Machine learning-driven prediction of compressive strength in sustainable geopolymer concrete using artificial neural network and random forest models

ABSTRACT

Geopolymer concrete is a sustainable substitute for ordinary Portland cement which minimizes carbon dioxide emissions and effectively utilizes the waste from industries. Proper predictive estimating compressive strength can assist in the mix deign optimization, structural reliability. The article presents a machine learning-based framework to predict the compressive strength of geopolymer concrete made with multiple industrial by-products as binders. This study investigated the subsequent strength of concrete when subjected to fly ash, ground granulated blast furnace slag, metakaolin, silica fume, and rice husk ash. A database was developed containing 243 experimentally prepared samples with different mix proportions. The study conducted the compressive strength prediction by implementing Artificial Neural Network (ANN) and Random Forest (RF) models in Python. The performance of model was analyzed through the coefficient of determination (R²) and mean absolute error (MAE) and root mean square error (RMSE). The RF model was found to be superior to the ANN model with R2 = 0.97, MAE = 1.9969, RMSE = 3.0586, which was an accurate result whereas ANN model was lower accurate R2 = 0.78. The results show that techniques using ensemble learning can capture complex non-linear relationships, reduce experimental efforts and assist in developing efficient and sustainable geopolymer concrete mix designs.

Keywords:
Geopolymer concrete; Machine learning; Artificial neural network; Random forest; Model performance.

1. INTRODUCTION

Using OPC as the major binding agent for concrete accounts for a significant amount of carbon emission. According to the estimates, about 7% of all CO₂ emissions globally are accounted for by the production of OPC. It is mainly assessed through the calcining of limestone along with burning of fossil fuel during clinker production [¹, ²]. As infrastructure demands increase, the environmental burden associated with OPC is expected to rise significantly, particularly in developing countries such as India. India is one of the world’s leading producers of cement. Due to OPC, the country’s environmental load is expected to increase rapidly. To address these challenges, GPC has emerged as a promising eco-friendly alternative [³]. Unlike regular OPC-based systems, GPC uses industrial by-products and can cut down carbon emissions by as much as 70%. Thus, GPC is ideal for sustainable construction applications.

In recent years, many Nations started using GPC for construction purposes. Australia has utilized GPC in various infrastructure projects, including precast bridge elements, road pavements and airport pavements to lessen carbon footprints associated with conventional concrete. Likewise, countries such as India, China and the USA have conducted large-scale field trials and pilot projects in which geopolymer materials were used in precast elements, building panels and transport infrastructure. These practical implementations demonstrate the growing global interest in geopolymer technology as a sustainable alternative to traditional Portland Cement-based concrete and highlight its potential for large-scale construction applications.

GPC typically consists of manufacturing residues and pozzolanic materials such as FA, GGBS, MK, RHA, and SF [⁴, ⁵]. Use of FA obtained from coal-fired power stations and slag of steelmaking is mostly industrial effluents, which avoids landfill waste and uses resources efficiently. Each ingredient of GPC has a specific function to improve its mechanical and durability properties. Class F fly ash, high in silica and alumina content but low in calcium content, is favored because it is effective under alkaline conditions [⁶,⁷,⁸]. GGBS and MK, respectively, offer higher strength and durability, whereas RHA and SF optimize the concrete’s internal structure and lower its permeability [⁹,¹⁰,¹¹].

In addition to material selection, GPC performance is affected significantly by aggregate type and curing conditions. Natural river sand is used as the fine aggregate in this study in place of factory-made sand (M-sand), and granite stone boulders are crushed to approximately 20 mm are utilized as the coarse aggregate. River sand is used for its better workability and particle grading, while crushed granite is used for its strength, durability, and good bonding with the binder, which improves the overall performance of the GPC. Predicting the CS in GPC is a complex undertaking that involves several interrelated variables, including binder composition, alkaline activator concentration, aggregate parameters, and curing conditions. Standard experimental and empirical methodologies are often time-consuming, expensive, and incapable of capturing the non-linear relationships between these parameters. ML techniques, in this context, have emerged as powerful tools for modelling complex material behavior. Data that is generated from experiments can be easily fed to ML Models. Moreover, the models discover hidden patterns and provide accurate predictions for the dataset with minimal computational and experimental effort. As a result, there has been a recent increase in efforts to apply ML in GPC for mix design optimization and improved prediction reliability. ML has been an essential tool in civil engineering in recent years, particularly for forecasting the mechanical behavior of various forms of concrete. In GPC, scientists employ ML models to predict CS without extensive physical testing, thereby improving prediction accuracy, reducing costs, and accelerating mix design optimization [¹²]. A few of the most used ML techniques used here include RF, ANN, Support Vector Machines (SVM), and Gradient Boosting [¹³]. It is observed that the performance of ensemble learning methods is better than individual.

Several recent studies have demonstrated the effectiveness of ML models in predicting the CS of GPC. There has been an increase in the use of agricultural waste materials such as sugarcane bagasse ash in geopolymer-based construction applications to promote sustainable infrastructural development. These materials have shown significant potential to enhance mechanical performance while promoting sustainability through waste utilization. A thorough review of ML techniques, such as RF, SVM, Gradient Boosting, and XGBoost, for fly ash-based GPC, reported that ensemble methods are invariably more accurate and robust than single-model predictions [¹⁴]. Recent research has focused on the durability behavior of advanced sustainable concrete systems, such as basalt fiber-reinforced recycled aggregate concrete confined with Carbon Fiber Reinforced Polymer (CFRP) under freeze–thaw conditions. The study results illustrate the necessity to examine strength and durability in designing next-generation eco-friendly construction materials [¹⁵]. Similarly, a performance benchmarking analysis of seven ML algorithms, such as MLR, ANN, SVM, KNN, Decision Trees, and ensemble learners based on boosting and bootstrapping for CS prognosis in fly ash–based GPC. They found that ensemble learners and ANNs consistently achieved higher accuracy and stability under different curing conditions [¹⁶]. Deep neural networks (DNN) and ResNet are deep learning models that have also been explored. These models achieved high prediction accuracy (R² > 0.93) while also identifying influencing factors like binder and curing conditions. Even with these advancements, nearly all works remain limited to binder combinations, narrow datasets, single models, or comparative analysis of ensemble and neural network models using various multi-source materials. However, most existing studies are limited to single binder systems, smaller datasets, or lack comprehensive comparisons between ensemble and neural network models under varying curing conditions. This limits the generalization and practical applicability of the developed models.

Traditional empirical and regression-based approaches have been extensively employed to predict the CS of concrete. Nonetheless, these approaches hardly capture the complex and nonlinear interactions among the multiple influencing parameters, such as binder composition, curing temperature and aggregate proportions. Using ML techniques can often be a better choice, as they learn from large datasets and model nonlinear relations. ANN and RF models have shown better predictive performances in concrete strength prediction, due to their abilities to deal with complex datasets and pick out major influencing features. ML strategies have become useful approaches for predicting the CS of GPC.

Despite significant advancements in ML applications for GPC, most existing studies are limited to specific binder combinations, restricted datasets, or single-model approaches. Furthermore, many studies lack comprehensive comparisons of ensemble learning techniques and neural network models across varied mix design conditions and curing environments. This limitation reduces the generalization capability and practical applicability of the developed prediction models.

The primary objective of this study is to find the most accurate and efficient ML model to predict the CS of GPC. To this end, the performance of RF and ANN models was evaluated and compared with the help of 243 experimental data. The study also involved various binder systems, mix proportions and curing temperatures (30 °C and 60 °C), introduced into the study for better generalization and predictive power of the model. The outcomes of this research advance data-driven methodologies for sustainable concrete design and provide practical insights for optimizing geopolymer mixtures in real-world applications.

