Abstract
Despite enormous efforts over the last decades to establish the relationship between concrete proportioning and strength, a robust knowledge-based model for accurate concrete strength predictions is still lacking. As an alternative to physical or chemical-based models, data-driven machine learning (ML) methods offer a new solution to this problem. Although this approach is promising for handling the complex, non-linear, non-additive relationship between concrete mixture proportions and strength, a major limitation of ML lies in the fact that large datasets are needed for model training. This is a concern as reliable, consistent strength data is rather limited, especially for realistic industrial concretes. Here, based on the analysis of a large dataset (>10,000 observations) of measured compressive strengths from industrially-produced concretes, we compare the ability of select ML algorithms to "learn" how to reliably predict concrete strength as a function of the size of the dataset. Based on these results, we discuss the competition between how accurate a given model can eventually be (when trained on a large dataset) and how much data is actually required to train this model.
Introduction
Predicting the 28-day compressive strength of concrete is a cornerstone of civil engineering, yet it remains notoriously difficult. The relationship between mixture proportions—including water, cement, aggregates, and admixtures—and final strength is non-linear and non-additive. While traditional physics-based models struggle with this complexity, data-driven Machine Learning (ML) offers a promising alternative. However, ML faces a significant "data hunger" problem; most high-performing algorithms require massive datasets to train effectively. In the construction industry, high-quality, consistent data is often rare. This paper investigates how much data is actually "enough" to train a reliable model and which algorithms are best suited for the sparse datasets typical of industrial applications.
Methodology
The study evaluated three archetypal learning algorithms representing different families of ML:
- Polynomial Regression (PR): A traditional statistical approach using a 3rd-degree polynomial function.
- Artificial Neural Network (ANN): A network-based model with one hidden layer and seven neurons.
- Random Forest (RF): A tree-based ensemble method consisting of 16 decision trees.
The researchers utilized a dataset of 10,264 observations of industrially produced commercial concretes. Models were trained on seven influential variables, including water-to-cementitious ratio, cement content, and various admixture dosages. The data was split into a 70% training set and a 30% hidden test set. Learning curves were employed to evaluate efficiency by plotting Mean Square Error (MSE) against an increasing number of training examples.
Results and Discussion
The analysis revealed a fundamental trade-off between a model's ultimate predictive ability and its learning efficiency. Rigid models like PR and ANN are more constrained by their mathematical structures, which limits their ultimate accuracy but allows them to reach their maximum potential with very few data points.
In contrast, the flexible Random Forest model is not constrained by a fixed formula, allowing it to capture more complex patterns. As shown in the accuracy comparison above, RF achieved the highest ultimate accuracy with an $R^2$ of 0.620. However, it required significantly more data—approximately 4,070 points—to reach this peak.
The Polynomial Regression model, while slightly less accurate ($R^2 = 0.596$), reached its peak performance with roughly 35% less data (~2,680 points) than the RF model. The comparative performance metrics are summarized below:
| Model Type | Ultimate Accuracy ($R^2$) | Data Points for Peak Accuracy |
|---|---|---|
| PR | 0.596 | ~2,680 |
| ANN | 0.591 | ~3,010 |
| RF | 0.620 | ~4,070 |
Conclusion
For industrial applications where data is scarce, simpler models like Polynomial Regression may be the more pragmatic choice due to their higher learning efficiency. While Random Forests offer superior accuracy given sufficient data, their high data requirements can be a limiting factor. Future research should explore hybrid models or transfer learning to combine the accuracy of flexible models with the efficiency of simpler ones.
Summary updated