Explainability analysis of a machine learning-based constitutive model for concrete

Resumo

Concrete, like other quasi-brittle materials, exhibits intricate mechanical behavior that defies simple mathe-
matical description and remains partially elusive. Limited comprehension of the governing mechanisms hampers

the development of a comprehensive theory and broader constitutive models. This limitation suggests that more

encompassing models could emerge through advanced techniques, like Machine Learning (ML) algorithms, capa-
ble of capturing material behaviors effectively. While various studies endorse ML-based constitutive models and

validate the proposed hypothesis, skepticism persists within academia due to concerns about ML models being
“black boxes” devoid of interpretable physical consistency. To challenge this perception, this paper introduces the
SHAP tool, employed to validate the physical coherence of an ML-based constitutive model focused on concrete
representation. The SHAP tool’s methodology is outlined, accompanied by illustrative applications showcasing
the correlation between input variables and model outputs. Clear demonstrations of physical consistency debunk

the notion of ML models as opaque “black boxes.” Ultimately, this study debunks skepticism, offering new per-
spectives on the utilization of ML-based constitutive models, thus fostering broader acceptance and integration in

the structural engineering community.