Using Artificial Intelligence Techniques to Predict Intrinsic Compressibility Characteristic of Clay

Abstract

Reconstituted clays have often provided the basis for the interpretation and modelling of the properties of natural clays. The term “intrinsic” was introduced to describe a clay remoulded or reconstituted at moisture content up to 1.5 times its liquid limit and consolidated one-dimensionally. In order to circumvent the difficulties of measuring an intrinsic constant called “intrinsic compressibility index” (C*c), a machine learning (ML) approach using traditional non-parametric tree-based and meta-heuristic ensembles was adopted in this study. Results indicated that tree-ensembles namely random decision forest (RDF) and boosted decision tree (BDT) performed better in C*c prediction (average R2 of 0.84 and root mean square error, RMSE of 0.51) compared to stand-alone models. However, models’ hyper parameters combined meta-heuristically, produced the highest accuracy (average R2 of 0.90 and root mean square error, RMSE of 0.34). The greatest capacity to distinguish between positive and negative soil classes (average accuracy of 0.95, precision and recall of 0.86) were demonstrated by meta-ensembles in multinomial classification.