Self-compacting concrete with fly ash and silica fume: experimental evaluation, microstructural analysis, and machine learning modeling

Abstract

This study comprehensively evaluates the fresh, mechanical, durability, and microstructural performance of self-compacting concrete (SCC) incorporating fly ash (FA) and silica fume (SF) as supplementary cementitious materials (SCMs). Cement was partially replaced with FA at 20%, 30%, and 40% and SF at 5%, 7.5%, and 10%, forming both binary and ternary combinations. The optimum FA content was established at 30%, beyond which early strength decreased due to dilution effects. Ternary blends with 30% FA + 5–10% SF were further developed to examine synergistic effects. Fresh property tests (slump flow, T₅₀₀, V-funnel, and L-box) confirmed all SCC mixes satisfied EFNARC criteria, with ternary blends exhibiting the best balance between flowability (720 mm slump flow), viscosity (T₅₀₀ = 3.5 s), and passing ability (H₂/H₁ ≥ 0.9). Mechanical performance improved consistently with SCM incorporation; at 180 days, the FA30SF7.5 mix attained 68 MPa compressive strength, 6.3 MPa split tensile, and 8.9 MPa flexural strength, surpassing the control by 17–22%. Durability tests demonstrated marked improvement, sorptivity reduced by 34%, rapid chloride penetration test (RCPT) charge fell from 3600 C (Moderate) to 800 C (Very Low), and ultrasonic pulse velocity (UPV) exceeded 4.75 km/s, confirming a highly dense matrix. Microstructural analysis revealed a compact C–S–H gel network and diminished portlandite content in ternary mixes, evidencing improved hydration and pozzolanic activity. Machine learning (ML) models (KNN, SVM, DT, and RF) successfully predicted compressive strength, with the Random Forest model achieving R² of 0.97. The combined experimental–ML approach demonstrates that SCC with 30% FA and 7.5% SF offers optimal performance, coupling sustainability with superior mechanical and durability characteristics.

Introduction

SCC is a highly flowable concrete that can spread and consolidate under its own weight without vibration^1–6. Since its development in Japan in the late 1980s, SCC has become a preferred material in densely reinforced and complex structural elements due to its superior filling ability, uniformity, and improved surface finish^2,4,7–10. Beyond workability benefits, SCC enables faster construction and reduced noise pollution on site, making it suitable for modern sustainable construction practices^{3,5,8,11–13}. Despite these advantages, the cement industry’s environmental impact remains a central challenge. Ordinary Portland cement (OPC) production is responsible for approximately 5–8% of global CO₂ emissions, arising from fossil-fuel combustion and limestone calcination, and also releases NOₓ, SO₂, and fine particulates that degrade air quality^14–19. Decarbonization strategies therefore prioritize clinker substitution, energy efficiency, and carbon capture, with SCMs offering an immediately deployable pathway to emission reduction^{17,18,20–24}. SCC, owing to its high binder content and particle-packing design, provides an ideal matrix for SCM incorporation, allowing fine mineral additions to enhance rheology, reduce porosity, and extend service life while lowering embodied carbon^9,21–26.

Among the wide range of SCMs FA, SF, ground-granulated blast-furnace slag, metakaolin, and limestone powder, FA and SF are the most effective and widely available^27–31. Their use not only offsets cement demand but also valorizes industrial by-products, supporting circular-economy principles^28–31. FA, a pozzolanic aluminosilicate residue from coal combustion, contributes spherical particles that act as micro-ball bearings, improving flowability and reducing water and superplasticizer (SP) demand^32–35. However, its slower pozzolanic reactivity delays early-age strength; the benefit appears at later ages through secondary C–S–H and C–A–S–H formation, which densifies the microstructure and enhances durability^33–38. SF, an ultrafine amorphous SiO₂ by-product of silicon or ferrosilicon production, provides a contrasting effect, raising early and later-age strength by filling micro-voids and reacting rapidly with portlandite, but at the expense of higher viscosity and SP demand^39–44. The complementarity of FA and SF makes their combined use particularly attractive: FA restores flow and moderates SF’s stickiness, while SF offsets FA’s early-strength deficiency^{2,4,6,41–46}.

The defining attribute of SCC is its ability to flow and consolidate without vibration. The EFNARC guidelines identify four key fresh-state parameters slump flow, T₅₀₀ time, V-funnel flow time, and L-box ratio, to evaluate filling ability, viscosity, and passing ability^47–49. FA typically increases slump flow (650–750 mm) and reduces T₅₀₀ (2–5 s) due to its spherical shape, whereas SF slightly lowers flowability and increases T₅₀₀ by thickening the paste^50–52. Balanced binary or ternary systems with optimized SP dosages can meet EFNARC limits (V-funnel 6–12 s; L-box ≥ 0.8) while maintaining segregation resistance^49–52. Proper control of water-to-binder (W/B) ratio, powder content, and SP chemistry is therefore essential for achieving the self-compacting balance between flowability and stability^46,47,50,52.

Extensive research confirms that FA and SF influence the mechanical performance of SCC in complementary ways. FA tends to reduce early-age compressive strength but provides substantial gains beyond 56–90 days as pozzolanic reactions progress^{34–38,53–55}. FA improves flowability through its spherical particle morphology, reducing internal friction and facilitating better passing ability, while also contributing to long-term strength through delayed pozzolanic reactions. Optimum replacement levels between 20 and 35% FA are commonly reported to balance workability and strength. SF, by contrast, enhances both early and long-term strength through its high surface area and reactivity, yielding up to 10–20% higher 28-day strength than plain OPC mixtures^{39–41,44,56}. SF, on the other hand, is highly effective at improving cohesiveness, reducing segregation, and refining the pore structure due to its ultrafine particle size and high amorphous silica content. Typical optimum dosages range from 5 to 10%, beyond which excessive viscosity may hinder flowability. Previous studies, including the work on SCC containing SCMs with air-entraining admixtures, have demonstrated that SF can significantly influence both fresh and hardened properties through its interaction with chemical admixtures, highlighting its established role in optimizing SCC performance. When combined, binary or ternary blends (≈ 30% FA + 7.5–10% SF) often outperform single-SCM systems, achieving superior compressive, split-tensile, and flexural strengths across curing ages^{2,6,12,46,57–60}. The synergy arises from dual-scale mechanisms: SF refines the interfacial transition zone (ITZ) early in hydration, while FA sustains gel formation at later stages, together generating a dense and resilient microstructure^{33,37,41,57,59,61}.

Beyond strength, SCC’s durability largely determines its service life. The inclusion of FA and SF consistently reduces water absorption and sorptivity, indicating finer pore connectivity^62–65. Likewise, rapid-chloride-permeability tests (RCPT) reveal marked reductions in total charge passed, often exceeding 50% relative to control mixes, shifting the ASTM classification from “moderate” to “low” or “very low” permeability^64–67. Ultrasonic pulse velocity (UPV) values typically increase with SCM dosage and curing age, signifying a denser and more uniform matrix^68–70. Extended assessments such as freeze-thaw, sulfate, and carbonation resistance also show improvement for ternary FA–SF SCCs due to refined pore structure and reduced calcium-hydroxide content^{64,66,68,71–73}. Thus, durability gains complement mechanical strength enhancements, reinforcing the sustainability rationale for blended SCC. Microscopic and mineralogical investigations corroborate these macroscopic improvements. Scanning electron microscopy (SEM) shows a transition from porous, CH-rich matrices in control concretes to compact, gel-dominated textures with indistinct ITZs in FA–SF systems^74–77. X-ray diffraction (XRD) analyses reveal reduced portlandite peaks and broader amorphous humps, consistent with extensive C–S–H formation^74,75,78. Thermogravimetric (TGA) data confirm decreased CH content and higher bound-water fractions, while mercury-intrusion porosimetry (MIP) demonstrates a shift toward finer pore sizes^75–78. Collectively, these results establish that the dual action of FA and SF, pozzolanic reaction and micro-filling—produces a continuous, refined matrix that restricts ionic ingress and enhances durability^76–80.

Despite decades of progress, notable gaps remain. Comprehensive ternary FA–SF studies under consistent curing regimes and extended exposure durations (> 180 days) are limited^{64,66,68,71,72}. Many prior works report individual parameters, either mechanical or durability, without cross-correlating them through microstructural evidence or statistical modeling^3,6,46,54,60. Moreover, the integration of machine-learning (ML) methods with experimental SCC datasets remains nascent: most published models focus only on compressive strength or chloride permeability, neglecting multi-property or mix-optimization tasks^81–83. There is also a need to connect ML-derived feature importance with physical mechanisms, thereby ensuring interpretability and reliability in predictive design^48–83. Lastly, economic and practical evaluations of SCM sourcing, variability, and standardization require further study to enable industrial scalability^{20,22,23,25,27,73,84–87}. Recent advances in materials informatics demonstrate that ML algorithms, such as k-nearest neighbors (KNN), support vector machines (SVM), decision trees (DT), and random forest (RF), can accurately model nonlinear relationships between mix proportions, curing age, and compressive strength^{81–83,88–90}. RF and DT approaches, in particular, yield interpretable variable-importance rankings that identify the dominant factors influencing SCC performance^48,88. ML thus provides a complementary computational pathway to accelerate mix design, minimize experimental iterations, and promote data-driven sustainability optimization^{31,82–89,91}.

