Discussion: A comparative study on the application of artificial intelligence networks versus regression analysis for the prediction of clay plasticity [Arab J Geosci (2021) 14(7), 534]

This article presents a discussion of the paper by Akbay Arama et al. (2021) published in the Arabian Journal of Geosciences, Vol. 14, No. 7, Article 534. Using simple and multiple regression analyses, as well as soft computing artificial neural network (ANN) methods, Akbay Arama et al. (2021) investigated a large database comprising of 1253 Atterberg limits test results for high and very high plasticity clay (CH) soils that are dominant on the European side of Istanbul Province, Turkey. Stated objectives included querying the attainment of the plasticity index (PI) for these soils directly from only liquid limit (LL) test results. Based on regression analyses results, Akbay Arama et al. (2021) produced various correlations relating the PI to LL. Then, employing the Noe (2005) experimental PI–LL dataset, Akbay Arama et al. (2021) reported that their proposed correlations have superior predictive performance compared to those of eight previously reported PI–LL correlations selected (by them). The authors of the present discussion article are of the views that (i) the produced PI–LL correlations for the Istanbul CH soils may not be sufficiently reliable, (ii) it does not seem intuitive to compare the predictive capabilities of these proposed correlations against those previously reported in the literature and deduced for diverse groups of fine-grained soils sourced from geographically different regions, and (iii) since PI–LL correlations cannot definitively and consistently establish the water content corresponding to the experimental plastic–brittle transition point, the thread-rolling test for plastic limit (PL) determination (and hence by association for PI calculation) needs to be retained. Each of these views is elaborated in this article.