2. MATERIALS AND METHODS

2.1. Materials

In this analysis, GPC was produced using various binders, namely GGBS, FA, MK, SF, and RHA, along with binary binder blends such as GGBS + FA, GGBS + MK, GGBS + RHA, and GGBS + SF. The physical and chemical properties of the above-mentioned materials determined were given in Tables 1–5. The alkaline activator consisted of sodium silicate (Na₂SiO₃) and sodium hydroxide (NaOH) solutions, with a fixed Na₂SiO₃/NaOH ratio of 2.5. River sand was replaced with manufactured sand (M-sand) as the fine aggregate, and broken granite stones were used as coarse aggregates. The fine aggregate had a maximum size of 4.75 mm, while the coarse aggregate consisted of graded fractions up to 20 mm. Water plays an equally vital role in geopolymerization during the middle and later stages of the reaction. Water helps dissolve aluminosilicate species and forms geopolymeric gels that are part of the final material but are not directly incorporated into its chemical structure. To ensure the strength and microstructural development of fly ash-based geopolymer systems, water content has to be properly controlled. The experimental program focused on M40 grade concrete. Specimens were cured for 28 days at two temperature conditions: 30 °C and 60 °C. A total of 243 concrete mixes were prepared. These mixes varied in binder type and amount, liquid-to-binder (L/B) ratio, and fine and coarse-aggregate content. This study aimed to understand how these factors affect CS. The large dataset enables the use of ML algorithms, including RF and ANN, to predict CS and find the best binder combinations for high-performance GPC [¹⁷].

Table 1
Physical properties of aluminosilicate materials.

SL. NO	PHYSICAL PARAMETERS	ALUMINOSILICATE MATERIALS
		FA	GGBS	MK	SF	RHA
1	Specific Gravity	2.35	3.2	2.30	2.202.10	3.14
2	Size (µm)	20–25	> 600	1–2	< 1	10–15
3	Colour	Light Grey	Blur White	White	white	Grey
4	Specific Surface Area (m2/kg)	350	500	14400	19600	58500

Table 2
Chemical properties of aluminosilicate materials.

SL. NO	ALUMINO- SILICATES	CHEMICAL PARAMETERS
		SiO2	Al2O3	Fe2O3	CaO	K2O	SO3	N2O	MgO	LOI
1	FA	69.35	13.55	10.75	4.25	–	–	–	–	2.10
2	GGBS	34.43	13.58	0.42	42.60	0.3	3.67	–	5	–
3	MK	52.0	46.0	0.60	0.09	0.03	–	0.10	0.03	1.00
4	SF	99.886	0.043	0.040	0.001	0.001	–	–	–	0.015
5	RHA	88.32	0.46	0.67	0.67	2.91	–	–	0.44	5.81

Table 3
Physical properties of aggregates.

SL. NO	PHYSICAL PARAMETERS	RS	BGS	BIS
1	Specific Gravity	2.65	2.75	2386-1968-part 3
2	Size (or) Gradation	II	–	383–1970
3	Moisture Content %	Nil	Nil	2386-1968-part 3
4	Fineness Modulus	2.8	7.32	383–1970
5	Crushing Value Index %	–	9.5	2386-1968-part 4
6	Water Absorption %	0.7	0.5	2386-1968-part 4

Table 4
Physical and chemical properties of sodium silicate.

SL. NO	PARAMETERS	VALUES
1	Na2O	15.5%
2	SiO2	29.5%
3	H2O	55%
4	Molecular Weight	184.04
5	Specific Gravity	1.6

Table 5
Physical and chemical properties of sodium hydroxide.

SL. NO	PARAMETERS	VALUES
1	Melting Point	318°C
2	Boiling Point	1388°C
3	Specific Gravity	1.43
4	Assay	97%
5	Na2CO3	2%
6	Cl	0.01%
7	SO2	0.05%
8	Lead (pb)	0.001%
9	Iron (Fe)	0.001%
10	Potassium (K)	0.1%
11	Zinc (Zn)	0.02%

Curing temperature significantly contributes to the strength development of GPC by affecting the rate of geopolymerization. At 60 °C, elevated temperature accelerates the dissolution of aluminosilicate species from binders like FA, MK, and RHA. This leads to quick polycondensation and the development of a more compact microstructure within the first few days of curing [¹⁸]. This effect is especially helpful for low-calcium binders such as FA and MK, which require additional thermal energy to gain substantial early strength. In contrast, high-calcium binders like GGBS and GGBS-based blends can achieve high strength even at ambient curing (30 °C) due to the combined creation of calcium silicate hydrate (C–S–H) and sodium aluminosilicate hydrate (N–A–S–H) gels. However, moderate heat curing at 60 °C can still enhance the overall CS of GGBS-rich mixes by further refining the pore structure [¹⁹]. The experimental results from the 243 mixes clearly demonstrate that optimal curing regimes vary by binder type, with GGBS-rich systems performing well at both ambient and elevated curing, whereas low-calcium binders benefit significantly from higher curing temperatures.

The compiled geopolymer concrete database used for this study is presented in Table 6. A summary of the input variables and the target variable used for the machine learning model development is presented in Table 7, which describes the dataset variables used for compressive strength prediction.

Table 6
Geopolymer concrete database.

