Uncertainty Quantification in Data-Driven Geotechnical Stratigraphic Modeling

Industry 4.0 will bring about a new Geotech digital era dominated by data science. In alignment with this trend, data-driven geotechnics is an emerging research field that contributes to this digital transformation. On the other hand, the Annex D in the latest edition of the international standard "General Principles on Reliability for Structures" (ISO2394:2015) recognizes that geotechnical reliability and risk analysis are sound only when the source uncertainty - the interpretation of observed data- is well quantified. Yet at present, it still heavily depends on engineers' subjective experience and may result in a less- or over-conservative design/decision. The challenges are two-fold: 1) how to better interpret geotechnical data that are multivariate, sparse, incomplete, and potentially corrupted in an algorithmic and smart manner at probed points, and 2) conditional on observed data, how to infer and model geotechnical information at vast unobserved locations accurately with quantified uncertainty. In this paper,the above two challenges are addressed to a certain extent using anin-house developed Bayesian approach for 2D soil stratigraphic interpretation. This new approach builds upon the author's previously developed one-dimensional hidden Markov random field (HMRF) model and 2D anisotropic Markov random field (MRF) simulation algorithm. Bayesian machine learning is implemented to jointly perform parameter estimation and stochastic simulation of soil stratigraphic profiles. The advantages of the developed approach are 1) inferring stratigraphic profile and associated uncertainty in an automatic manner and 2) both aleatoric and epistemic uncertainties are taken into consideration. This paper contributes to the techniques leveraging limitedsite investigation data for informed decision-making in geo-risk management

Keywords: stratigraphic modeling, Bayesian machine learning, uncertainty quantification, Markov random field.