Ground settlement prediction for open caisson shafts in sand using a neural network constrained by empiricism

Abstract

Accurate prediction of soil settlements induced by open caisson construction in sand is essential for safe and reliable delivery of critical underground urban infrastructure. This paper presents a novel prescriptive design approach using a neural network (NN) constrained by empirical relationships, referred to as an ’empiricism-constrained neural network’. The proposed approach is benchmarked using a traditional closed-form empirical expression. Both methods are calibrated using experimental data from reduced-scale laboratory testing for the prediction of surface and subsurface settlement trough shape and magnitude. The outcomes demonstrate that while both methods accurately capture the primary effect of caisson depth on surface and subsurface soil settlements, the NN approach exhibits superior prediction accuracy. These methods are developed in a form amenable for routine design use in industry and have the potential for broader applicability in other design scenarios, such as building damage assessment and risk-assessment exercises.