MIX NO	BINDER		FACTORS CONSIDERED					ALKALINE LIQUIDS (kg/m3)		AGGREGATES (kg/m3)		COMPRESSIVE STRENGTH (N/mm2)
	BINDER TYPE	AMOUNT (kg/m3)	NaOH (M)	Na2SiO3/ NaOH	L/P	COARSE AGGREGATE (%)	COMBINED AGGREGATE (%)	Na2SiO3	NaOH	RS	BGS	30ᵒ	60ᵒ
M1	FA	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	15.00	26.10
M2	FA	600	10	2.5	0.6	65	60	257.14	102.86	504	936	17.50	30.50
M3	FA	600	10	2.5	0.6	60	60	257.14	102.86	576	864	21.00	39.60
M4	FA	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	15.00	26.00
M5	FA	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	17.40	30.50
M6	FA	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	22.30	36.00
M7	FA	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	13.00	22.80
M8	FA	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	15.30	26.50
M9	FA	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	17.50	30.50
M10	FA	640	10	2.5	0.5	65	60	228.57	91.429	504	936	10.00	17.40
M11	FA	640	10	2.5	0.5	60	60	228.57	91.429	576	864	11.60	20.30
M12	FA	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	13.30	23.20
M13	FA	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	12.50	21.70
M14	FA	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	14.50	25.40
M15	FA	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	15.00	28.30
M16	FA	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	10.90	19.00
M17	FA	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	12.70	22.30
M18	FA	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	14.50	25.40
M19	FA	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	6.70	14.60
M20	FA	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	7.80	17.00
M21	FA	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	9.00	19.40
M22	FA	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	7.50	16.20
M23	FA	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	8.70	19.00
M24	FA	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	10.00	21.30
M25	FA	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	6.50	14.20
M26	FA	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	7.50	16.50
M27	FA	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	8.70	18.90
M28	GGBS	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	61.50	68.70
M29	GGBS	600	10	2.5	0.6	65	60	257.14	102.86	504	936	71.70	80.20
M30	GGBS	600	10	2.5	0.6	60	60	257.14	102.86	576	864	82.00	91.60
M31	GGBS	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	65.50	73.20
M32	GGBS	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	76.50	85.50
M33	GGBS	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	87.40	97.70
M34	GGBS	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	55.70	62.30
M35	GGBS	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	65.00	72.70
M36	GGBS	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	74.30	83.10
M37	GGBS	640	10	2.5	0.5	65	60	228.57	91.429	504	936	48.70	54.50
M38	GGBS	640	10	2.5	0.5	60	60	228.57	91.429	576	864	56.80	63.60
M39	GGBS	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	65.00	72.70
M40	GGBS	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	52.50	58.00
M41	GGBS	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	61.20	68.50
M42	GGBS	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	70.00	78.50
M43	GGBS	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	46.00	51.40
M44	GGBS	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	53.60	60.00
M45	GGBS	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	61.00	68.50
M46	GGBS	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	37.60	43.62
M47	GGBS	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	43.80	50.80
M48	GGBS	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	50.00	58.10
M49	GGBS	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	40.70	47.00
M50	GGBS	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	47.50	55.00
M51	GGBS	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	55.00	61.40
M52	GGBS	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	35.60	41.40
M53	GGBS	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	41.50	48.20
M54	GGBS	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	47.50	55.10
M55	MK	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	11.50	23.30
M56	MK	600	10	2.5	0.6	65	60	257.14	102.86	504	936	12.90	27.20
M57	MK	600	10	2.5	0.6	60	60	257.14	102.86	576	864	14.80	31.00
M58	MK	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	9.50	24.00
M59	MK	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	11.10	27.30
M60	MK	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	15.50	32.40
M61	MK	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	8.30	20.40
M62	MK	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	9.70	22.50
M63	MK	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	11.10	26.50
M64	MK	640	10	2.5	0.5	65	60	228.57	91.429	504	936	8.30	15.50
M65	MK	640	10	2.5	0.5	60	60	228.57	91.429	576	864	9.70	18.00
M66	MK	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	11.10	20.00
M67	MK	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	8.70	16.00
M68	MK	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	8.70	20.50
M69	MK	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	8.70	23.40
M70	MK	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	8.70	15.30
M71	MK	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	8.70	17.80
M72	MK	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	8.70	19.40
M73	MK	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	8.70	6.70
M74	MK	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	8.70	7.80
M75	MK	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	8.70	9.00
M76	MK	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	8.70	8.80
M77	MK	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	8.70	12.50
M78	MK	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	8.70	16.00
M 79	MK	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	8.70	6.50
M80	MK	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	8.70	7.60
M81	MK	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	8.70	8.70
M82	SF	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	0.00	0.00
M83	SF	600	10	2.5	0.6	65	60	257.14	102.86	504	936	0.00	0.00
M84	SF	600	10	2.5	0.6	60	60	257.14	102.86	576	864	0.00	0.00
M85	SF	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	0.00	0.00
M86	SF	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	0.00	0.00
M87	SF	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	0.00	0.00
M88	SF	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	0.00	0.00
M89	SF	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	0.00	0.00
M90	SF	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	0.00	0.00
M91	SF	640	10	2.5	0.5	65	60	228.57	91.429	504	936	0.00	0.00
M92	SF	640	10	2.5	0.5	60	60	228.57	91.429	576	864	0.00	0.00
M93	SF	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	0.00	0.00
M94	SF	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	0.00	0.00
M95	SF	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	0.00	0.00
M96	SF	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	0.00	0.00
M97	SF	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	0.00	0.00
M98	SF	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	0.00	0.00
M99	SF	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	0.00	0.00
M100	SF	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	0.00	0.00
M101	SF	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	0.00	0.00
M102	SF	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	0.00	0.00
M103	SF	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	0.00	0.00
M104	SF	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	0.00	0.00
M105	SF	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	0.00	0.00
M106	SF	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	0.00	0.00
M107	SF	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	0.00	0.00
M108	SF	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	0.00	0.00
M109	RHA	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	4.00	11.00
M110	RHA	600	10	2.5	0.6	65	60	257.14	102.86	504	936	4.60	12.80
M111	RHA	600	10	2.5	0.6	60	60	257.14	102.86	576	864	5.30	14.60
M112	RHA	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	4.00	11.20
M113	RHA	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	4.70	13.00
M114	RHA	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	6.30	16.00
M115	RHA	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	3.50	9.80
M116	RHA	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	4.10	11.40
M117	RHA	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	4.70	13.00
M118	RHA	640	10	2.5	0.5	65	60	228.57	91.429	504	936	3.10	8.50
M119	RHA	640	10	2.5	0.5	60	60	228.57	91.429	576	864	3.60	10.10
M120	RHA	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	4.50	11.50
M121	RHA	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	3.40	9.30
M122	RHA	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	3.90	10.91
M123	RHA	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	4.30	12.30
M124	RHA	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	3.00	8.20
M125	RHA	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	3.40	9.50
M126	RHA	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	4.00	10.90
M127	RHA	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	2.30	5.90
M128	RHA	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	2.50	6.90
M129	RHA	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	2.80	7.90
M130	RHA	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	2.70	7.50
M131	RHA	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	3.10	8.70
M132	RHA	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	3.60	9.60
M133	RHA	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	2.33	6.50
M134	RHA	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	2.50	7.60
M135	RHA	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	3.00	8.30
M136	FA + GGBS	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	48.90	59.70
M137	FA + GGBS	600	10	2.5	0.6	65	60	257.14	102.86	504	936	57.10	69.70
M138	FA + GGBS	600	10	2.5	0.6	60	60	257.14	102.86	576	864	65.30	79.67
M139	FA + GGBS	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	49.80	70.70
M140	FA + GGBS	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	58.10	82.50
M141	FA + GGBS	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	66.40	81.00
M142	FA + GGBS	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	37.40	53.00
M143	FA + GGBS	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	42.80	60.80
M144	FA + GGBS	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	49.00	69.50
M145	FA + GGBS	640	10	2.5	0.5	65	60	228.57	91.429	504	936	34.50	43.20
M146	FA + GGBS	640	10	2.5	0.5	60	60	228.57	91.429	576	864	40.30	50.40
M147	FA + GGBS	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	46.00	57.60
M148	FA + GGBS	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	35.00	50.00
M149	FA + GGBS	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	40.80	57.90
M150	FA + GGBS	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	50.00	61.60
M151	FA + GGBS	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	30.50	43.40
M152	FA + GGBS	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	35.50	50.70
M153	FA + GGBS	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	40.80	57.00
M154	FA + GGBS	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	25.50	30.70
M155	FA + GGBS	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	29.90	35.90
M156	FA + GGBS	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	34.20	41.00
M157	FA + GGBS	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	28.00	39.70
M158	FA + GGBS	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	32.30	41.30
M159	FA + GGBS	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	35.00	43.50
M160	FA + GGBS	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	24.50	34.60
M161	FA + GGBS	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	28.50	40.50
M162	FA + GGBS	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	32.00	43.00
M163	MK + GGBS	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	33.90	41.10
M164	MK + GGBS	600	10	2.5	0.6	65	60	257.14	102.86	504	936	39.60	47.90
M165	MK + GGBS	600	10	2.5	0.6	60	60	257.14	102.86	576	864	45.30	54.80
M166	MK + GGBS	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	34.20	41.40
M167	MK + GGBS	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	40.00	48.30
M168	MK + GGBS	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	46.30	57.40
M169	MK + GGBS	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	30.00	36.30
M170	MK + GGBS	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	35.00	42.30
M171	MK + GGBS	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	40.00	48.30
M172	MK + GGBS	640	10	2.5	0.5	65	60	228.57	91.429	504	936	25.80	31.20
M173	MK + GGBS	640	10	2.5	0.5	60	60	228.57	91.429	576	864	30.10	36.40
M174	MK + GGBS	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	34.40	41.60
M175	MK + GGBS	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	28.50	34.50
M176	MK + GGBS	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	33.30	40.30
M177	MK + GGBS	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	35.00	43.50
M178	MK + GGBS	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	25.00	30.20
M179	MK + GGBS	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	29.00	35.30
M180	MK + GGBS	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	33.30	40.30
M181	MK + GGBS	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	16.80	21.30
M182	MK + GGBS	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	19.60	24.90
M183	MK + GGBS	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	22.50	28.50
M184	MK + GGBS	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	18.50	23.70
M185	MK + GGBS	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	22.00	27.70
M186	MK + GGBS	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	25.00	31.50
M187	MK + GGBS	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	16.40	20.70
M188	MK + GGBS	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	19.00	24.20
M189	MK + GGBS	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	21.50	27.70
M190	SF + GGBS	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	0.00	27.48
M191	SF + GGBS	600	10	2.5	0.6	65	60	257.14	102.86	504	936	0.00	32.08
M192	SF + GGBS	600	10	2.5	0.6	60	60	257.14	102.86	576	864	0.00	36.64
M193	SF + GGBS	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	0.00	29.28
M194	SF + GGBS	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	0.00	34.20
M195	SF + GGBS	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	0.00	39.08
M196	SF + GGBS	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	0.00	24.90
M197	SF + GGBS	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	0.00	29.08
M198	SF + GGBS	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	0.00	33.24
M199	SF + GGBS	640	10	2.5	0.5	65	60	228.57	91.429	504	936	0.00	21.80
M200	SF + GGBS	640	10	2.5	0.5	60	60	228.57	91.429	576	864	0.00	25.44
M201	SF + GGBS	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	0.00	29.08
M202	SF + GGBS	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	0.00	23.20
M203	SF + GGBS	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	0.00	27.40
M204	SF + GGBS	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	0.00	31.40
M205	SF + GGBS	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	0.00	20.56
M206	SF + GGBS	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	0.00	24.00
M207	SF + GGBS	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	0.00	27.40
M208	SF + GGBS	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	0.00	17.45
M209	SF + GGBS	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	0.00	20.32
M210	SF + GGBS	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	0.00	23.24
M211	SF + GGBS	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	0.00	18.80
M212	SF + GGBS	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	0.00	22.00
M213	SF + GGBS	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	0.00	24.56
M214	SF + GGBS	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	0.00	16.56
M215	SF + GGBS	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	0.00	19.28
M216	SF + GGBS	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	0.00	22.04
M217	RHA + GGBS	600	10	2.5	0.6	70	60	257.14	102.86	432	1008	19.30	22.40
M218	RHA + GGBS	600	10	2.5	0.6	65	60	257.14	102.86	504	936	22.50	26.10
M219	RHA + GGBS	600	10	2.5	0.6	60	60	257.14	102.86	576	864	25.80	29.80
M220	RHA + GGBS	450	10	2.5	0.6	70	70	192.86	77.143	504	1176	16.50	22.50
M221	RHA + GGBS	450	10	2.5	0.6	65	70	192.86	77.143	588	1092	19.30	26.20
M222	RHA + GGBS	450	10	2.5	0.6	60	70	192.86	77.143	672	1008	25.50	30.20
M223	RHA + GGBS	300	10	2.5	0.6	70	80	128.57	51.429	576	1344	14.50	19.70
M224	RHA + GGBS	300	10	2.5	0.6	65	80	128.57	51.429	672	1248	16.90	23.00
M225	RHA + GGBS	300	10	2.5	0.6	60	80	128.57	51.429	768	1152	19.50	26.30
M226	RHA + GGBS	640	10	2.5	0.5	65	60	228.57	91.429	504	936	13.60	15.50
M227	RHA + GGBS	640	10	2.5	0.5	60	60	228.57	91.429	576	864	16.00	18.40
M228	RHA + GGBS	640	10	2.5	0.5	70	60	228.57	91.429	432	1008	18.20	21.00
M229	RHA + GGBS	480	10	2.5	0.5	65	70	171.43	68.571	588	1092	13.80	18.50
M230	RHA + GGBS	480	10	2.5	0.5	60	70	171.43	68.571	672	1008	16.10	21.90
M231	RHA + GGBS	480	10	2.5	0.5	70	70	171.43	68.571	504	1176	19.20	23.30
M232	RHA + GGBS	320	10	2.5	0.5	65	80	114.29	45.714	672	1248	12.00	16.40
M233	RHA + GGBS	320	10	2.5	0.5	60	80	114.29	45.714	768	1152	14.00	19.20
M234	RHA + GGBS	320	10	2.5	0.5	70	80	114.29	45.714	576	1344	16.10	21.90
M235	RHA + GGBS	685.71	10	2.5	0.4	60	60	195.92	78.367	576	864	10.00	10.50
M236	RHA + GGBS	685.71	10	2.5	0.4	70	60	195.92	78.367	432	1008	11.60	12.30
M237	RHA + GGBS	685.71	10	2.5	0.4	65	60	195.92	78.367	504	936	13.30	14.00
M238	RHA + GGBS	514.29	10	2.5	0.4	60	70	146.94	58.776	672	1008	10.20	12.50
M239	RHA + GGBS	514.29	10	2.5	0.4	70	70	146.94	58.776	504	1176	12.00	14.50
M240	RHA + GGBS	514.29	10	2.5	0.4	65	70	146.94	58.776	588	1092	13.70	16.70
M241	RHA + GGBS	342.86	10	2.5	0.4	60	80	97.959	39.184	768	1152	9.70	11.20
M242	RHA + GGBS	342.86	10	2.5	0.4	70	80	97.959	39.184	576	1344	11.30	13.00
M243	RHA + GGBS	342.86	10	2.5	0.4	65	80	97.959	39.184	672	1248	12.90	14.80