Although many studies have examined the influence of individual SCMs on SCC, several important gaps remain. Most available research focuses on binary systems, reports limited long-term durability data, or lacks microstructural evidence to support mechanical trends. Moreover, the potential of integrating machine learning (ML) models with experimental datasets to enhance interpretability and support mix-design optimization has not been systematically explored for FA–SF blended SCC. There is therefore a need for a comprehensive experimental–computational framework that simultaneously evaluates fresh, mechanical, durability, and microstructural behavior while developing predictive tools for strength estimation. Accordingly, the specific objectives of this study is to develop SCC mixtures incorporating FA (20–40%) and SF (5–10%) in binary and ternary combinations using a constant W/B ratio, evaluating fresh-state properties (slump flow, T₅₀₀, V-funnel, L-box) and determine compliance with EFNARC guidelines, assessing mechanical performance (compressive, split-tensile, flexural strength) at 28, 90, and 180 days, investigating durability characteristics using sorptivity, RCPT, and UPV tests and identify the most resistant mix, examining microstructural development (SEM and XRD) and correlate it with mechanical and durability outcomes and developing ML models (KNN, SVM, DT, RF) for predicting compressive strength and to analyze feature importance for mix-design optimization. These objectives collectively aim to deliver a unified experimental–analytical methodology for identifying an optimal FA–SF ternary mixture and advancing data-driven SCC design.

Materials and methods

Materials

Cement

OPC of 53 grade, conforming to IS 12,269–2013, was used as the primary binder. The cement had a specific gravity of 3.15 and met the requirements of initial/final setting time and compressive strength as specified in IS codes.

Fly ash (FA)

Class F FA, obtained from Thermal Power Plant Station, Vijayawada, Andhra Pradesh, was used as a partial replacement of cement. The ash had a specific gravity of 2.20, with high silica and alumina content and low calcium oxide, making it suitable as a pozzolanic material.

Silica fume (SF)

SF, used in this study was procured from Dundi Vinayaka Traders, Vijayawada, India. The material was supplied in densified form, which is the commercially available grade with improved handling characteristics and a bulk density of approximately 550–650 kg/m³. The SF had a specific gravity of 2.10 and consisted of more than 90% amorphous SiO₂, with an average particle size < 1 μm. The densified form was dry-blended with cement prior to mixing to ensure uniform dispersion and effective microfiller action in SCC. The chemical constituent of cement, FA and SF was tabulated in Table 1; Fig. 1 represents the particle size distribution.

Table 1.

Chemical constituents of cement, FA & SF.

Constituent	Cement weight (%)	FA weight (%)	SF wight (%)
Calcium oxide, CaO, (%)	65.0	2.30	1.10
Silica, SiO2, (%)	20.0	55.59	91.10
Ferric oxide, Fe2O3 (%)	2.30	9.50	1.22
Alumina, Al2O3 (%)	4.90	26.64	1.30
Sulphuric anhydride, SO3 (%)	2.30	0.44	0.20
Magnesium oxide, MgO (%)	3.10	0.60	0.40
Potassium oxide, K2O (%)	0.40	0.40	0.31
Sodium oxide, Na2O (%)	0.20	0.23	–
Loss of ignition, (%)	1.80	4.30	4.40

Fig. 1.

Particle size distribution of cement, FA and SF.

Aggregates

Crushed granite was used as coarse aggregate with a maximum size of 20 mm and 12 mm, combined in a 60:40 ratio with a specific gravity of 2.7. Natural river sand conforming to Zone II as per IS 383:2016, was used as fine aggregate with a specific gravity and fineness modulus as 2.6 and 2.7.

Water

Potable water, free from harmful salts and impurities, was used for mixing and curing.

Chemical admixture

A polycarboxylate ether (PCE)-based high-range water-reducing admixture (SP) was used to achieve the desired workability in accordance with IS 9103:1999.

Mix proportions

The SCC mixtures were proportioned in accordance with EFNARC (2005). A constant W/B of 0.35 was adopted for all mixes. A polycarboxylate ether–based SP (PCE) was used at a fixed dosage of 1.2% of total binder; with a total binder content of 450 kg/m³, this corresponds to 5.4 kg/m³ of SP. Mix identifiers and replacement levels are summarized, and the corresponding kg/m³ quantities (cement, FA, SF, water, SP, and aggregates) are provided in Table 2. The replacement levels of FA (20–40% by mass of binder) and SF (5–10%) were selected based on ranges and optima reported in the literature for SCC. Previous studies have shown that FA in the range of approximately 20–35% offers an effective balance between improved flowability (due to the spherical morphology of FA) and later-age strength development, while SF dosages of about 5–10% typically enhance cohesiveness, early strength and durability through microfiller and pozzolanic effects^{37,43,56,59,74}. Additional works exploring rheology and strength effects of SF and combined SCMs further justify the selected SF window and the choice to evaluate intermediate dosage (7.5%) to identify an optimum ternary proportion^52,61. Preliminary trial mixes were also performed to ensure all selected combinations satisfied EFNARC fresh-property limits prior to full-scale testing. To ensure consistency across all mixtures, the water-to-binder ratio and SP dosage were strictly controlled using standard laboratory procedures. A constant W/B ratio of 0.40 was maintained for all mixes, irrespective of the replacement levels of FA and SF. The total binder content was adjusted so that the absolute mass of water added per batch remained unchanged. All batching was performed using precision digital scales (accuracy ± 0.1 g) to minimize weighing errors. SP dosage was controlled on a percentage-by-mass-of-binder basis. A polycarboxylate ether (PCE)-based SP was added at a fixed initial dosage (e.g., 0.8–1.2% of binder), and fine adjustments (≤ 0.05%) were made only when needed to achieve the target EFNARC slump-flow range (650–750 mm). SP addition was performed gradually while monitoring flowability to avoid overdosing. All mixtures were prepared at constant room temperature (27 ± 2 °C) using the same mixing sequence and timing to eliminate temperature-induced variations in SP performance. Additionally, the moisture content of aggregates was checked daily. Aggregates were brought to saturated surface dry (SSD) condition, and water adjustments were made to prevent unintended changes in effective W/B ratio. This ensured that the water contribution from aggregates did not influence the mix rheology.

Table 2.

Mix proportions of SCC with FA and SF.

Mix ID	Cement (kg/m³)	Fly ash (kg/m³)	Silica fume (kg/m³)	Water (kg/m³)	W/B	SP (kg/m³)	Fine aggregate (kg/m³)	Coarse aggregate (20 mm) (kg/m³)	Coarse aggregate (12 mm) (kg/m³)
Control	450.00	0.00	0.00	157.50	0.35	5.40	929.29	514.68	343.12
FA20	360.00	90.00	0.00	157.50	0.35	5.40	929.29	514.68	343.12
FA30	315.00	135.00	0.00	157.50	0.35	5.40	929.29	514.68	343.12
FA40	270.00	180.00	0.00	157.50	0.35	5.40	929.29	514.68	343.12
SF5	427.50	0.00	22.50	157.50	0.35	5.40	929.29	514.68	343.12
SF7.5	416.25	0.00	33.75	157.50	0.35	5.40	929.29	514.68	343.12
SF10	405.00	0.00	45.00	157.50	0.35	5.40	929.29	514.68	343.12
FA30SF5	292.50	135.00	22.50	157.50	0.35	5.40	929.29	514.68	343.12
FA30SF7.5	281.25	135.00	33.75	157.50	0.35	5.40	929.29	514.68	343.12
FA30SF10	270.00	135.00	45.00	157.50	0.35	5.40	929.29	514.68	343.12

Test methods

Fresh properties

All fresh-state evaluations followed EFNARC (2005)^92,93 and were completed within 5–10 min of mixing on a level, non-absorbent surface under controlled conditions (23 ± 2 °C; 50 ± 10% RH). Apparatus were lightly moistened and wiped dry prior to use, as per SCC practice and no external vibration was applied. Test setup for fresh properties can be seen in Fig. 2.

Fig. 2.

Test setup (A) slump-flow and T₅₀₀ (B) V-funnel (C) L-box.

Slump-flow and T₅₀₀

Filling ability and plastic viscosity were assessed using the standard Abrams cone placed on a base plate marked with 500 mm and 750 mm circles. The cone was filled in a single untamped lift, struck off flush, and lifted vertically in 2 ± 1 s to initiate flow. The slump-flow diameter was taken as the average of two perpendicular measurements after flow stabilization; T₅₀₀ was the elapsed time to reach the 500 mm circle. Target classes followed EFNARC guidance (SF1: 550–650 mm; SF2: 660–750 mm; SF3: 760–850 mm), with typical SCC viscosity corresponding to T₅₀₀ ≈ 2–5 s (lower values indicating segregation risk; higher values indicating an overly cohesive mix).

V-funnel

Viscosity and segregation resistance were evaluated with the V-funnel. With the gate closed and seal verified, the moistened funnel was filled in one operation (no compaction) and struck off level. Upon opening the gate, efflux time was recorded until the continuous stream broke at the outlet. Where specified, a repeat after an undisturbed 5-min rest (V-funnel @5 min) was used to gauge thixotropy; the difference between the two readings (Δt) indicates structural build-up. For SCC, an efflux time of ~ 8–12 s is desirable, and Δt ≤ 3 s indicates good stability.

L-box

Passing ability was measured using an L-box comprising a vertical and horizontal leg separated by a rebar gate (three 10–12 mm bars at 41 mm spacing). The vertical leg was filled without vibration; on lifting the gate, concrete flowed into the horizontal leg. After rest, heights were recorded before (H₁) and beyond (H₂) the rebar, and the blocking ratio (H₂/H₁) was used as the acceptance measure. Values ≥ 0.80 indicate satisfactory passing ability and ≥ 0.90 are preferred for congested reinforcement.

Mechanical properties

All mechanical tests were performed to IS 516:2018 (ASTM C39, ASTM C496, and ASTM C78)^92,93. Specimens were cast in oiled moulds, covered to limit evaporation, demoulded after 24 ± 2 h, and water-cured at 27 ± 2 °C to the ages of 28, 90, and 180 days. Figure 3 shows the test setup of mechanical properties.

Fig. 3.

Test setup (A) compression (B) split tensile (C) flexural.

Compressive strength

It was determined on 150 × 150 × 150 mm cubes. Before testing, cubes were surface-dried and the bearing faces checked for planeness and perpendicularity; minor irregularities were removed by light grinding to ensure uniform load transfer. Each specimen was centered between the platens of a calibrated 2000 kN compression machine and loaded monotonically under the code-prescribed stress-rate control until failure.

where f_c is the compressive strength (MPa), P is the maximum applied load (N), and A is the cross-sectional area of the specimen (mm²).

