling data. It models observed jacking forces using a deterministic force model and estimates four soil parameters: the face-resistance factor, residual friction angle, pipe–soil interface friction coefficient, and soil unit weight.<br /><br />The Bayesian framework uses borehole-based prior information and the No-U-Turn Sampler (NUTS) to update parameter estimates as drilling data accumulate, producing posterior distributions and credible intervals rather than single estimates. Before inference, the data are cleaned using Savitzky–Golay filtering, STL decomposition, and Q–Q analysis to reduce noise and identify outliers. The method is evaluated on unlubricated sections of the Alvarado Trunk Sewer and Morris Avenue Tunnel projects and compared with genetic algorithm (GA) and particle swarm optimization (PSO) methods.<br /><br />All three approaches predict measured jacking forces reasonably well. However, Bayesian inference produces more stable and interpretable parameter estimates, with uncertainty generally narrowing as more observations are incorporated. In contrast, GA and PSO estimates fluctuate more and can behave unstably. The Bayesian method also captures changing conditions along the tunnel alignment, supporting continuous subsurface characterization.<br /><br />A key limitation is that bentonite lubrication can affect inferred properties: estimates from lubricated sections may represent the combined behavior of soil and lubricant rather than the native ground. The authors recommend explicitly modeling lubrication effects in future work. Overall, the results suggest Bayesian inverse modeling can improve real-time geotechnical assessment and decision-making in trenchless construction.

Keywords

Bayesian inference

microtunneling

subsurface properties

jacking forces

No-U-Turn Sampler

posterior distributions

soil parameters

noise filtering

genetic algorithm

particle swarm optimization