An AI-Driven, Mechanistically Grounded Framework for Geospatial Modelling of Soil Liquefaction

Regional-scale predictions of soil liquefaction are useful, both prior to an event for policy and planning, and immediately after for response and recovery. Such predictions could ideally be made: (1) rapidly, (2) at high resolution, and (3) over the regional scope of cities and transportation networks. Given the growth of community geotechnical data, remote sensing, and algorithmic learning, the existence of such a model is conceivable. Towards that end, interest in “geospatial” models has recently grown. In lieu of directly measuring below-ground traits, as is traditionally done by liquefaction models, geospatial models infer them from above-ground parameters. Field tests of such models have demonstrated both their provocative potential and significant room for improvement. Informed by these tests, this paper develops a geospatial modelling framework driven by algorithmic learning (benefiting from machine- and deep-learning insights, or ML/DL) but pinned to a physical framework (benefiting from mechanics and the knowledge of regression modelers). In effect, subsurface geotechnical test data are predicted remotely and then input to a well-known mechanistic framework for probabilistically predicting ground failure. This has three advantages: (1) a mechanistic underpinning; (2) a significantly larger training set, with the model trained on in situ test data, rather than on ground failures; and (3) insights from ML/DL, with greater potential for geospatial data to be exploited. While liquefaction is a physical phenomenon best predicted by mechanics, subsurface traits lack theoretical links to above-ground parameters, but surely correlate in complex, interconnected ways—a prime problem for ML/DL. To demonstrate, a preliminary model is developed from a modest dataset, is tested on recent earthquakes, and is shown to provide efficient predictions. Ultimately, this approach could be used to develop national and global geospatial models that provide improved regional-scale predictions of liquefaction.