Data Cleansing & Overfitting Check for Interpretable ML in Concrete Design – a punching shear paradigm

Abstract

Evaluation of shear capacity is crucial for existing concrete buildings and bridges. Shear failure can cause an abrupt and brittle loss of load‐carrying capacity, leading to load redistribution and progressive collapse. An internationally research and scietific discourse is actively ongoing on the best prediction model for shear, torsion, and punching shear capacity, with no clear consensus. A machine learning‐based strength model is suggested for accurate prediction of punching shear capacity, using Gaussian Process Regression and Support Vector Regression models. Data elaboration and thorough evaluation are needed to enhance the model's performance and interpretability. The paper discusses critical methodological steps based on the authors' experience of developing ML‐models for structural concrete design against punching shear failure using 544 full‐scale tests.