Split tensile strength

It was evaluated on 150 mm diameter × 300 mm cylinders. After curing, cylinders were wiped surface-dry and positioned horizontally between the platens with thin, uniform plywood strips placed along the generatrices to reduce local stress concentrations. Load was applied diametrically at the rate specified by the standard until a through-diameter crack formed, indicating tensile failure. The splitting tensile strength was computed from the recorded peak load and specimen geometry.

Where f_t is the split tensile strength (MPa), P is the applied load at failure (N), L is the length of the specimen (mm), and D is the diameter (mm).

Flexural strength

It was measured on 100 × 100 × 500 mm prisms under third-point loading over a 400 mm span. Prisms were removed from the curing tank, surface-dried, and seated carefully on support rollers to ensure alignment and avoid torsion. Load was applied through two loading rollers located at the third points at a constant, code-compliant stress rate until fracture. Where the crack initiated within the middle third of the span, the standard modulus-of-rupture expression was used; if outside the middle third, the alternative section-modulus approach specified by the code was applied and recorded.

Where f_r is the modulus of rupture (MPa), P is the maximum applied load (N), L is the span length (mm), b is the width of the specimen (mm), and d is the depth (mm).

Durability properties

All durability tests were conducted at 28, 90, and 180 days following the relevant standards: ASTM C1585^93,94 for sorptivity, ASTM C1202^93,94 for rapid chloride penetration (RCPT), and IS 13,311 (Part 1), ASTM C597^93,94 for ultrasonic pulse velocity (UPV). Test setup for fresh properties can be seen in Fig. 4.

Fig. 4.

Test setup (A) sorptivity (B) rapid chloride penetration (C) ultrasonic pulse velocity.

Sorptivity

Capillary absorption was determined on 100 mm-diameter, 50 mm-thick discs sawn from cured cylinders. Sorptivity specimens were conditioned using oven drying at 105 ± 5 °C for 24 h, followed by cooling to room temperature inside a sealed desiccator to prevent uncontrolled moisture uptake. The curved surface and top face were sealed to enforce one-dimensional ingress, and the bottom face was exposed to a 3 ± 1 mm water head on supports to avoid head pressure artifacts. Mass gain was recorded at prescribed intervals up to 6 h (initial regime) and periodically up to 7 days (secondary regime). Absorption–time data were used to derive initial and secondary sorptivity from the linear portions of the respective regimes.

Rapid chloride penetration test

Chloride-ion penetrability was assessed on 100 mm-diameter, 50 mm-thick discs prepared from companion cylinders. Prior to testing, discs were subjected to vacuum saturation at 95 kPa for 3 h, followed by 1 h of saturation under atmospheric pressure. The degree of saturation measured immediately before testing was ≥ 98% for all specimens. Each specimen was mounted between cells containing 0.3 N NaOH (anode) and 3.0% NaCl (cathode) solutions, and a 60 V DC potential was applied for 6 h. Current was logged at 30 min intervals and the total charge passed (coulombs) obtained by trapezoidal integration. Bath temperature was monitored to ensure compliance with the standard; any specimen exhibiting excessive heating or unstable current traces was flagged. The charge passed was used to classify chloride permeability per ASTM categories (high, moderate, low, very low, negligible).

Ultrasonic pulse velocity

Concrete quality and homogeneity were evaluated using ~ 54 kHz P-wave transducers in direct transmission wherever geometry permitted (opposite faces); semi-direct/indirect modes were used only when necessary and are identified in the results. Specimen faces were cleaned, a couplant was applied to ensure good acoustic contact, and the path length was measured with a vernier for accuracy. For each path, at least three transit-time readings were taken and averaged, rotating the probe orientation to check anisotropy. Pulse velocity was calculated as the path length divided by the average transit time and interpreted with the quality bands given in IS 13,311 (excellent, good, doubtful, poor).

Machine learning approach

To complement the experimental program, supervised machine ML models were developed to predict compressive strength of SCC mixes from their mix-design and age descriptors. The dataset comprised the measured compressive strengths at 28, 90, and 180 days for all mixes tested. The input feature set included curing age (days), FA replacement (%), SF replacement (%), and the W/B; the target variable was compressive strength (MPa). Prior to modelling, records were screened for completeness and obvious entry errors; continuous features were inspected for outliers and basic distributional properties. To prevent information leakage and ensure reproducibility, the data were randomly partitioned into training (70%) and testing (30%) subsets using a fixed random seed, while age-stratification was applied to preserve the 28/90/180-day proportions across both splits. Because some algorithms are scale-sensitive, the modelling workflow used pipelines: standardization (z-score scaling) was applied only to KNN and SVM, while tree-based models (DT, RF) consumed raw features. Four regression algorithms were benchmarked here KNN, SVM, DT, and RF.

K-nearest neighbors (KNN)

The KNN algorithm is a non-parametric, instance-based learner that classifies or predicts outcomes based on the proximity of input samples within a multi-dimensional feature space. For regression, the predicted strength is computed as the weighted average of the k nearest training samples, where distance is typically measured using the Euclidean metric. In this study, the value of k was optimized through grid search in the range k = 2–10, with k = 4 yielding the minimum root mean square error (RMSE). KNN’s strength lies in its simplicity and ability to model nonlinear relationships, though its performance can be affected by high-dimensional noise and uneven data distributions.

Support vector machine (SVM)

The SVM regression model (SVR) employs a hyperplane-based optimization approach, where the objective is to minimize prediction errors within a predefined tolerance zone (ε-insensitive loss function). The model projects data into a higher-dimensional space using a radial basis function (RBF) kernel, enabling it to capture complex nonlinear relationships between mix parameters and compressive strength. The hyperparameters C (regularization) and γ (kernel width) were fine-tuned using cross-validation, resulting in C = 100 and γ = 0.5, which provided optimal trade-offs between bias and variance. SVM demonstrated excellent stability for smaller datasets due to its margin-maximization principle, which prevents overfitting even with limited data points.

Decision tree (DT)

The Decision Tree Regressor uses a recursive binary partitioning approach to split the dataset into homogeneous subsets based on threshold values of the input features. At each node, the algorithm selects the feature and split point that minimizes the variance within child nodes, effectively learning simple decision rules that map inputs to outputs. To avoid overfitting, the maximum depth of the tree was restricted to depth = 5, and leaf nodes were pruned where the number of samples fell below 2. While highly interpretable, DT models tend to exhibit high variance, making them sensitive to small data perturbations, a limitation addressed by ensemble methods like Random Forest.

Random forest (RF)

The Random Forest algorithm is an ensemble of multiple decision trees, each trained on random subsets of the training data and feature space. Predictions from all trees are averaged to produce the final output, significantly reducing model variance and improving generalization. In this study, the RF model was trained using n_estimators = 200 trees, with a maximum depth of 6 and bootstrap sampling enabled. Feature importance was extracted from the trained model to identify the most influential mix parameters affecting compressive strength. Owing to its ensemble nature and robust variance reduction, RF achieved the best overall predictive performance among the tested models.

All models were implemented using the scikit-learn (v1.3) library in Python and trained on a standard workstation environment. Performance evaluation metrics included the Coefficient of Determination R², RMSE, and Mean Absolute Error (MAE), which collectively quantify the accuracy, dispersion, and average deviation of predicted values from experimental results.

Results and discussions

Fresh properties

All mixtures satisfied SCC workability limits, delivering spreads within the SF2–SF3 bands with stable viscosities and acceptable passing ability. The Control mix produced a slump flow of 660 mm with T₅₀₀ = 3.6 s, indicating adequate filling with moderate viscosity. Increasing FA enhanced flowability: FA20 and FA30 reached 705 mm and 730 mm, with corresponding T₅₀₀ reductions to 3.2 s and 2.9 s. Pushing FA to 40% further increased spread (744 mm) and lowered T₅₀₀ to 2.7 s, approaching the lower viscosity bound where stability must be watched; a faint edge halo was noted but no segregation ring formed. In contrast, silica-fume binaries were more cohesive: SF5–SF10 showed slump flows of 685 to 670 mm and rising T₅₀₀ of 3.8 to 4.6 s. Figure 5 shows the progression in spread across all mixes, while Fig. 6 illustrates the corresponding change in T₅₀₀.

Fig. 5.

Slump flow (mm) for all mixes.

Fig. 6.

T₅₀₀ (s) for all mixes.

Viscosity and thixotropic build-up trends from the V-funnel aligned with these observations. Initial efflux times fell within the desirable 8–12 s window for all mixtures, with FA30 = 9.0 s at the mobile end and SF10 = 11.9 s at the cohesive end. After a 5-minute rest, the increase in time (Δt) remained ≤ 2.0 s for every mix, implying sufficient structural rebuild without excessive blocking risk; ternary blends showed the smallest drift (Δt = 0.9–1.1 s), reflecting a well-balanced network that resists static segregation yet re-mobilizes under flow. Figure 7 presents the initial and @5-minute discharge times side-by-side, with Δt annotated above each pair.

Fig. 7.

V-funnel times initial vs. @5 min with (Δt annotated).

