Benchmark examples for data-driven site characterisation

TLDR

Geotechnical engineering decisions depend on site‑specific projects, making data‑driven site characterization a central focus in data‑centric geotechnics. The paper introduces eight benchmark examples and a procedure to evaluate machine‑learning methods for data‑driven site characterization, aiming to produce 3D subsurface maps and property estimates from site data and Big Indirect Data. The study uses a 20 m × 20 m × 10 m full‑scale ground model, generating virtual grounds with varied soil layers and synthetic CPT data, and defines benchmark examples from both virtual and real CPT datasets to assess DDSC methods.

Abstract

Decision making in geotechnical engineering is always related to a project carried out at a specific site. It is natural for data-driven site characterization (DDSC) to attract the most attention in data-centric geotechnics. This paper proposed eight benchmark examples and a benchmarking procedure to support unbiased and competitive evaluation of emerging ML methods. The primary goal of DDSC is to bring the value of a "data first" agenda to practice, specifically to produce a 3D stratigraphic map of the subsurface volume below a full-scale project site and to estimate relevant engineering properties at each spatial point based on site investigation data and other relevant Big Indirect Data (BID). A reasonable full-scale ground 20 m long × 20 m wide × 10 m deep is adopted. Virtual grounds containing horizontal, inclined, or discontinuous soil layers and spatially varying synthetic cone penetration test data are created to test the performance of DDSC methods over a range of ground conditions. A benchmark example is defined by a combination of a virtual ground ("reality") and a training dataset (measured "reality"). An additional benchmark example based on actual CPT data is included to check whether performance under virtual ground conditions holds under real ground conditions.