Table 7
Description of dataset variables used for compressive strength prediction.

SL. NO	VARIABLE TYPE	PARAMETER	DESCRIPTION	UNIT
1	Input	Fly Ash	Class F fly ash used as binder	kg/m3
2	Input	Ground Granulated Blast Furnace Slag	Supplementary cementitious material	kg/m3
3	Input	Metakaolin	Aluminosilicate binder material	kg/m3
4	Input	Rice Husk Ash	Agricultural waste binder material	kg/m3
5	Input	Silica Fume	Highly reactive pozzolanic material	kg/m3
6	Input	Liquid-to-Binder Ratio	Ratio of alkaline solution to binder	–
7	Input	NaOH Molarity	Sodium hydroxide concentration	M
8	Input	Na2SiO3/NaOH Ratio	Ratio of sodium silicate to sodium hydroxide	–
9	Input	River Sand	Fine aggregate	kg/m3
10	Input	Coarse Aggregate	Broken granite stone	kg/m3
11	Input	Curing Temperature	Temperature used for curing	0C
12	Output	Compressive Strength	Target variable	N/mm2

2.2. Methodology

The GPC CS forecasting in this research is conducted using a structured and systematic ML framework designed to enhance both accuracy and reliability. The dataset used in this research consists of 243 experimentally generated GPC mix proportions (Table 6), obtained through controlled laboratory investigations. Such experimentally derived datasets are considered more reliable because they capture realistic material behavior under practical conditions. All experimental samples were prepared according standardized mixing, curing, and testing procedures to ensure consistency and repeatability of results. The mix design strategy was developed to systematically vary binder compositions and curing conditions, thereby capturing a broad spectrum of material behavior. The dataset encompasses a wide range of input variables that influence the mechanical performance of GPC.