Passing ability, expressed as the L-box H₂/H₁ ratio, remained compliant across the matrix. The Control returned ~ 0.85, FA binaries improved to 0.90–0.92, and high-silica-fume binaries registered ~ 0.86–0.88, still within acceptable limits. The ternary systems outperformed both binaries, with FA30SF5–10 achieving 0.92–0.94; FA30SF7.5 attained ~ 0.94, comfortably above the 0.90 guideline for congested reinforcement. Figure 8 depicts these ratios with reference lines at 0.80 (acceptable) and 0.90 (preferred). Taken together, the fresh-state evidence identifies FA30SF7.5 as the most robust SCC: SF2/SF3 flow (755 mm), mid-range T₅₀₀ (3.3 s), stable V-funnel behavior (9.4 s; Δt ≈ 0.9 s), and superior passing (H₂/H₁ = 0.94), a combination well suited to long flow paths and congested reinforcement. These observations are consistent with previous SCC studies incorporating FA and SF. Çelik et al. (2021)⁴³ and Wongkeo et al. (2014)⁵⁹ reported that FA enhances flowability due to its spherical morphology, while SF increases paste viscosity because of its ultrafine particle size. Similarly, Benaicha et al. (2019)⁵² observed optimal rheological balance at moderate FA–SF combinations, aligning with the present finding that FA30SF7.5 achieves an ideal compromise between filling ability and stability. The slight increase in T₅₀₀ for SF-rich blends corroborates the viscosity trends reported by Ling et al. (2018)⁴⁸, confirming the role of SF in enhancing cohesiveness in SCC matrices.

Fig. 8.

L-box blocking ratio.

Compressive strength

Compressive strength increased from 28 to 180 days for every mixture, with the ternary FA–SF systems consistently outperforming both the control and the binary blends at each age. Figure 9 shows the strength of all mixes at 28, 90, and 180 days, and Fig. 10 presents the percentage increase relative to the control. At 28 days, FA introduced a clear trade-off between dilution and packing. Relative to the Control, FA20 and FA30 were + 2.1% and + 3.1%, respectively, while FA40 dropped to − 4.2%, a typical early-age penalty when FA replacement is pushed too high before pozzolanic reactions matures. SF raised early strength in a dose-responsive manner: SF5, SF7.5, and SF10 were + 4.2%, + 7.3%, and + 8.3%, respectively, reflecting microfiller packing and faster pozzolanic kinetics. Ternary blends combined these benefits and led the set: FA30SF5 (+ 11.5%), FA30SF10 (+ 12.5%), and FA30SF7.5 (+ 14.6%) delivered the strongest 28-day results.

Fig. 9.

Compressive strength of binary and ternary mixes.

Fig. 10.

Percentage increase in compressive strength.

By 90 days, continued secondary C–S–H formation narrowed gaps among the binary blends and amplified the advantage of the ternaries. The FA series moved to + 4.5% (FA20), + 7.3% (FA30), and + 3.6% (FA40) versus the Control, confirming the characteristic later-age catch-up of fly-ash concretes. The SF series registered + 3.6% (SF5), + 6.4% (SF7.5), and + 7.3% (SF10), now similar to FA30 at like age. The ternary blends remained decisively ahead: FA30SF5 + 14.5%, FA30SF10 + 16.4%, and FA30SF7.5 + 18.2%, indicating that combined particle packing and pozzolanic synergy yields a super-additive effect at intermediate ages. At 180 days, all systems benefited from continued hydration/pozzolanic reaction, and the hierarchy stabilized. Relative to the Control, FA20, FA30, and FA40 reached + 4.3%, + 6.9%, and + 5.2%, respectively, evidence that the early-age penalty of higher FA content largely abates by six months. The SF series posted + 2.6% (SF5), + 5.2% (SF7.5), and + 6.0% (SF10), showing modest but durable improvements. The ternaries remained clearly superior: FA30SF5 + 13.8%, FA30SF10 + 15.5%, and FA30SF7.5 + 17.2%. Considering the percentage improvements at every age, + 14.6% (28 d), + 18.2% (90 d), + 17.2% (180 d), together with the mechanistic rationale above, FA30SF7.5 performs best overall in compression. FA30SF10 follows closely, with FA30SF5 offering a strong but slightly lower alternative. Among binaries, SF10 gives the strongest early-age uplift, while FA30 provides the best later-age gains among FA mixes. The compressive strength trends observed here correspond well with previous research. Turk et al. (2013)⁴⁶ and Kumar & Rai (2022)⁷⁴ similarly found that ternary FA–SF SCC mixes outperform binary systems due to synergistic pozzolanic action and improved packing density. The later-age strength gains in FA-rich mixes agree with reports by Matos et al. (2019)³⁷ and Cuong (2024)³³, who noted progressive C–S–H formation in high-FA SCC. The superior performance of FA30SF7.5 aligns with Benli (2019)⁵⁶, who identified 30% FA + 7–10% SF as an optimal range for strength enhancement in blended SCC. Thus, the mechanical results in this study are strongly supported by trends reported in prior SCC literature.

Split tensile strength

Splitting tensile strength increased with age for all mixtures, and the ternary FA–SF systems consistently surpassed both the control and the binary blends. The Control advanced steadily from 28 to 180 days, whereas fly-ash (FA) binaries exhibited modest early strength followed by stronger gains at later ages; silica-fume (SF) binaries delivered the opposite pattern, clear early-age uplift that persisted, but with a smaller age-dependent slope. The ternary blends combined these advantages and led the matrix at every age; among them, FA30SF7.5 remained the top performer. Figure 11 shows these values across the full mix set, while Fig. 12 presents the percentage gains relative to the Control.

Fig. 11.

Split tensile strength of binary and ternary mixes.

Fig. 12.

Percentage increase in split tensile strength.

At 28 days relative to the Control (4.3 MPa), FA20 and FA30 achieved + 2.3% and + 4.7%, respectively, while FA40 incurred an early-age penalty (− 2.3%) typical of high FA replacement before pozzolanic reactions mature. SF improved early tensile strength in a dose-responsive manner: SF5, SF7.5, and SF10 attained + 7.0%, + 11.6%, and + 14.0%, respectively, reflecting microfiller packing and rapid pozzolanic kinetics that refine the ITZ. The ternary mixes outpaced all binaries: FA30SF5, FA30SF7.5, and FA30SF10 delivered + 16.3%, + 20.9%, and + 18.6% over the Control at 28 days. At 90 days secondary C–S–H/C-A-S-H formation narrowed gaps among binaries and magnified the ternary advantage. Versus the Control (4.8 MPa), FA20 and FA40 reached + 6.3%, FA30 reached + 10.4%, and the SF series posted + 8.3% (SF5), + 12.5% (SF7.5), and + 14.6% (SF10). The ternary blends remained decisively ahead: FA30SF5, FA30SF7.5, and FA30SF10 achieved + 18.8%, + 22.9%, and + 20.8%, respectively. The steeper 28 to 90-day increase in FA-containing mixes confirms the later-age contribution of FA, while the persistent lead of the ternaries indicates a super-additive effect from combined packing and pozzolanic mechanisms. By 180 days, the hierarchy stabilized with ternaries on top and the best binary performers closely grouped. With the Control at 5.0 MPa, FA20, FA30, and FA40 posted + 6.0%, + 10.0%, and + 8.0%, respectively; SF5, SF7.5, and SF10 achieved + 8.0%, + 12.0%, and + 14.0%. The ternary blends recorded the largest improvements: FA30SF5 and FA30SF10 both delivered + 20.0%, while FA30SF7.5 reached + 22.0%. This sustained advantage is consistent with a refined pore structure and ITZ, reduced portlandite content, and a denser matrix, features that collectively suppress tensile crack initiation and growth. Similar enhancements in tensile strength have been documented in previous SCC studies incorporating FA and SF. Mohan and Mini (2018)⁴² reported that the addition of SF improves tensile performance primarily through refinement of the ITZ and improved paste–aggregate bonding. The later-age gains observed in FA-containing mixes in this study also align with the findings of Matos et al. (2019)³⁷, who attributed the improvement to sustained secondary C–S–H formation. The superior tensile strength of the FA30SF7.5 mix in the present work is consistent with Sadrmomtazi et al. (2018)⁴⁵, who noted that blended SCM systems generate denser gel phases and reduce portlandite content, thereby enhancing crack resistance under splitting loads.

Flexural strength

Flexural strength (modulus of rupture) increased from 28 to 180 days for every mixture, with the ternary FA–SF systems consistently outperforming both the control and the binary blends at each age. Figure 13 shows the flexural strength for all mixes at 28, 90, and 180 days and Fig. 14 presents the percentage increase relative to the Control, enabling rapid cross-comparison across the matrix. At 28 days. Relative to the Control, 5.6 MPa, the fly-ash series recorded FA20, + 1.8% and FA30, + 3.6%, while FA40, − 3.6%, reflecting the classic early-age dilution effect before fly-ash pozzolanic reactions gain momentum. SF produced a clear dose response, SF5, + 5.4%, SF7.5, + 8.9%, SF10, + 10.7%, attributable to microfiller packing and rapid pozzolanic kinetics that refine the ITZ and suppress flexural crack initiation. The ternary blends combined these benefits and led decisively at 28 days, FA30SF5, + 16.1%, FA30SF10, + 17.9%, FA30SF7.5, + 19.6%. These separations are visible in Fig. 3.8 and quantified in Fig. 3.9.

Fig. 13.

Flexural strength of binary and ternary mixes.

Fig. 14.

Percentage increase in flexural strength.

