Abstract:
Geospatial models for predicting soil liquefaction infer subsurface traits via satellite remote sensing and mapped information, rather than directly measure them with subsurface tests. Field tests of such models have demon- strated both promising potential and severe shortcomings. Informed by these tests, this paper develops geospatial models that are driven by algorithmic learning but pinned to a physical framework, thereby benefiting both from machine and deep learning, or ML/DL, and the knowledge of liquefaction mechanics developed over the last 50 years. With this approach, subsurface cone penetration test (CPT) measurements are predicted remotely within the framing of a popular CPT model for predicting ground failure. This has three potential advantages: (i) a mechanistic underpinning; (ii) a significantly larger training set, with the model principally trained on in-situ test data, rather than on ground failures; and (iii) insights from ML/DL, with greater potential for geospatial data to be exploited. While liquefaction is a phenomenon best predicted by mechanics, subsurface traits lack theoretical links to above-ground parameters, but correlate in complex, interconnected ways - a prime problem for ML/DL. Preliminary models are trained using ML/DL and a modest U.S. dataset of CPTs to predict liquefaction-potential index values via 12 geospatial variables. The models are tested on recent earthquakes and are shown – to a statistically significant degree – to perform as well as, or better than, the current leading geospatial model. The models are coded in free, simple-to-use Windows software. The only input is a ground-motion raster, downloadable minutes after an earthquake or available for countless future scenarios. Ultimately, the proposed approach and models, which warrant further application and evaluation, could be improved upon using additional training data and new predictor variables. Users of the models should understand key limitations, asdiscussed in detail herein.