A causal perspective on Machine Learning for concrete quality predictions and data-driven mixture optimization

Zenodo (CERN European Organization for Nuclear Research)

Abstract

Machine Learning (ML) predictions of cement and concrete quality and subsequent data-driven mixture optimization has been advertised for almost three decades. However, supervised ML leverages correlations, not causal relationships. Aiming for hybrid models, we derive the first causal mixture model for concrete, combining causal discovery and domain knowledge. This causal model highlights that mixture optimization is adversarial in cost, strength and workability; and omitting any of these yields trivial solutions. Challenging the literature consensus, we find that mixture design has therefore few degrees of freedom and strictly requires a causally interventional evaluation, rather than the correlation-based evaluation employed in most papers. Applying a perspective of formal causality, we identify common confounders and (spurious) correlations in different dataset types, among laboratory, literature, and production datasets. These considerations lead to specific guidelines for the experimental setup and evaluation of ML on concrete mixture data. A subsequent literature review reveals widespread malpractice and a quantitative link of apparently excellent model performance to poor evaluation practice. Finally, we demonstrate how these conceptual mistakes induce overestimation of model performance, optimal complexity, and usability by comparing the common spurious setup to our framework on two datasets. We find simple models to improve predictions of Compressive Strength in production, yet only moderately, whereas Slump is not predictable – contrary to the conclusions with the common setup. Overall, this work aims to advance data-driven mix design with ML by introducing causal reasoning to the field, raising awareness of ML limitations, and providing tools for unbiased performance evaluation.

Authors 6

1. Thorsten Kalb

   University of Padua

   Affiliation as printed

   University of Padua

   Statwolf Data Science

2. Anil Esen Aachen Institute for Building Materials Research (ibac)

   RWTH Aachen University

   Affiliation as printed

   Everox

   Institute for Building Materials Research at RWTH Aachen University

3. Elsa Qoku

   Technische Universität Braunschweig

   Affiliation as printed

   Institute of Building Materials, Concrete Construction and Fire Safety, Technical University of Braunschweig

4. Thomas Matschei

   Technische Universität Braunschweig

   Affiliation as printed

   Institute of Building Materials, Concrete Construction and Fire Safety, Technical University of Braunschweig

5. Chiara Masiero
   Affiliation as printed

   Statwolf Data Science

6. Gian Antonio Susto

   University of Padua

   Affiliation as printed

   University of Padua

   Statwolf Data Science