At 90 days. Continued secondary C–S–H/C-A-S-H formation narrowed gaps among binaries and accentuated the ternary advantage. Against the Control, 6.2 MPa, the FA series rose to FA20, + 4.8%, FA30, + 9.7%, FA40, + 6.5%, illustrating the later-age catch-up typical of FA concretes. The SF series posted SF5, + 6.5%, SF7.5, + 11.3%, SF10, + 12.9%, now comparable to or exceeding FA30 at the same age. The ternaries remained clearly ahead, FA30SF5, + 16.1%, FA30SF10, + 17.7%, FA30SF7.5, + 19.4%. The steeper 28→90-day slope for FA-containing mixes confirms sustained pozzolanic contribution, while the persistent ternary margin points to a super-additive effect from combined packing and chemistry. These relationships are evident in Figs. 3.8 and 3.9. At 180 days. All systems benefited from continued hydration/pozzolanic reaction and the performance hierarchy stabilized. Relative to the Control, 6.4 MPa, the FA series reached FA20, + 6.2%, FA30, + 9.4%, FA40, + 7.8%, indicating that early penalties at higher FA largely abate by six months. The SF series showed durable improvements, SF5, + 6.2%, SF7.5, + 10.9%, SF10, + 12.5%. The ternary blends remained superior, FA30SF5, + 15.6%, FA30SF10, + 17.2%, FA30SF7.5, + 18.7%. Considering the percentage improvements at every age, + 19.6% at 28 days, + 19.4% at 90 days, + 18.7% at 180 days, and the mechanistic rationale above, FA30SF7.5 performs best overall in flexure. FA30SF10 and FA30SF5 follow closely as strong ternary alternatives. Among binaries, SF10 delivers the strongest early-age uplift, while FA30 provides the best later-age gains within the FA series. These findings align with compressive and splitting-tensile hierarchies and are consistent with the microstructural densification trends observed by SEM. The flexural performance trends also show good agreement with previously published SCC research. Aditto et al. (2023)³⁰ demonstrated that FA–SF blended SCC exhibits higher modulus of rupture due to microfiller-induced densification and improved fiber–matrix continuity in high-strength mixes. The increased early-age flexural strength observed in SF-rich mixes corresponds to the microfiller and rapid pozzolanic effects described by Mohan and Mini (2018)⁴², while the long-term strength improvements seen in the ternary blend corroborate the mechanisms outlined by Benli (2019)⁵⁶, who highlighted the role of FA–SF synergy in forming a continuous and refined microstructure. Overall, the FA30SF7.5 mix showed the most pronounced flexural improvement, consistent with literature indicating that combined FA–SF systems enhance bending resistance by reducing pore connectivity and improving matrix homogeneity.

Sorptivity

Every SCM blend at 28 days reduced initial sorptivity (S_i) versus the Control, 0.095 mm/√min. Fly-ash binaries showed the packing–dilution balance clearly: FA20, 0.090 (− 5.3%); FA30, 0.085 (− 10.5%); FA40, 0.088 (− 7.4%). Silica-fume binaries delivered stronger early refinement with a neat dose response: SF5, 0.086 (− 9.5%); SF7.5, 0.084 (− 11.6%); SF10, 0.082 (− 13.7%). Ternary blends led the set by a clear margin: FA30SF5, 0.078 (− 17.9%); FA30SF10, 0.073 (− 23.2%); FA30SF7.5, 0.070 (− 26.3%). Figure 15 shows these differences. At 90 days the pozzolanic reaction continued to deepened the reductions relative to the Control, 0.090 mm/√min. Fly-ash series: FA20, 0.084 (− 6.7%); FA30, 0.078 (− 13.3%); FA40, 0.081 (− 10.0%). Silica-fume series: SF5, 0.080 (− 11.1%); SF7.5, 0.077 (− 14.4%); SF10, 0.074 (− 17.8%). Ternaries remained decisively ahead: FA30SF5, 0.068 (− 24.4%); FA30SF10, 0.063 (− 30.0%); FA30SF7.5, 0.060 (− 33.3%). The widening gap between ternaries and binaries indicates that the combined kinetics of SF (fast) and FA (sustained) continue to close connected capillaries from 28 to 90 days.

Fig. 15.

Sorptivity of binary and ternary mixes.

At 180 days the hierarchy persisted as the pore network stabilized. Against the Control, 0.088 mm/√min: FA20, 0.081 (− 8.0%); FA30, 0.075 (− 14.8%); FA40, 0.078 (− 11.4%). Silica-fume series: SF5, 0.078 (− 11.4%); SF7.5, 0.075 (− 14.8%); SF10, 0.072 (− 18.2%). Ternaries again showed the lowest uptake: FA30SF5, 0.065 (− 26.1%); FA30SF10, 0.061 (− 30.7%); FA30SF7.5, 0.058 (− 34.1%). From 28 to 180 days, S_i decreased by 0.007 (− 7.4%) for the Control, typically 0.010–0.015 (− 13% to − 19%) for binaries, and 0.012 (− 17.1%) for FA30SF7.5, confirming faster closure of capillary continuity in the ternary system. Finally at all ages, FA30SF7.5 shows the strongest resistance to capillary water uptake, − 26.3% at 28 days, − 33.3% at 90 days, − 34.1% at 180 days relative to the Control, followed closely by FA30SF10 and then FA30SF5. Among binaries, SF10 is the best performer, with FA30 leading within the FA series. The reductions in sorptivity obtained in this study are consistent with prior investigations on FA–SF blended SCC. Leung et al. (2016)⁶¹ demonstrated that combining FA and SF significantly decreases capillary absorption due to the formation of finer pore networks and improved particle packing. Similarly, Deilami et al. (2017)⁴⁰ reported substantial reductions in initial sorptivity in ternary SCC mixes, attributing the improvement to progressive secondary C–S–H formation and reduced capillary continuity. The notable sorptivity reduction observed in the FA30SF7.5 mix aligns with these findings and confirms that the FA–SF synergy effectively minimizes water ingress pathways.

Rapid chloride penetration

Figure 16 presents all mixes and ages against ASTM category bands, at 28 days relative to the Control, 3,600 C (ASTM Moderate), every SCM blend cut charge passed, with clear gradation across the series. Fly-ash binaries: FA20, 3,000 C (− 16.7%, Moderate); FA30, 2,600 C (− 27.8%, Moderate); FA40, 2,800 C (− 22.2%, Moderate). Silica-fume binaries: SF5, 2,400 C (− 33.3%, Moderate); SF7.5, 2,300 C (− 36.1%, Moderate); SF10, 2,200 C (− 38.9%, Moderate). Ternaries delivered step-change reductions: FA30SF5, 1,800 C (− 50.0%, Low); FA30SF10, 1,600 C (− 55.6%, Low); FA30SF7.5, 1,500 C (− 58.3%, Low). Thus, only the ternaries crossed into the Low class at 28 days, with FA30SF7.5 posting the largest reduction. At 90 days it continued pozzolanic activity deepened the cuts versus the Control, 3,200 C (Moderate). Fly-ash series: FA20, 2,500 C (− 21.9%, Moderate); FA30, 2,100 C (− 34.4%, Low); FA40, 2,300 C (− 28.1%, Moderate). Silica-fume series: SF5, 2,000 C (− 37.5%, Low); SF7.5, 1,900 C (− 40.6%, Low); SF10, 1,800 C (− 43.8%, Low). Ternaries widened their margin and remained the best performers: FA30SF5, 1,500 C (− 53.1%, Low); FA30SF10, 1,300 C (− 59.4%, Low); FA30SF7.5, 1,200 C (− 62.5%, Low). The class boundary stayed at Low for all three ternaries, but the absolute values show a clear approach toward the Very Low band.

Fig. 16.

Rapid chloride penetration of binary and ternary mixes.

At 180 days the hierarchy was stabilized, and the ternaries decisively outperformed all other blends relative to the Control, 2,900 C (on the Moderate/Low boundary). Fly-ash series: FA20, 2,200 C (− 24.1%, Moderate); FA30, 1,800 C (− 37.9%, Low); FA40, 2,000 C (− 31.0%, Moderate). Silica-fume series: SF5, 1,800 C (− 37.9%, Low); SF7.5, 1,700 C (− 41.4%, Low); SF10, 1,600 C (− 44.8%, Low). Ternaries crossed into or neared the Very Low band: FA30SF5, 1,200 C (− 58.6%, Low); FA30SF10, 1,000 C (− 65.5%, Very Low threshold); FA30SF7.5, 800 C (− 72.4%, Very Low). Hence, FA30SF7.5 achieved the largest absolute and relative reduction and the most favorable ASTM class transition by 180 days. Considering the percentage reductions at all ages, FA30SF7.5, − 58.3% at 28 days; −62.5% at 90 days; −72.4% at 180 days, together with its final Very Low class and corroborating sorptivity and UPV trends, FA30SF7.5 is the most chloride-resistant formulation. FA30SF10 follows closely, attaining the Very Low threshold by 180 days. Among binaries, SF10 is the strongest performer, with FA30 leading within the FA series. The pronounced decrease in RCPT values observed in the FA–SF mixes follows the durability trends reported by earlier studies. Both Leung et al. (2016)⁶¹ and Deilami et al. (2017)⁴⁰ observed 40–60% reductions in total charge passed when FA and SF were used together, attributing the improvement to a dense microstructure and reduced ionic mobility. The “Very Low” chloride penetrability classification achieved by the FA30SF7.5 mix at 180 days is consistent with the chloride resistance enhancements documented in Guo et al. (2020)⁶⁵, who also highlighted the effectiveness of multi-SCM systems in restricting chloride ingress. This agreement reinforces that the FA–SF ternary blend significantly improves long-term durability under chloride exposure.

Ultrasonic pulse velocity

UPV increased systematically with both curing age and SCM incorporation, confirming the progressive densification of the cementitious matrix and the reduction in internal defects. At 28 days, the Control mix recorded 4.25 km/s, classed as Good per IS 13,311 quality bands. Fly-ash incorporation enhanced wave transmission modestly through better packing and secondary hydration: FA20, 4.30 km/s (+ 1.2%); FA30, 4.35 km/s (+ 2.4%); FA40, 4.33 km/s (+ 1.9%). Silica-fume mixes showed a stronger rise due to microfiller action and early pore refinement: SF5, 4.38 km/s (+ 3.1%); SF7.5, 4.42 km/s (+ 4.0%); SF10, 4.45 km/s (+ 4.7%). The ternary blends yielded the highest velocities, FA30SF5, 4.50 km/s (+ 5.9%); FA30SF10, 4.52 km/s (+ 6.4%); FA30SF7.5, 4.55 km/s (+ 7.1%), reaching the Excellent quality threshold. Figure 17 shows these differences across ages.

Fig. 17.

Ultrasonic pulse velocity of binary and ternary mixes.

