Predicting the Compressive Strength ofConcrete up to 28 Days-Ahead: Comparisonof 16 Machine Learning Algorithms onBenchmark Datasets

Concrete is the most important and widely consumed construction material.
Concrete parts are produced by a mixing process, followed by casting and a
certain curing time. To assess the quality of concrete, its compressive strength
is usually measured (typically after 28 days curing time). Several factors affect
the compressive strength of concrete, including environmental factors, the type,
quality, and quantity of the constituents, the order of the mixing process, and
the curing conditions. Due to the multitude of factors effecting compressive
strength and partially known chemical reactions during mixing and curing,
in this contribution, data-driven methods are used to model the behavior of
the concrete production process. Three different benchmark datasets from the
concrete manufacturing field are used for the modeling procedure. 16 typical
learning algorithms were selected based on their simplicity and their performance
in predicting compressive strength. The results show that 1) repeated
cross-validation is more reliable than repeated hold-out in this configuration, 2)
the interaction and power terms (2nd order) of the inputs have a positive effect
on model prediction, 3) the kernel type of the models is of crucial importance,
and 4) gradient boosting and kernel ridge are the most appropriate models for
predicting compressive strength.