Input parameters were selected due to their known effects on the CS of GPC. As the binder materials control the geopolymerization process and directly influence strength development, FA, GGBS, MK, RHA and SF were included. The liquid-to-binder ratio and alkaline activator parameters (NaOH molarity and Na₂SiO₃/NaOH ratio) were selected for their significant influence on the reaction kinetics and microstructural formation. Aggregate characteristics, including river sand and coarse aggregate, were considered because they influence workability and load transfer behavior. In addition, curing temperature was incorporated as a critical parameter affecting the rate of geopolymerization and final CS. The selection of these parameters is supported by previous studies, which identify binder composition, activator concentration, and curing conditions as key factors governing GPC performance.

The mix parameters include different binder combinations such as Class F Fly ash, GGBS, MK, RHA, SF, GGBS + FA, GGBS + MK, GGBS + RHA, GGBS + SF, the liquid-to-binder ratio, and alkaline activator characteristics such as NaOH molarity and the Na₂SiO₃/NaOH ratio. Furthermore, river sand was used as fine aggregate while 20 mm BGS was used as coarse aggregate and specimens were cured at 30°C and 60°C. The target variable for model development is the quantified CS obtained from these experimental mixes. Before model development, the dataset underwent validation procedures to ensure quality and consistency. The data were checked for missing values, inconsistencies, and recording errors. Outlier analysis was performed using statistical methods, and no significant abnormal values were identified; therefore, all experimental data points were retained for further analysis. This helps avoid bias in the dataset caused by too much data filtering and reflects actual experimental behavior. In addition, although explicit feature selection algorithms were not employed, the input parameters were carefully chosen based on their physical significance and their established influence on GPC strength, ensuring an effective and meaningful feature set for model development.

Consistent with standard practices in ML-based concrete studies, the sequence of operational stages considered in the model analysis is illustrated in Figure 1. Before modelling, the raw data undergoes a preprocessing phase to ensure it is reliable and consistent. This stage starts with data cleaning to address missing or inconsistent values and with normalizing continuous variables to a common scale. Categorical data, such as different binders, are encoded so that the ML algorithms can properly interpret them. The evidence dataset was split into a training and testing dataset before building the model and checking its predictive power on test dataset.

Figure 1
Flowchart for research methodology.

Consistent with standard practices in ML-based concrete studies, the dataset undergoes a preprocessing stage before model development to ensure data quality and consistency. The preprocessing procedure entails data cleaning to rectify missing or inconsistent values, normalizing continuous variables to the 0–1 range for computational efficiency, and encoding categorical variables like binder combinations. To allow fair performance evaluation, the processed dataset has been divided randomly into a 70% training and 30% testing dataset. The testing dataset, which was not used during training, was used to assess predictive performance and ensure that the models do not overfit the training data. A dataset of 243 samples was randomly split into training and testing subsets. In this work, 70% of samples (170) were used for model training, while 30% (73) were used for performance evaluation. To minimize bias in the data, random partitioning was used so that both datasets would equally reflect the entirety of data.

Since the experimentally generated data were not highly noisy and did not have sharp fluctuations, data smoothing procedures were not undertaken. The retention of every data point served to preserve dataset integrity and ensure that the ML models were trained on realistic material behaviours. Therefore, the model’s performance metrics indicate reliable, unbiased predictive accuracy.

All ML models were implemented in Python (version 3.10). The data preprocessing includes using libraries NumPy and Pandas for numerical calculation. The various ML algorithms like the RF were created using Scikit-learn Library. The ANN model is created using TensorFlow and Keras. Matplotlib and Seaborn were used for graphics and data visualization that represent the results. The use of highly popular and open-source libraries ensures transparent and reproducible ML experiments.

Two predictive models are constructed for this study: the RF model, which builds a collection of decision trees and averages their predictions to gain more accuracy, and the ANN, which is highly effective at modelling intricate, nonlinear relationships between input variables and CS [²⁰]. Both models are trained on the cleaned dataset with their parameters set to optimize learning outcomes. The selection of the RF and AN model in this study is based on their proven effectiveness in modelling complex nonlinear relationships in concrete technology, as reported in recent literature. RF was chosen for its robustness, ability to handle high-dimensional data, and resistance to overfitting through ensemble learning. ANN was selected for its capability to capture nonlinear interactions between input parameters and CS. The use of these two models also enables a comparative analysis between an ensemble learning approach and a single predictive model. The model parameters were selected based on standard practices and iterative tuning to achieve optimal performance. To minimize the risk of overfitting, appropriate model parameters were selected during hyperparameter tuning. In the RF model, using multiple decision trees and random feature selection helps reduce overfitting and improve generalization. Similarly, for the ANN model, normalization and controlled network architecture were used to prevent excessive model complexity.

Hyperparameter tuning was conducted on both RF and ANN models for better performance. To improve prediction accuracy, we altered the number of trees, maximum depth of tree and minimum number of samples required to split RF model. To achieve optimal performance, the ANN model was tuned by varying the number of hidden layers, neuron count, learning rate, and number of training epochs. All combinations under consideration were evaluated using the R², RMSE and MAE metrics obtained on the testing dataset. The optimization process guarantees reliable and generalized predictive performance for the developed models.

Model performance is evaluated using the coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE) [²¹]. These metrics help in quantifying the extent to which predicted values are consistent with experimental observations. A comparison of RF and ANN reveals the capabilities and constraints of each model, allowing the best-performing model to be chosen as the final prediction tool. After verification, this selected model is applied to predict the CS of new GPC mixes, providing a faster, data-driven solution in place of lengthy, traditional lab tests. Not only does this approach speed up the mix design cycle, but it also encourages green building practices by enabling researchers and engineers to refine GPC formulations with improved accuracy and confidence.

Figure 2 presents the frequency distribution of the relative frequencies of various binder and activator materials employed in GPC, i.e., FA, GGBS, MK, Na₂SiO₃, NaOH, RHA, and SF. The histograms of FA, GGBS, MK, RHA, and SF all show a similar trend where the relative frequency is highest at the lower dosage level of approximately 300 kg, decreasing as the dosage increases to 700 kg. This can be interpreted as lower levels of these pozzolanic materials being used more often in the dataset. However, the alkaline activators follow a contrasting trend. Sodium silicate (Na₂SiO₃), between 100 and 250 kg, exhibits a nearly flat distribution with slightly elevated frequencies at 150–200 kg, which creates a weak bell-shaped curve. Sodium hydroxide (NaOH), between 40 and 100 kg, also shows a distinct bell-shaped distribution with peak counts around 60–70 kg, indicating that intermediate dosages are most frequently occurring in use. Generally, the frequency distribution analysis shows a tendency to use lower amounts of supplementary binders and a moderate level of alkaline activators, which aligns with common practice mix design preferences for maintaining strength and workability in GPC.

Figure 2
Relative frequency distribution of the input variables.

2.3. Multicollinearity analysis of input features

Before model development, the input variables were examined for multicollinearity to ensure the robustness of the machine learning model. Pearson correlation analysis was performed to evaluate the linear relationship among the input features. Additionally, the Variance Inflation Factor (VIF) was calculated to quantify multicollinearity among predictors. A VIF value below 5 indicates that multicollinearity is not significant and the features can be retained for modelling. The results indicated that all selected input variables have acceptable correlation values and the VIF values are within the acceptable limit, showing that multicollinearity has no significant effect on the predictive model.

VIF multicollinearity analysis was done to ensure the reliability of the ML model and predictions. Multicollinearity occurs when two or more predictor variables are highly correlated, which can negatively affect the stability and interpretability of predictive models. Generally, a VIF value greater than 10 indicates severe multicollinearity, while values below 5 suggest low multicollinearity. The VIF results for the input features are presented in Figure 3. Among the evaluated variables, water shows the highest VIF value of 7.8, followed by aggregate with a VIF of around 7.4, and cement with a VIF value of 6.4. Although these values indicate a moderate level of correlation among the mixture components, they remain below the critical threshold of 10, suggesting that multicollinearity does not significantly affect the model performance.