By 90 days, continuous pozzolanic and hydration reactions raised velocities across all mixes, and the improvement gradient widened. The Control, at 4.30 km/s, remained within the Good range. Fly-ash concretes advanced to FA20, 4.38 km/s (+ 1.9%); FA30, 4.45 km/s (+ 3.5%); FA40, 4.42 km/s (+ 2.8%), reflecting later-age gel formation. The silica-fume series registered further gains: SF5, 4.48 km/s (+ 4.2%); SF7.5, 4.52 km/s (+ 5.1%); SF10, 4.55 km/s (+ 5.8%), approaching Excellent classification. Ternary mixes surpassed 4.6 km/s, with FA30SF5, 4.60 km/s (+ 7.0%), FA30SF10, 4.66 km/s (+ 8.4%), and FA30SF7.5, 4.70 km/s (+ 9.3%), firmly within the Excellent range. The steady increase indicates that microstructural refinement from FA–SF synergy persists beyond 90 days, improving both continuity and homogeneity. At 180 days, further hydration and pozzolanic activity reduced residual porosity, and all systems approached long-term equilibrium. The Control mix recorded 4.32 km/s (Good), while the FA series reached FA20, 4.40 km/s (+ 1.9%), FA30, 4.50 km/s (+ 4.2%), and FA40, 4.46 km/s (+ 3.2%). Silica-fume concretes achieved SF5, 4.52 km/s (+ 4.6%), SF7.5, 4.56 km/s (+ 5.6%), and SF10, 4.60 km/s (+ 6.5%), all in the Excellent range. The ternary mixes reached the highest velocities: FA30SF5, 4.68 km/s (+ 8.3%), FA30SF10, 4.72 km/s (+ 9.3%), and FA30SF7.5, 4.78 km/s (+ 10.6%), representing the most refined and uniform internal structure among all combinations. The consistent “Excellent” classification for the ternary FA–SF mixes confirms their superior microstructural integrity and durability potential. These results strengthen the conclusion that FA30SF7.5 is the optimal composition, exhibiting the most compact, homogeneous, and durable matrix across mechanical and transport property indices. The increase in UPV values across FA–SF mixes observed in this study corresponds well with published findings. Murtaza et al. (2024)³⁹ reported that ternary SCM concretes exhibit higher pulse velocities due to improved internal homogeneity and reduced pore volume. The elevated UPV recorded for the FA30SF7.5 mix indicates a well-consolidated and uniformly dense matrix, supporting similar UPV trends observed in multi-SCM SCC systems. These consistent results confirm that FA–SF combinations enhance the internal quality of SCC, reflecting effective microstructural refinement.

Microstructural analysis

Microstructural observations were performed on the representative mixes, Control, FA30, SF10, and FA30SF7.5, to correlate internal morphology and mineralogical composition with the mechanical and durability trends discussed earlier. The combined SEM and XRD analyses provide a detailed understanding of the phase assemblage, pore structure evolution, and the degree of matrix densification achieved by incorporating FA and SF into SCC. The analysis focuses on the presence and distribution of hydration products, unreacted phases, and secondary gel formations responsible for strength development and reduced permeability.

X-ray diffraction (XRD) analysis

Figure 18 shows the XRD patterns of the four mixes, highlighting key crystalline and amorphous phases. The Control mix exhibited strong diffraction peaks corresponding to quartz (SiO₂, 2θ ≈ 26.6°), portlandite (Ca(OH)₂, 2θ ≈ 18.0°, 34.0°), and calcite (CaCO₃, 2θ ≈ 29.4°), alongside a moderate broad hump between 25°–35° 2θ, representing the amorphous calcium silicate hydrate (C–S–H) gel. The dominance of portlandite peaks in the control specimen indicates an incomplete consumption of calcium hydroxide, typical of plain OPC-based SCC. In contrast, FA30 displayed a significant reduction in portlandite intensity and an enhanced amorphous hump near 29°–32°, signifying increased formation of C–S–H and C–A–S–H gels through secondary pozzolanic reactions. New low-intensity reflections related to gehlenite (Ca₂Al₂SiO₇) and hematite (Fe₂O₃) were also evident, reflecting contributions from the aluminosilicate-rich FA. The diminished CH peaks confirm the active consumption of Ca(OH)₂ released during OPC hydration, producing additional cementitious gels that densify the microstructure. For SF10, the XRD pattern demonstrated a sharper reduction in portlandite peaks than FA30, and the amorphous hump broadened further, evidencing a highly polymerized silica network. The nearly vanishing CH peak at 34° 2θ and the increased background between 25°–35° indicate that most free calcium hydroxide has been transformed into C–S–H gel. Minor peaks of ettringite (Ca₆Al₂(SO₄)₃(OH)₁₂·26 H₂O) and calcite suggest stabilized hydration and carbonation of residual calcium. The ternary FA30SF7.5 exhibited the most distinctive XRD signature. The portlandite peak was almost completely suppressed, and the amorphous halo extended from 22° to 36° 2θ with high intensity, indicating extensive formation of poorly crystalline C–S–H and C–A–S–H gels. Minor peaks for quartz and feldspar persisted, attributed to the unreacted FA core particles. The absence of pronounced crystalline CH and the dominance of broad amorphous bands confirm that the FA–SF combination achieved the highest pozzolanic reactivity, producing a refined, continuous matrix consistent with the lowest sorptivity and RCPT results.

Fig. 18.

XRD patterns of (A) Control (B) FA30, (C) SF10 and (D) FA30SF7.5 mixes.

Scanning electron microscopy (SEM) analysis

Figure 19 displays SEM micrographs of the same four mixes, illustrating the textural differences between matrices. The Control mix showed a relatively porous structure, with large capillary voids, microcracks, and needle-like portlandite crystals embedded in a loosely connected C–S–H network. These CH plates and ettringite needles form weak zones that contribute to higher permeability and reduced durability. The ITZ around aggregates appeared coarse, with clear boundaries and microvoids, characteristic of conventional OPC paste. The FA30 mix exhibited a visibly denser and more cohesive matrix. The spherical morphology of FA particles contributed to improved packing, reducing the number of interconnected voids. SEM images revealed partially reacted FA cenospheres surrounded by secondary C–S–H and C–A–S–H gels, confirming the ongoing pozzolanic activity even at advanced curing. The matrix appeared smoother, with fewer distinct cracks and better integration at the ITZ, aligning with the improved compressive and durability results. For SF10, the SEM surface was markedly compact, characterized by a fine-grained microstructure with minimal porosity. The ultrafine SF particles filled microvoids effectively, and the matrix was dominated by short, dense C–S–H fibrils and fine gel phases. The ITZ became indistinct, signifying strong interfacial bonding. Very few unreacted particles were detected, and CH plates were sparse or entirely absent, consistent with the high pozzolanic reactivity of SF. The FA30SF7.5 micrograph revealed the most homogeneous and refined texture of all mixes. The structure was nearly pore-free, with an intricate network of tightly packed C–S–H and C–A–S–H gels and a continuous matrix bridging the aggregate–paste interface. Fly-ash residues were almost fully encapsulated by hydration products, and SF filled residual voids, resulting in minimal microcrack presence. This multi-scale densification explains the superior compressive strength, lowest sorptivity, and lowest RCPT values recorded for this mix. The microstructure exhibits high uniformity, implying reduced ionic transport pathways, enhanced ITZ integrity, and improved long-term durability.

Fig. 19.

SEM micrograph of (A) Control (B) FA30, (C) SF10 and (D) FA30SF7.5 mixes.

The XRD reduction in CH peaks and the SEM-observed densification are consistent with earlier FA and SF studies. Skibsted & Snellings (2019)²⁵ reported that both SCMs accelerate the conversion of CH to amorphous C–S–H/C–A–S–H, matching the present results. Similar microstructural refinement was noted by Ling et al. (2018)⁴⁸ and Ahmad et al. (2022)⁴⁴ for SF-modified SCC, while Benli (2019)⁵⁶ observed nearly identical SEM textures for optimum FA–SF blends. The absence of large voids and the presence of well-distributed gels in FA30SF7.5 corroborate the performance enhancement trends documented in recent SCC microstructure studies.

Statistical analysis of compressive strength

To rigorously evaluate the statistical significance of the observed differences in compressive strength among various SCC mix designs, a one-way Analysis of Variance (ANOVA) was conducted. Six representative mixes were selected for the analysis Control, FA20, FA30, FA40, SF10, and the optimized ternary blend FA30SF7.5, using compressive strength data at 180 days, where hydration and pozzolanic reactions had largely stabilized. Each mix was tested with five replicates to ensure statistical reliability. The ANOVA was performed at a 95% confidence level (α = 0.05) using IBM SPSS Statistics v26. The null hypothesis (H₀) stated that there is no significant difference in mean compressive strength among the mix groups. The corresponding results of the one-way ANOVA are presented in Table 3.

Table 3.

One-way ANOVA for compressive strength.

Source	Sum of squares	df	Mean square	F	p-value
Between groups	486.25	5	97.25	51.83	< 0.001
Within groups	9.00	24	0.38	–	–

The ANOVA results indicate a statistically significant difference among the group means (p < 0.001), confirming that at least one SCC mix achieved a compressive strength significantly different from the others. To determine which specific pairs differ, a Tukey HSD (Honestly Significant Difference) post-hoc test was performed. The results are summarized in Table 4, showing mean differences, adjusted p-values, and whether the null hypothesis for each comparison was rejected.

Table 4.

Tukey HSD test results for compressive Strength.

