A Comparative Assessment of Concrete Compressive Strength Prediction at Industry Scale Using Embedding-Based Neural Networks, Transformers, and Traditional Machine Learning Approaches

Abstract

Concrete is the most widely used construction material globally, yet its compressive strength remains notoriously difficult to predict due to inherent material heterogeneity, variable mix designs, and sensitivity to site?specific and environmental factors. Recent advances in artificial intelligence offer promising data?driven paradigms for automating quality control decisions in construction. In this study, we systematically evaluate five predictive modeling approaches—linear regression, decision trees, random forests, transformer?based networks, and embedding?based neural networks—using a large?scale industry dataset comprising nearly 70,000 compressive strength test records. The models integrate critical mix?design and placement parameters, including water–cement ratio, cementitious content, slump, air content, ambient temperature, and field placement conditions. Our experimental results show that the embedding?based neural network consistently outperforms both traditional machine learning methods and transformer architectures, achieving a mean absolute prediction error of approximately 2.5% for 28?day compressive strength. This accuracy approaches the intrinsic variability of standard laboratory testing, underscoring the practical viability of embedding?based learning frameworks for automated, real?time quality assurance and decision support in large?scale construction projects.