Smart computing models of California bearing ratio, unconfined compressive strength, and resistance value of activated ash-modified soft clay soil with adaptive neuro-fuzzy inference system and ensemble random forest regression techniques

Sugeno or Takagi–Sugeno–Kang (TSK) type fuzzy inference system ANFIS proposed by Jang and ensemble random forest (ERF) regression, an extension of bootstrap aggregation of decision trees, has been employed to forecast the triple targets of strength properties of a hydrated-lime activated rice husk ash stabilized soft clay soil. This was necessitated to deal with the incessant failure being recorded on flexible pavements around the world and the efforts being made to tackle the situation in a more smart and sustainable approach. The independent variables of this model protocol were the HARHA—hydrated-lime-activated rice husk ash, $${w}_{L}$$ —liquid limit, $${w}_{P}$$ —plastic limit, $${I}_{P}$$ —plasticity index, $${w}_{\mathrm{OMC}}$$ —optimum moisture content, $${A}_{C}$$ —clay activity, $${\delta }_{\mathrm{max}}$$ —maximum dry density, while CBR—California bearing ratio, $${\mathrm{UCS}}_{28}$$ —unconfined compressive strength at 28 days curing and $$R$$ —resistance value were estimated and employed as the targets (dependent variables). The natural clayey expansive soil used for this research work was investigated through preliminary experiments and classified as A-7–6 group according to AASHTO. It exhibits a very high plasticity index with high clay content, hence needed modification to be rendered as a foundation material. The soil was treated with varying percentages of HARHA, and the effect on the consistency limits, compaction, CBR, UCS, and R-value was studied. These observed values gave rise to 61 datasets. The observed datasets were deployed on the learning capacity of ANFIS and ERF regression to proposed models for the targets. The outcome of the results showed that both the models presented a close correlation between the parameters used in the model execution. Evaluation of the models was performed using a variety of statistical errors, Kendall and Spearman’s rank correlations. The results of ERF regression outclasses ANFIS model yielding a 100% coefficient of determination (R) for the triple targets. The performance evaluation and validation tests show that the coefficient of determination was more than 0.94 with minimized errors. It was concluded that ERF regression and ANFIS learning techniques are viable smart approaches to forecasting engineering problems for a more sustainable design and performance evaluation.