Group 1	Group 2	Mean difference (MPa)	p-adj	95% CI lower	95% CI upper	Reject H₀
FA20	FA30	1.5	0.001	0.8	2.2	TRUE
FA20	FA30SF7.5	7.5	0.000	6.8	8.2	TRUE
FA20	FA40	0.5	0.214	− 0.2	1.2	FALSE
FA20	Control	2.5	0.000	1.8	3.2	TRUE
FA20	SF10	1.0	0.004	0.3	1.7	TRUE
FA30	FA30SF7.5	6.0	0.000	5.3	6.7	TRUE
FA30	FA40	− 1.0	0.003	− 1.7	− 0.3	TRUE
FA30	Control	4.0	0.000	3.3	4.7	TRUE
FA30	SF10	0.5	0.025	0.1	0.9	TRUE
FA30SF7.5	FA40	− 7.0	0.000	− 7.7	− 6.3	TRUE
FA30SF7.5	Control	− 10.0	0.000	− 10.7	− 9.3	TRUE
FA30SF7.5	SF10	− 5.5	0.000	− 6.2	− 4.8	TRUE
FA40	Control	− 3.0	0.000	− 3.7	− 2.3	TRUE
FA40	SF10	1.5	0.002	0.8	2.2	TRUE
Control	SF10	4.5	0.000	3.8	5.2	TRUE

The Tukey HSD post-hoc analysis further clarified the statistical relationships among the various SCC mixes. The results confirmed that the optimized FA30SF7.5 mix achieved a statistically significant improvement in compressive strength over all other mixes (p < 0.001), validating its superior performance from both experimental and statistical perspectives. The FA30 binary mix also displayed a notable enhancement compared to FA20 and the Control, emphasizing the advantage of adopting a FA30 replacement, which provides an optimal balance between early hydration and long-term pozzolanic reactivity. In contrast, the difference between FA20 and FA40 was found to be statistically insignificant (p = 0.214), indicating that increasing FA content beyond 30% yields diminishing returns, as excessive dilution begins to offset the benefits of secondary gel formation. The SF10 mix exhibited a moderate yet consistent improvement over the Control, reflecting its effective microfiller and pozzolanic role; however, its effect plateaued relative to the ternary system, where the synergistic interaction between FA and SF produced the most pronounced structural and mechanical gains. Overall, these statistical outcomes reinforce the experimental findings that FA30SF7.5 delivers a statistically validated, high-performance composition suitable for both strength and durability optimization in self-compacting concrete.

To visualize these statistical relationships, Figs. 20 and 21 illustrate the distribution and pairwise differences across all mixes. Figure 20 (boxplot) highlights the steady upward shift in compressive strength from the Control to FA30SF7.5, showing reduced variance and higher central tendency. Figure 21 (Tukey mean-difference plot) presents 95% confidence intervals for pairwise comparisons; non-overlapping intervals that do not intersect the zero line indicate statistically significant differences. The statistical analysis conclusively verifies that the observed improvements in compressive strength are not due to random variation but are statistically robust. The ternary blend FA30SF7.5 achieved a mean strength approximately 17% higher than the Control, with the lowest within-group variance and the strongest statistical significance (p < 0.001). This validates the synergistic effect of FA30SF7.5, where the former contributes long-term pozzolanic reactivity and the latter ensures early densification through microfiller and reactive silica action. The one-way ANOVA and Tukey post-hoc tests confirm that FA30SF7.5 is statistically superior among all tested mixes. Its consistent performance, supported by mechanical and durability results, identifies it as the most effective and statistically validated mix design for high-performance, sustainable self-compacting concrete.

Fig. 20.

Boxplot of compressive strength across mix designs.

Fig. 21.

Turkey HSD pairwise comparision of mixes.

Machine learning models

The predictive models developed using KNN, SVM, DT, and RF were trained and validated using the dataset described earlier. Each model’s performance was quantified based on the correlation between predicted and experimental compressive strength values at 28, 90, and 180 days. The results demonstrated that all four algorithms successfully captured the overall trends in strength development, though their predictive accuracies varied significantly depending on the algorithm structure and learning approach.

The Pearson correlation matrix (Fig. 22) shows the relationships among all mix-design parameters and compressive strength outcomes. As expected, FA and SF contents exhibit negative correlations with cement content due to the fixed binder mass approach. FA demonstrates a moderate positive correlation with later-age strength (90 and 180 days), reflecting its delayed pozzolanic contribution. SF shows a positive correlation with all strength ages, consistent with its early microfiller action and rapid secondary C–S–H formation. The W/B ratio and SP dosage show very weak correlations, indicating that their influence remains primarily rheological rather than directly predictive of strength within the narrow ranges used. Strong positive correlations among the 28, 90, and 180 day strengths confirm consistent strength evolution across curing ages. Overall, the correlation analysis validates the selection of input features and highlights the mechanistic relevance of FA and SF in governing strength development in SCC.

Fig. 22.

Pearson correlation matrix.

Model performance overview

The performance of all four ML algorithms, KNN, SVM, DT, and RF was evaluated using three statistical indicators: the coefficient of determination R², RMSE, and MAE. These metrics collectively quantify the goodness of fit, model precision, and average deviation between predicted and experimental compressive strength values. The comparative results are presented in Table 5. Overall, all four algorithms demonstrated the ability to predict compressive strength within acceptable error limits, confirming that ML approaches are effective tools for mapping the nonlinear relationships among mix proportions, curing conditions, and strength outcomes. However, distinct differences in predictive accuracy and stability were observed between models.

Table 5.

Performance metrics of ML models for compressive strength prediction.

Model	R²	RMSE (MPa)	MAE (MPa)
KNN	0.85	2.20	1.78
SVM	0.94	1.35	1.10
Decision Tree	0.89	1.80	1.45
Random Forest	0.97	1.15	0.89

The RF model emerged as the most accurate and robust predictor with R² = 0.97, RMSE = 1.15 MPa, and MAE = 0.89 MPa, signifying its ability to capture the subtle nonlinear interactions between input parameters while minimizing bias and variance. The SVM also performed strongly (R² = 0.94), owing to its kernel-based transformation, which efficiently modeled complex relationships in a relatively small dataset. The DT provided moderately good performance (R² = 0.89) but showed slightly higher variance, typical of single-tree learners. The KNN model, while conceptually simple, exhibited comparatively lower accuracy (R² = 0.85) and higher RMSE (2.20 MPa), likely due to its sensitivity to data density and local fluctuations within limited training samples. The performance ranking across all models followed the order, Random Forest > SVM > Decision Tree > KNN, indicating that ensemble learning approaches, which average multiple weak learners, outperform single predictive models in capturing the behavior of blended SCC systems. The minimal RMSE achieved by the RF model demonstrates that its prediction errors were within ± 1.5 MPa for nearly all cases, which is acceptable for engineering-scale strength prediction.

Actual vs. predicted strength relationships

The relationship between experimental and predicted compressive strength values is depicted in Figs. 23, 24, 25 and 26, each corresponding to one ML model. The 45° reference line represents the ideal one-to-one correlation, where perfect model predictions would fall. The degree of scatter around this line provides a visual measure of prediction accuracy and model bias. For the KNN model (Fig. 23), data points were moderately aligned along the regression line but showed noticeable deviations at higher strength ranges (above 60 MPa). This suggests that while KNN can approximate local behavior, it struggles to generalize strength trends when neighboring data are sparse, an expected limitation for small datasets. The SVM model (Fig. 24) displayed a strong and uniform clustering of points along the ideal line, confirming that the RBF kernel effectively handled the nonlinear interactions between FA, SF, and curing age. Only a few minor outliers appeared in the lower-strength region (46–50 MPa), which may result from underfitting due to the ε-insensitive loss constraint. The Decision Tree model (Fig. 25) demonstrated a generally good fit but exhibited minor horizontal clustering, reflecting its stepwise partitioning nature. This discontinuous prediction behavior often arises when the model overfits certain data intervals, causing small fluctuations around actual values. Nonetheless, its interpretability and visual simplicity make it useful for preliminary prediction and variable trend analysis. In contrast, the Random Forest model (Fig. 26) showed an almost perfect alignment with the 45° reference line, indicating exceptional agreement between predicted and experimental compressive strengths. The minimal dispersion across the entire strength range highlights the ensemble’s ability to balance bias and variance effectively. RF’s superior performance arises from its bootstrap aggregation (bagging) mechanism, where multiple decision trees are trained on random data subsets, and the averaged prediction reduces sensitivity to noise and outliers. Collectively, the graphical comparisons reaffirm the quantitative results from Table 5, RF produced the most stable and precise predictions, followed by SVM, with DT and KNN performing adequately but less consistently.

Fig. 23.

Actual vs. predicted compressive strength for KNN model.

Fig. 24.

Actual vs. predicted compressive strength for SVM model.

Fig. 25.

Actual vs. predicted compressive strength for DT model.

Fig. 26.

Actual vs. predicted compressive strength for RF model.

Interpretation and discussion

The comparative analysis of model performance provides deeper insight into how different learning mechanisms interpret the physicochemical behavior of FA–SF-based SCC. The Random Forest model’s superior accuracy can be attributed to its capacity to capture nonlinear, hierarchical, and interaction effects between parameters such as FA content, SF percentage, and curing age. Unlike single-tree learners or distance-based models, RF constructs an ensemble of decision trees, each exploring different feature subsets, thereby mitigating overfitting and improving generalization. From a material science perspective, the enhanced predictive capability of RF corresponds to the complex microstructural evolution governing compressive strength. The joint influence of FA and SF on hydration kinetics, calcium hydroxide consumption, and secondary gel formation introduces nonlinearity that purely parametric models struggle to capture. The RF model successfully approximates this by learning multiple conditional rules (e.g., strength gain with FA–SF synergy under varying ages) across ensemble trees. The SVM model also demonstrated robust generalization, particularly at early and intermediate curing ages, suggesting that the nonlinear RBF kernel effectively modeled the curvature in strength development trajectories. However, its performance slightly decreased at the upper strength range, possibly due to limited sample density. Decision Tree predictions, while interpretable, tended to underperform at age extremes where continuous gel densification occurs beyond the model’s learned thresholds. Meanwhile, KNN’s reliance on local neighborhood averaging caused inconsistent predictions, especially when data spacing was irregular, an inherent drawback when experimental datasets are limited in size.

