Application of Machine Learning Techniques for Strength Properties of Self-Compacting High-Performance Concrete

Abstract Strength verification of concrete still depends on casting and crushing specimens at 7 and 28 days, which is slow and resource-intensive, especially for self-compacting concrete (SCC) containing several supplementary materials. This paper reports an experimental study of high-performance SCC (HPSCC) of grades M60 and M65 in which cement was partly replaced by ground granulated blast-furnace slag (GGBS) at 0%, 7.5% and 15% by mass. Fresh properties (slump flow, L-box, V-funnel) and hardened properties (compressive, flexural and split tensile strength at 7 and 28 days) were measured and compiled into a structured dataset. Raising the GGBS content from 0% to 15% increased slump flow by 19.8–28.2%, shortened V-funnel time by 3 s and increased 28-day compressive strength by 6.4–8.3%, while 7-day compressive strength fell slightly (1.6–2.1%). The highest 28-day compressive (75.27 MPa) and flexural (8.79 MPa) strengths were recorded for M65 with 15% GGBS. The dataset is analysed statistically, and a machine-learning framework (ANN, SVM, decision tree and random forest) for predicting strength from mixed proportions is set out as the next stage of the work. Keywords: self-compacting concrete; machine learning; compressive strength; tensile strength; GGBS; strength prediction 1. INTRODUCTION 1.1 Background Concrete is the most widely used construction material in the world. Its appeal lies in low cost, local availability of constituents, durability and high compressive strength, together with the ability to be placed as a fluid and then hardened into a load-bearing solid through cement hydration. Reactive powder concrete, with strengths reported between 200 and 800 MPa, shows how far the material has been pushed in high-performance structures. Self-compacting concrete (SCC) extends this versatility further: it consolidates under its own weight, fills congested reinforcement and complex formwork without vibration, and reduces labour and noise on site [1]. Its fresh behaviour, however, depends on a delicate balance of paste volume, water-to-binder ratio, superplasticizer and fine materials, and its hardened strength is sensitive to binder type and curing [2]. 1.2 Limitations of conventional quality control Strength is routinely verified on cubes or cylinders taken from the same batch as the site concrete, cured under controlled conditions and tested at 7 and 28 days. The 7-day result gives an early indication of calcium silicate hydrate (C–S–H) development, and the 28-day result remains the reference for design acceptance. The drawback is time: decisions on formwork removal, stressing or loading must often be made before the 28-day result exists. The problem is compounded by modern mixes. Fly ash, silica fume, GGBS, metakaolin and chemical admixtures each modify hydration and rheology in different ways [3,4], so empirical equations calibrated on plain cement–water–aggregate mixes lose reliability. Industrial by-products are also being adopted to cut carbon emissions and waste [5], but they make strength gain less predictable and more dependent on curing. Trial-and-error proportioning of SCC is correspondingly costly in time, materials and labour [6,7]. 1.3 Data-driven prediction Predictive models estimate concrete properties from many variables at once, including water-to-binder ratio, binder composition, aggregate content, admixture dosage, curing and age. They can capture non-linear interactions that simple closed-form relations assume. The consequences of poor prediction are practical: overestimated strength can lead to premature formwork removal and a safety risk, while underestimated strength causes unnecessary delay and cost. The aim of such models is to support laboratory testing with earlier, better-informed decisions, not to replace it. Artificial neural networks (ANNs) have been applied to estimate the strength of concrete and structural members [8], and ensemble and hybrid machine-learning models have been reported for self-compacting concrete containing industrial by-products [9,10]. Genetic programming can return explicit equations rather than a black box, fuzzy logic can represent uncertain site conditions, and hybrid or evolutionary networks that tune ANN weights and architecture with genetic algorithms reduce training effort. Once trained on historical mix designs and their test results, such models can give an estimate of 28-day strength at the time of mixing. 1.4 Objectives The objectives of this work are: (i) to evaluate the effect of GGBS replacement (0%, 7.5%, 15%) on the fresh and hardened properties of M60 and M65 HPSCC; (ii) to assemble the measurements into a structured dataset and analyse it statistically; and (iii) to define a machine-learning framework for predicting HPSCC properties from mix variables. 