Figure 3
VIF analysis of input features used in the ML model to assess multicollinearity.

The age parameter exhibits the lowest VIF value of 3.5, indicating a relatively weak correlation with the other input variables. This suggests that curing age behaves as an independent predictor in the dataset. The overall VIF examines the input features selected, and since they are acceptable, they will be retained for the ML model development. This analysis demonstrates the robustness of the feature selection process and ensures that the model’s predictive performance is not adversely affected by multicollinearity among the input variables.

2.4. Machine learning algorithms

ML algorithms are now becoming a very efficient means of solving sophisticated engineering problems, especially when conventional statistical approaches fail to identify highly complex inter-variable relations. In concrete technology, machine learning enables the analysis of large experimental databases, uncovering hidden patterns and making accurate predictions without the need for extensive trial-and-error experiments. We make use of RF and ANN to forecast the CS of GPC in this research, considering variables such as binder types, curing temperatures, and mix design parameters. These algorithms not only help identify the factors that contribute most to strength but also support the upgrading of mix designs, leading to better-performing GPC [²²].

2.4.1. Random forest regression

RFR is an efficient ML technique for modelling output responses as a function of input variables. RF is an ensemble model that combines multiple decision trees to improve prediction accuracy and reliability [²³]. The RF model significantly reduces overfitting, and that is its greatest strength. Likewise, it can handle nonlinear relationships between input parameters. This makes it an appropriate choice for material engineering applications, given the CS multiple factors affecting our factors. In this study, RFR is used to capture the combined effect of binder composition, curing conditions and other input variables. The RF model’s output is the average of the predictions of the individual RF trees. This improves the reliability of the RF prediction. The regression prediction of the RF is expressed in Eqn (1) [²⁴]:

(1) $$\hat{y}=\frac{1}{T}\sum _{t=1}^T{h}_t\left(x\right)$$

2.4.2. Artificial neural network

ANN consist of an input layer, one or more hidden layers, and an output layer. They were inspired by the human brain and are interconnected computational models of the same [²⁵, ²⁶]. Through weights in the different layers of the model, the relationship between input parameters and the output response will be captured. The present research treats the experimental variables as input, whereas the network evaluates the CS by computing in hidden layers. The weighted sum of inputs for the j^th neuron is calculated using Eqn (2) [²⁷]:

(2) $${\left(\text{net}\right)}_j={\displaystyle \sum }_{\text{i}=1}^{\text{n}}{\text{w}}_{\text{ij}}{\text{x}}_{\text{i}}+\text{b}$$

where w_ij represents the weight connecting the i^th neuron in the previous layer to the j^th neuron in the current layer, x_i denotes the output from the i^th neuron of the preceding layer, b is the bias term, and n is the total number of inputs. This value is then passed through a nonlinear activation function, such as the sigmoid function, to generate the output, as given in Eqn (3) [²⁸]. This process enables the ANN model to capture complex nonlinear relationships between mix design parameters and CS.

(3) $${\left(\text{out}\right)}_j=\text{f}\left({\left(\text{net}\right)}_j\right)=\frac{1}{1+{\text{e}}^{-\text{a}{\left(\text{net}\right)}_j}}$$

where, a is a constant controlling the slope of the activation function in its nonlinear region.

For predicting the CS of GPC, an ANN model was developed using Python and trained through a backpropagation learning algorithm. Figure 4 illustrates the architecture of the proposed ANN model, comprising an input layer, hidden layers, and an output layer for CS prediction. The dataset was randomly partitioned into training and testing subsets, with 70% of the data used for training and the remaining 30% reserved for model evaluation. Prior to training, all input and output variables were normalized to the range 0 to 1 to improve model performance. The network architecture consists of eight input parameters, two hidden layers with 96 neurons, and one output neuron representing the 28-day CS.

Figure 4
Artificial neural network architecture.

2.4.3. Model performance

Model performance in this study is assessed using standard regression metrics like coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE) [²⁹] to quantify and compare the predictive exactness of RFR and ANN models for GPC CS. The R² value indicates the fraction of variability in the actual data accounted by the model, with values near 1 signifying a stronger fit, while MAE, MSE, and RMSE evaluate the average degree of prediction inaccuracies: MAE provides the average absolute deviation, MSE penalizes larger errors by squaring deviations, and RMSE converts the error to the original units, facilitating physical interpretation [³⁰, ³¹]. Lower values of MAE, MSE, and RMSE indicate better exactness predictions. Both RFR and ANN can capture nonlinear relationships. Still, their comparative performance is objectively judged through these metrics, enabling identification of which algorithm yields closer agreement with experimental CS data under the given mix design and curing conditions. The metrics are defined in Eqns (4–7) [³²]:

(4) $$R^2=1-\frac{{{\displaystyle \sum }}^{\text{}}{\left(x_1-y_1\right)}^2}{{{\displaystyle \sum }}^{\text{}}{(y_1-{\overline{y}}_1)}^2}$$

(5) $$\text{MAE}=\frac{1}{\text{n}}{\displaystyle \sum }_{i=1}^n|x_i-x|$$

(6) $$MSE=\frac{1}{n}{\displaystyle \sum }_{i=1}^n(y_{pred}-y_{ref}){}^2$$

(7) $$RMSE=\sqrt{{{\displaystyle \sum }}^{\text{}}\frac{{(y_{pred}-y_{ref})}^2}{n}}$$

where n = total number of data samples, x, y_ref = reference values in the data sample, x_i, y_pred = predicted values from models.

5. RESULTS AND DISCUSSION

5.1. Results for compressive strength

Figure 5 shows the influence of curing temperature on the CS of GPC. The results indicate that specimens cured at 60 °C exhibit higher CS compared to those cured at 30 °C. Elevated curing temperature accelerates the geopolymerization process, resulting in improved strength development.

Figure 5
Effect of curing temperature (30 °C and 60 °C) on the average CS of GPC.

5.2. Results for random forest regression

The RF model exhibited superior forecasting performance for GPC strength (Table 8), achieving an R² of 0.979, an MAE of 1.997 MPa, an MSE of 9.355 MPa², and an RMSE of 3.059 MPa. Table 9 and Figure 6 present a comparison of actual and predicted strengths across various binder types—including FA, GGBS, MK, RHA, SF, and their GGBS combinations—showing percentage errors within ±7.15%, indicating strong model generalization. These metrics indicate that the model simulates nearly 98% of the variability in CS with minimal prediction error (Figure 7), validating its reliability and robustness [³³, ³⁴].

Table 8
Statistical data of training phase prediction models (compressive strength) with available RF and ANN models.

MODEL	TESTING RESULTS (COMPRESSIVE STRENGTH)
	R2	MAE	MSE	RMSE	95% CONFIDENCE INTERVAL
RF	0.979021	1.996885	9.355274	3.058639	±0.70
ANN	0.785439	7.555548	95.67903	9.781566	±2.24

Table 9
Cross-validation performance of RF and ANN models.

MODEL	TESTING RESULTS (COMPRESSIVE STRENGTH)
	R2	MAE	MSE	RMSE	95% CONFIDENCE INTERVAL
RF	0.967	2.05	8.1	3.10	±0.71
ANN	0.783	7.60	95.2	9.50	±2.18

Figure 6
Relationship between actual and predicted results obtained from the RF model.

Figure 7
Error distribution of the actual and predicted results obtained from the RF model.

