FEATURE TOKENIZER TRANSFORMER BASED ADVANCED DEEP LEARNING METHOD FOR PREDICTION OF THE BEARING CAPACITY OF BRANCHED PILES ON PEAT

Abstract

Peat is regarded as one of the most complex category of soil, in terms of construction, owing to its very soft and highly compressible nature. Thus, in order to overcome shortcomings of peat, deep foundations are most commonly adopted to transfer load through soft deposits to harder strata. However, such techniques may be uneconomical, particularly for low to medium-rise structures, pavements, embankments, and other lightly loaded critical infrastructures. This shortcoming is of particular concern especially in peat deposits, where the highly compressible peat can induce substantial deformation even under relatively light structural loads. Hence, there is a need to develop alternative ground improvement techniques that are cost-effective, practically feasible, and capable of mitigating excessive deformation associated with peat while addressing limitations of conventional stabilization methods. In this regard, the present study evaluates the bearing capacity of peat reinforced with branched piles. Large-scale experimental studies were conducted to investigate the effectiveness of the proposed technique under controlled loading conditions, while machine learning (ML) techniques were simultaneously incorporated to develop predictive models for estimating bearing capacity and to assess the influence of key soil and pile parameters on the overall foundation response. To predict the bearing capacity of branched piles (BPs) and fine-tune the prediction outcomes, four ML models, namely the Feature Tokenizer Transformer (FTT), Bayesian Regularized Feed Forward Neural Network (BRFFNN), Generalized Regression Neural Network (GRNN), and Multi-Layer Perceptron Regressor (MLPR), were employed. The predictive performance of the four models was comparatively evaluated to identify the most accurate and reliable approach for estimating bearing capacity of branched piles (BPs). Model training was performed using a dataset comprising of 570 experimental observations, each representing five key bearing-related attributes of BPs subjected to static vertical loading. Among the four models, FTT demonstrated the highest predictive accuracy, outperforming the other approaches, as evidenced by the statistical evaluation with a coefficient of determination (R²) value of 0.9940, approaching unity. The high predictive efficiency of the FTT model was further substantiated through rank and sensitivity analyses, which provided insights into the relative importance and influence of the input parameters on BP bearing capacity.