Title:
Bayesian Machine Learning for Modeling and Design of Ultra-High-Performance Concrete
Author(s):
Kurtis
Publication:
Web Session
Volume:
ws_F23_Ogulcan-KimKurtis.pdf
Issue:
Appears on pages(s):
Keywords:
DOI:
Date:
10/29/2023
Abstract:
Designing a UHPC mix is a complex problem due to the range of material feedstocks that can be used and the rigorous requirements for performance metrics. Bayesian methods are a powerful approach to machine learning that allow for robust embedding of domain knowledge and provide accurate predictions of material properties as well as uncertainty estimates. Here, using compilations of published UHPC data, two Bayesian machine learning methods were employed to demonstrate the application of this type of modeling for cementitious materials. Materials were represented in a framework of hierarchical machine learning, and a fundamental goal in this study was to compare the accuracy and generalizability of models parameterized by compositional variables with those parameterized by latent variables based on empirical models from the literature. The data were first modeled by an ensemble ridge regression, and miscalibration area (a Bayesian error metric) indicated that models parameterized by latent variables have improved generalizability compared to those parameterized by composition. Then, Gaussian process regression based on an expanded feature set was used to develop predictions of compressive strength that, counterintuitively, were found to have higher accuracy for models parameterized by compositional variables (test R2 = 0.91) than by latent variables (test R2 = 0.77). However, the latter were found to more accurately predict the properties of materials produced with untested fine aggregate and to predict novel compositions with high compressive strength, which is consistent with the significant reduction in miscalibration error in these models. These results demonstrate that latent variables in a framework of Bayesian machine learning are a powerful tool in modeling cement and concrete, providing greater generalizability across the variable space, making robust predictions on untested feedstocks, and predicting new compositions with optimal properties.