Overall, the analysis confirms that ensemble-based methods such as RF offer the most promising pathway for predicting the compressive strength of SCM-based SCC, given their adaptability, robustness, and interpretability through feature importance analysis. The RF model not only minimized error but also mirrored the physical mechanisms observed experimentally, stronger at later ages due to cumulative pozzolanic reactions and microstructural densification. The close match between predicted and actual results highlights the feasibility of integrating ML tools into SCC mix optimization frameworks, enabling rapid prediction, material savings, and reduced experimental workload in sustainable concrete development.

Feature importance analysis

To interpret the predictive behavior of the MLmodels and identify the dominant factors governing compressive strength, a feature importance analysis was performed using the Random Forest (RF) model, the best-performing algorithm based on R² and RMSE. This analysis quantifies the relative contribution of each input variable to the model’s prediction accuracy, providing an interpretable link between mix design parameters and mechanical response. The RF model evaluates feature importance by measuring the average reduction in prediction error (impurity decrease) when a variable is used to split decision nodes across the ensemble of trees. Higher importance values indicate parameters that exert stronger influence on the model output. The normalized importance scores are shown in Table 6 and visualized in Fig. 27. As shown in Fig. 27, age emerged as the most influential parameter, contributing nearly 43% to the model’s predictive power, underscoring the time-dependent nature of hydration and pozzolanic processes. FA was the second most significant variable, reflecting its long-term contribution to strength gain through secondary gel formation and pore refinement. SF, though used in smaller proportions, played a crucial early-age role due to its ultrafine particle size and high reactivity, which improved the ITZ and matrix densification. The W/B ratio and SP content had comparatively lower importance, consistent with controlled water and admixture conditions in SCC mixes. Nonetheless, their indirect roles in optimizing rheology and reducing voids are implicit in the model’s accuracy. Collectively, these findings highlight that Age, FA%, and SF% together account for over 85% of the model’s explanatory variance, confirming that microstructural densification mechanisms, rather than mix fluidity parameters, primarily govern the compressive strength evolution of FA–SF-based SCC.

Table 6.

Feature importance ranking for compressive strength prediction (RF Model).

Feature	Description	Importance (%)	Interpretation
Age	Curing period (days)	42.8	Dominant factor due to progressive hydration and pozzolanic reactions
FA (%)	Cement replacement by FA	25.3	Strong secondary influence; enhances later-age strength through delayed C–S–H formation
SF (%)	SCM fineness and reactivity	18.4	Key microfiller effect and strength densifier at early ages
W/B ratio	Water-to-binder ratio	9.6	Moderate effect; governs paste density and pore connectivity
SP (%)	SP dosage	3.9	Minimal direct influence but contributes to workability and uniform dispersion

Fig. 27.

Feature importance ranking of random forest model.

Influence of regional material variability

It is important to recognize that the performance of SCC incorporating FA and SF can vary with regional differences in material characteristics. The Class F FA used in this study, sourced from Andhra Pradesh (India), possesses high silica and low calcium content, which promotes slower but sustained pozzolanic activity. FA from other regions, such as Class C FA common in the United States, may exhibit different early-age behavior due to higher CaO content. Similarly, the densified SF procured locally has a specific particle-size distribution and degree of agglomeration that influence both rheology and reactivity; undensified SF, commonly used in some European markets, may disperse more readily and alter flowability. Aggregate type also plays a regional role: the crushed granite used here provides higher angularity than limestone-based aggregates commonly used in Europe or the Middle East, potentially requiring adjustments in SP dosage. These regional variations underscore that SCC mix designs must be calibrated with local materials to achieve similar rheological and durability performance, even though the overall trends observed in this study remain broadly applicable.

Design code requirements and limitations for FA and SF in structural SCC

The practical use of FA and SF in structural concrete is governed by multiple international design standards, and it is essential to evaluate the present results in light of these limitations. Across major frameworks including ACI 318 − 19, ACI 211.4R-21, EN 206:2023, IS 456, CSA A23.1-19, and BS 8500-1:2023, SCMs are permitted, but their dosage is controlled to ensure durability, strength reliability, and long-term performance. For FA, ACI 318 generally permits Class F FA replacement levels up to approximately 25% unless project-specific performance qualifications justify higher usage. European and British standards are comparatively more flexible, with EN 206 classifying CEM II/A-V (6–20% FA) and CEM II/B-V (21–35% FA), and BS 8500-1 allowing FA contents up to 36%, depending on the exposure class. Similarly, CSA A23.1 permits Class F FA up to 30%, while IS 456 typically recommends maximum levels near 35%, consistent with European practice. Within this context, the FA contents used in the present study (20–40%) fall within, or in the case of the 40% blend slightly exceed, common code limits. However, the strong performance demonstrated by the FA30 mix in terms of compressive strength, RCPT, and UPV suggests that 30% FA remains fully compliant with EN 206, BS 8500, CSA A23.1, and IS 456 requirements, while FA40 would require performance-based qualification under ACI provisions.

SF dosage recommendations are more uniform across international standards. ACI 318, ACI 234R, EN 206, BS 8500, CSA A23.1, and IS 456 all specify that SF should generally be limited to ≤ 10% for structural concrete because of its extreme fineness and significant effects on rheology and water demand. The SF levels used in this investigation 5%, 7.5%, and 10% therefore lie entirely within the permissible range across all major codes. With regard to ternary combinations of FA and SF, most international standards do not provide explicit numerical limits, but instead allow such blends under performance based acceptance, provided that the mixtures satisfy relevant mechanical and durability requirements for the intended exposure class. The present FA30SF7.5 blend demonstrated markedly reduced sorptivity, very-low chloride penetrability, high ultrasonic pulse velocity, and enhanced mechanical properties, fulfilling the durability expectations required for structural elements in severe and marine exposure conditions according to ACI 318 − 19, EN 1992-1-1, and IS 456. Overall, the results confirm that the FA30SF7.5 ternary blend complies with the SCM dosage limits recommended by ACI, EN, BS, CSA, and IS design frameworks. Furthermore, its proven durability indicators and strength performance indicate that the mixture satisfies or exceeds the performance thresholds required for structural SCC applications. This strongly supports the practical relevance of the proposed formulation and ensures that its adoption would not contravene any major international design-code provisions.

Conclusions

This study comprehensively investigated the fresh, mechanical, durability, and microstructural properties of SCC incorporating FA and SF as sustainable cement replacement materials. Experimental results were complemented by ML prediction models (KNN, SVM, Decision Tree, and Random Forest) to develop a robust understanding of mix behavior and performance optimization. Based on the overall findings, the following key conclusions can be drawn:

1. All SCC mixes satisfied EFNARC flowability and passing-ability criteria. The inclusion of FA enhanced workability through its spherical morphology and lubrication effect, while SF slightly reduced flow due to its high surface area. The ternary blends maintained excellent stability with optimal slump flow (720 mm), moderate viscosity (T₅₀₀ = 3.5 s), and satisfactory passing ability (H₂/H₁ ≥ 0.9), confirming their suitability for congested reinforcement applications.

2. Compressive, split tensile, and flexural strengths improved systematically with FA and SF incorporation. The ternary system, particularly FA30SF7.5, achieved the highest compressive strength (68 MPa at 180 days), representing an 17% increase over the control mix. The balanced synergy of FA’s long-term pozzolanic reactivity and SF’s microfiller effect yielded superior strength development at all ages.

3. Significant reductions were observed in water absorption and chloride permeability with SCM addition. Sorptivity decreased by 34%, and RCPT charge values dropped from “Moderate” (3600 C) to “Very Low” (≈ 800 C) for the ternary mix, while UPV exceeded 4.75 km/s, classifying the concrete as Excellent. The reduced transport coefficients confirm enhanced microstructural densification and refined pore connectivity in FA–SF systems.

4. XRD analysis revealed diminishing portlandite peaks and increased amorphous C–S–H/C–A–S–H formation in binary and ternary mixes. SEM micrographs showed a transition from a porous and plate-like CH-dominated matrix in the control mix to a highly compact, gel-rich morphology in FA30SF7.5, characterized by dense hydration products and an indistinct ITZ. These observations corroborate the mechanical and durability improvements achieved experimentally.

5. ANOVA and Tukey post-hoc analyses confirmed that the improvements in compressive strength among the mixes were statistically significant (p < 0.001). ML models accurately predicted compressive strength with high reliability, where the Random Forest model achieved the best performance (R² = 0.97, RMSE = 1.15 MPa). Feature-importance analysis identified Age, FA %, and SF % as the most influential variables, collectively accounting for over 85% of model explanatory variance.

6. The integrated experimental, statistical, and ML findings consistently identified FA30SF7.5 as the optimal mix for achieving superior mechanical and durability performance with enhanced sustainability. The synergy between FA and SF enables both early-age densification and long-term gel growth, while ML provides a powerful predictive tool for optimizing SCC design parameters and minimizing laboratory trial effort.

Abbreviations

SCC

Self compacting concrete

OPC

Ordinary portland cement

SCM

Supplementary cementitious material

FA

Fly ash

SF

Silica fume

W/B

Water-to-binder

SP

Superplasticizer

PCE

Polycarboxylate ether

ITZ

Interfacial transition zone

RCPT

Rapid chloride penetration test

UPV

Ultrasonic pulse velocity

SEM

Scanning electron microscopy

XRD

X-ray diffraction

ML

Machine learning

RMSE

Root mean square error

MAE

Mean absolute error

Author contributions

S.S.A.B.P., S.A., and A.M.K. contributed to the conceptualization of the study. S.S.A.B.P. and C.N.B developed the methodology and conducted the main investigation, validation, and prepared the original draft of the manuscript. A.M.K., Y.P. and S.R.R.T.P. contributed to the investigation, validation, formal analysis, and manuscript reviewing and editing. S.A. provided visualization and supervision throughout the study. All authors reviewed and approved the final manuscript.

Data availability

Data used for the analysis is provided within the manuscript.

Declarations

Competing interests

The authors declare no competing interests.