The highest accuracy was observed for SF (0.00% error) and RHA+GGBS (+1.26%), while the largest deviation was for FA+GGBS (−7.15%). Despite the complexity of geopolymer systems, the RF model performed well across various binder combinations. This strong performance can be attributed to the ensemble nature of the RF algorithm, which combines multiple decision trees to capture nonlinear relationships and effectively reduce overfitting. These results demonstrate its ability to accurately predict CS, confirming its value in optimizing mix designs for GPC through ensemble learning methods [³⁵, ³⁶]. Compared to the ANN model, RF demonstrates more stable and reliable predictions for the given dataset. ABDELLATIEF et al. [¹⁷] found that XGBoost slightly outperforms RF, delivering strong predictive accuracy (R² ≈ 0.887), better generalization than conventional regression approaches, and effective feature importance analysis, underlining the robust performance of ensemble models in complex concrete systems.

We have performed a 95% confidence interval (CI) analysis using the prediction errors of the test dataset to further validate the predictive performance of our models. The RF model is less confident than the ANN model as it has a narrower confidence interval. Specifically, the RF model has a confidence interval of ±0.70. Meanwhile, the ANN model has a confidence interval of ±2.24. The RF model’s smaller CI signifies that it is likely to produce more stable predictions. According to this statistical analysis, it is likely that the RF model is generally a better predictive model than the ANN model for estimating the CS of GPC.

The cross-validation results demonstrate that the RF model maintains consistent performance with low fold-to-fold variability, indicating strong generalization capability. In contrast, the ANN model shows greater variation, suggesting lower prediction stability, as represented in Table 10. To better understand how the input variables affect the RF model’s CS prediction of concrete, a feature importance analysis was conducted. Based on the findings, the most critical parameters affecting CS are binder composition, alkaline activator concentration, and curing temperature. This analysis reveals how much each variable is contributing to the outcomes and improves subsequent interpretation. To assess the reliability of the predictions, uncertainty analysis was performed based on the distribution of prediction errors. The standard deviation of the residuals was used as an indicator of prediction uncertainty. The RF Model with relatively low RMSE shows that the predicted values were closely grouped around the experimental results. Hence, it indicates low uncertainty and a highly reliable prediction. This reinforces the model’s prediction that GPC strength is stable and robust.

Table 10
Actual vs predicted CS for high-performance binder combinations.

SL. NO	BINDER TYPE	ACTUAL STRENGTH	PREDICTED STRENGTH
1	Fly Ash	39.6	37.41
2	GGBS	97.7	93.54
3	Metakaolin	32.4	34.62
4	Silica Fume	0	0
5	Rice Husk Ash	16	15.21
6	GGBS + Fly Ash	82.5	76.60
7	GGBS + Metakaolin	57.4	54.35
8	GGBS + Silica Fume	39.08	37.53
9	GGBS + Rice Husk Ash	30.2	30.58

Even though RF model has better predictive accuracy for predicting the CS of GPC, it has limitations. RF and their hyperparameters are essential to avoid overfitting and to make the predictions better. The computational expense may also be increased with larger data sets or more input features. In addition, the model’s performance may depend on tuning the hyperparameters. As a result, it is necessary to validate the model thoroughly and optimize its parameters.

5.3. Results for artificial neural network

The ANN model used here was trained and tuned carefully to fine-tune its parameters. The model yielded an R² of 0.785, an MAE of 7.56 MPa, an MSE of 95.68 MPa², and an RMSE of 9.78 MPa. All the above values indicate a moderate level of accuracy, indicating that the model’s forecast is reasonably close to the actual measurements of CS. While the ANN can provide reasonable estimates, its accuracy is not sufficient for high-strength GPC mixtures [³⁷,³⁸,³⁹]. The model often underestimates CS, especially in higher-strength specimens, suggesting limitations in capturing the complex, nonlinear interactions in binder behaviour.

A thorough comparison of measured and predicted strengths for various binder compositions is presented in Table 11 and Figure 8, supporting this observation. The model underestimated the strength by 33.2% for GGBS, 26.1% for FA–GGBS, and 36.2% for MK. The largest deviation occurred with the SF–GGBS mix, where the prediction fell short by 47.3%. In contrast, the RHA GGBS mix had the smallest error, at 9.7% (Figure 9). Interestingly, the only overprediction happened with the SF specimen, where the model estimated 2.33 MPa, where no strength was recorded. The ANN model’s comparatively lower performance can be attributed to its sensitivity to dataset size and model parameters. ANN models typically require larger datasets and careful hyperparameter tuning to learn complex nonlinear relationships effectively. In the present study, the dataset size and variability in binder composition may have limited the model’s ability to generalize accurately. KINA et al. [⁴⁰] compared several ML models, including ANN, Bagging, least-Squares Boosting (LSBoost), and others, for predicting the CS of FA/GGBS-based or GPC. While LSTM achieved the highest R² (~0.98), the ANN still showed good but slightly lower accuracy (R² ≈ 0.94–0.95) compared to the best deep learning model. This supports the observation that ANN models may underperform ensemble or deep architectures in complex datasets.

Table 11
Actual vs predicted compressive strength for high-performance binder combinations.

SL. NO	BINDER TYPE	ACTUAL STRENGTH	PREDICTED STRENGTH
1	Fly Ash	39.6	31.81
2	GGBS	97.7	65.26
3	Metakaolin	32.4	20.68
4	Silica Fume	0	2.33
5	Rice Husk Ash	16	13.99
6	GGBS + Fly Ash	82.5	60.98
7	GGBS + Metakaolin	57.4	45.37
8	GGBS + Silica Fume	39.08	20.58
9	GGBS + Rice Husk Ash	30.2	27.28

Figure 8
Relationship between actual and predicted results obtained from the ANN model.

Figure 9
Error distribution of the actual and predicted results obtained from the ANN model.

For the given dataset, the ANN model exhibits reduced predictive capability and stability when compared with the RF model. This evidence suggests that while the ANN model provides a reasonable estimate for moderate-strength mixes, it struggles to capture the strength development of high-performance binder systems reliably. Further model refinement or the use of hybrid methods may be necessary to enhance its predictive ability [²²].

The developed ML framework presents valuable practical implications for construction engineering and industry applications. Traditional experimental methods for determining CS are time-consuming and require extensive laboratory testing, material preparation, and curing periods. In contrast, the developed RF model enables rapid, high-accuracy prediction of CS, thereby reducing the need for repetitive experimental trials. This results in considerable time savings during mix design optimization.

Furthermore, predictive modelling helps minimize material waste and reduce overall testing costs by limiting trial-and-error. Engineers and practitioners can use the developed model to design GPCs to achieve desired strength properties efficiently, improving decision-making in real-world applications. Additionally, the ability to optimize the use of industrial by-products such as FA, GGBS, and RHA contributes to sustainable construction practices by reducing environmental impact and promoting resource efficiency.

The superior performance of the RF model can be attributed to its ensemble structure, which reduces overfitting and improves generalization. In contrast, the ANN model’s performance is influenced by its sensitivity to dataset size and parameter tuning, which may limit its predictive performance on complex and variable datasets.

A more detailed comparison between the RF and ANN models reveals significant differences in their predictive capabilities, particularly for high-strength GPC mixes. While the RF model demonstrates consistent accuracy across all binder combinations, the ANN model shows noticeable limitations in predicting higher CS values. This can be attributed to the ANN model’s dependence on sufficient training data and its sensitivity to hyperparameter configuration, which may restrict its ability to capture complex nonlinear relationships in highly variable binder systems.

