Optimised Machine Learning for MASW-Based Zonation of Weak Alluvial Soils Using Anchor-Based SPT/SCPT Calibration

Weak alluvial soils are difficult to delineate where penetration-test data are sparse and stiffness varies rapidly with depth. This study presents a novel goal-attainment-optimised machine learning framework that combines multichannel analysis of surface waves (MASW) with anchor-based standard penetration test (SPT) and seismic cone penetration test (SCPT) calibration to improve the depth continuity of weak-soil zonation. A dimensionless potential weak-soil index, Pweak, was derived from SPT-N, SCPT cone resistance and MASW-derived shear-wave velocity for 265 site-depth observations from 13 Taipei Basin sites. SPT/SCPT measurements at predefined anchor depths within the 0–10, 10–20 and 20–30 m zones were condensed into a site-zone calibration coefficient and combined with representative depth and MASW-derived Vs as artificial neural network (ANN) inputs. An exhaustive architecture search evaluated 650 candidate architectures, after which the goal-attainment procedure selected a two-hidden-layer network with 16 and 22 neurons, yielding Rval=0.896, MSEval=0.019 and an overall correlation coefficient of 0.875. The ANN-derived zonation reproduced the principal weak and relatively strong intervals of the reference calculation and preserved local depth-dependent transitions more effectively than simple interpolation between sparse anchor levels. The framework therefore provides a practical means of extending anchor-based penetration-test calibration along MASW profiles for preliminary ground-risk interpretation.