2. METHODOLOGY The study combines laboratory experiments with a machine-learning workflow: data generation, pre-processing, model development, evaluation, validation and benchmarking against conventional prediction methods. In the present paper the first stage is complete and its dataset (six mixes) is analysed in Section 4; the model-training stages described below define the planned prediction framework, for which a larger dataset is required. 2.1 Experimental database HPSCC mixes were produced by systematically varying the supplementary cementitious material content, aggregate proportions, water and superplasticizer dosage, using materials that conform to the relevant standards so that batches are consistent and repeatable. Two groups of properties were recorded: fresh-state behaviour (flowability, passing ability and segregation resistance) and hardened-state behaviour (compressive, flexural and split tensile strength at 7 and 28 days). 2.2 Input and output variables The inputs are cement content, SCM dosage, fine and coarse aggregate content, water-to-binder ratio (w/b), superplasticizer dosage, curing condition and age at testing. These govern hydration kinetics, fresh workability, microstructure development and load-bearing capacity. The outputs are slump-flow spread, V-funnel time and L-box ratio (fresh state), and compressive, splitting tensile and flexural strength (hardened state). 2.3 Data pre-processing Before modelling, the data are checked for missing values, duplicates, outliers and entry errors; outliers are treated with robust statistical methods. Variables are normalised or standardised so that features with large numerical ranges do not dominate the loss function. The data are then partitioned into a training set (model calibration), a validation set (hyper-parameter tuning) and an independent test set (final generalisation check). 2.4 Machine-learning models Artificial neural network (ANN): approximates highly non-linear, multi-variable relationships, such as the interaction between binder, water and admixture in cementitious matrices. Support vector machine (SVM): builds regression functions in a high-dimensional feature space through kernel functions and generalises well on small non-linear datasets. Decision tree (DT): produces transparent rule-based models by recursively splitting the data at key input thresholds. Random forest (RF): averages many decision trees, which lowers variance and overfitting and improves stability. 2.5 Training, evaluation and validation Hyper-parameters are tuned by grid search, random search or genetic algorithms, with k-fold cross-validation to limit overfitting. Models are compared using the coefficient of determination (R2), root-mean-square error (RMSE) and mean absolute error (MAE); a better model has R2 closer to 1 and lower RMSE and MAE. Predictions are validated against independent laboratory results using correlation plots and statistical tests. Sensitivity analysis and feature-importance ranking are used to reveal the influence of each constituent and to make the models interpretable. 2.6 Comparison with conventional methods The best models are benchmarked against empirical formulas and code-based regression equations in terms of prediction error, computational cost and scalability, and flexibility across mix compositions. This comparison also covers the prediction of compressive and flexural strength development with binder type, water content, aggregate grading and curing age. 3. MATERIALS AND MIX DESIGN 3.1 Materials Cement. Ordinary Portland cement of 53 grade conforming to IS 12269 was used throughout. Its specific gravity was 3.15, initial setting time 30 minutes, final setting time 10 hours and normal consistency 26–33%, all within the limits of the Indian Standard. A single batch was used for the whole programme to avoid variation between mixes. Fine aggregate. Sand from a nearby source, free of clay lumps, organic matter and silt, was used (specific gravity 2.65) after testing to the relevant Indian Standard procedures. It fills the voids between coarse particles and improves cohesion and workability, which is particularly important in SCC. Coarse aggregate. Clean, crushed angular aggregate from a local quarry was used in two fractions (20 mm and 12.5 mm; Table 3), with a specific gravity of 2.72 and water absorption within permissible limits. The smaller sizes support flow through narrow gaps between reinforcing bars, and the grading was suitable for HPSCC. GGBS. Ground granulated blast-furnace slag is obtained by rapidly quenching molten iron-making slag in water to form granules, which are then ground. Its chemical composition and physical properties are given in Tables 1 and 2 (Fig. 1). Table 1. Chemical composition of GGBS Material CaO (%) Al₂O₃ (%) Fe₂O₃ (%) SiO₂ (%) GGBS 37.30 14.12 1.01 37.68 Table 2. Physical properties of GGBS No. Test Result Requirement (IS 16714:2018) 1 Specific gravity 2.81 Not specified 2 Fineness (m2/kg) 361 Minimum 320 3 Slag activi