In contrast, the RF model, due to its ensemble learning structure, effectively handles data variability and reduces overfitting by aggregating multiple decision trees. This enables RF to maintain stable and accurate predictions even for high-performance geopolymer mixes. The observed underestimation of strength by ANN in high-strength specimens further highlights its limited generalization capability under complex material interactions. The cross-validation results further confirm the robustness of the RF model, with an average R² of 0.96 and low fold-to-fold variation. In contrast, the ANN model exhibited higher variability, indicating lower generalization capability. Therefore, the results suggest that while ANN can provide reasonable predictions for moderate-strength mixes, ensemble models such as RF are more reliable for accurately predicting the CS of high-strength GPC.

Along with prediction accuracy, the interpretation of the model also matters for ML applications. The RF model is more interpretable than the ANN model. RF allows us to evaluate how important each feature is, which helps us to identify the most influential parameters on the predicted output. On the contrary, an ANN functions as a black box model, such that one is unable to identify the relationship between input features and output predictions. Thus, while both models offer strong prediction ability, the RF model offers more insight into the contribution of each feature.

5.4. Model interpretability using SHAP analysis

To improve the interpretability of the developed machine learning model, SHAP (SHapley Additive exPlanations) analysis was performed to evaluate the contribution of each input parameter to the predicted CS. The SHAP summary plot (Figure 10) illustrates the relative importance of the input variables and their influence on the model output. From the SHAP plot, cement content is identified as the most influential parameter in predicting CS. Higher cement content is associated with positive SHAP values, indicating a significant positive contribution to CS. This behavior is expected, as increased cement content enhances hydration and improves strength development in cement-based materials.

Figure 10
SHAP summary plot illustrating the contribution of input features to the predicted CS of the ML model.

The aggregate content also has a noticeable impact on the model prediction. Higher aggregate values generally correspond to positive SHAP values, suggesting that aggregate proportion contributes to strength development by improving the load-bearing capacity and internal structure of the composite material. Unlike other parameters, water content displays a mixed influence, as indicated by both positive and negative SHAP values, while higher water contents tend to reduce CS. This trend reflects the well-known effect of excess water in cementitious systems, where an increased water-to-cement ratio leads to higher porosity and reduced strength.

The age of the specimen shows relatively lower SHAP values than the other parameters, indicating a comparatively lower influence on the predicted CS within the dataset used for model training. However, slight positive contributions are observed as curing age increases, consistent with the gradual strength gain due to ongoing hydration reactions.

The SHAP analysis provides a comprehensive understanding of the model’s decision-making process by measuring the impact of each input feature on CS prediction. The results confirm that cement content and aggregate proportion are the dominant factors governing strength development, whereas water content negatively influences strength at higher amounts. The agreement between these findings and the fundamental behavior of cementitious materials underscores the reliability, robustness, and interpretability of the proposed predictive model.

6. CONCLUSION

This study compares the performance of an ensemble ML model, RF, with an individual ML model, ANN. Also, it sought to develop a model for casting GPC with different binder types, including GGBS, FA, MK, RHA, SF, and binary combinations of GGBS with these binders. RF and ANN models were proposed to predict strength, and their effectiveness was evaluated using R², MAE, MSE, RMSE, and percentage error for different binder types.

The RF model ensemble performed satisfactorily with an R² of 0.979, demonstrating high predictive ability and minimal error. This was complemented by an MAE of 1.997 MPa, a MSE of 9.355 MPa², and a RMSE of 3.059 MPa. Prediction errors for all binder types were within ±7.15%, indicating the model’s reliability, generality, and strength in coping with a variety of binder combinations. The best accuracy was observed for the SF and RHA–GGBS mixtures, highlighting the RF model’s performance capability with intricate geopolymer formulations.

As a comparison, the ANN model performed moderately, with R² of 0.785 and much higher errors—MAE of 7.56 MPa, MSE of 95.68 MPa², and RMSE of 9.78 MPa. The ANN tended to underestimate CS across most binder mixes, with the largest discrepancy observed for the SF–GGBS combination, where the predictions were 47.3% lower than actual values. Notably, the ANN also overestimated strength in the SF mixes, predicting a non-zero value despite no measurable strength being detected in tests. These findings point out the limitations of the ANN model in accurately capturing the complex nonlinear interactions in high-performance geopolymer systems.

Considering the overall accuracy and predictive reliability, the study concludes that ensemble learning methods, particularly the RF model, are more appropriate for forecasting the mechanical performance of GPC. In addition to improved prediction accuracy, the proposed machine learning framework offers significant practical benefits. It reduces the need for extensive laboratory testing, thereby saving time and experimental effort. Furthermore, it minimizes material waste and lowers overall project costs by reducing trial-and-error mix design approaches. This makes the model highly suitable for real-world applications, enabling engineers to design sustainable GPC mixtures with improved performance and reliability efficiently.

This study shows that ensemble learning methods, particularly RF, provide superior predictive performance compared to single models like ANN in modelling multi-binder GPC systems. It highlights the importance of ensemble methods in capturing complex nonlinear interactions among materials. In practice, the developed model can be used by civil engineers and concrete manufacturers to optimize GPC mix designs, saving time, reducing laboratory testing, minimizing material waste, and promoting sustainable construction practices.

In practical terms, the ML models developed in this work can be used as useful decision-support tools for engineers and researchers involved in the GPC design. The RF model was the best-performing model. It can be used for quickly estimating CS based on mix composition and curing conditions, thereby reducing the intensive laboratory testing processes. This approach helps engineers to conduct optimization of GPC mix design more effectively, reduce experimental costs and accelerate the development of sustainable construction materials. Moreover, the use of data-driven prediction models in concrete design practices can further assist in the wider placement of GPC.

7. LIMITATIONS AND FUTURE WORK

Despite the promising results obtained in this study, certain limitations should be acknowledged. The dataset used for model development consists of 243 experimentally generated samples, which, although reliable, may still be limited in size for training highly complex ML models such as an ANN. In addition, the study considered specific binder combinations and curing conditions, which may limit the model to other geopolymer systems. The ANN model’s performance may also be influenced by the selection of hyperparameters and the network architecture, both of which could be further optimized.

With the addition of more experimental and field data, the model could become more robust, a future research focus. Employing state-of-the-art strategies, including hybrid ML models, deep learning techniques, and feature selection, can further improve predictive accuracy. Also, interpretability tools such as SHAP, sensitivity analysis and PDP can be used to understand better how different inputs affect the results. By extending the study to cover durability properties and real-time implementation in construction practices, the applicability of the present approach would be further strengthened.

8. DATA AVAILABILITY

Data will be available on request.

9. BIBLIOGRAPHY

• ^[1] SBAHIEH, S., MCKAY, G., NURDIAWATI, A., et al., “The sustainability of partial and total replacement of Ordinary Portland Cement: a deep dive into different concrete mixtures through life cycle assessment”, Journal of Building Engineering, v. 108, pp. 112830, Aug. 2025. doi: https://doi.org/10.1016/j.jobe.2025.112830.
  » https://doi.org/10.1016/j.jobe.2025.112830
• ^[2] ALSALMAN, A., ASSI, L.N., KAREEM, R.S., et al., “Energy and CO2 emission assessments of alkali-activated concrete and Ordinary Portland Cement concrete: a comparative analysis of different grades of concrete”, Cleaner Environmental Systems, v. 3, pp. 100047, Dec. 2021. doi: https://doi.org/10.1016/j.cesys.2021.100047.
  » https://doi.org/10.1016/j.cesys.2021.100047
• ^[3] IMTIAZ, L., REHMAN, S.K.U., ALI MEMON, S., et al., “A Review of recent developments and advances in eco-friendly geopolymer concrete”, Applied Sciences (Basel, Switzerland), v. 10, n. 21, pp. 7838, Nov. 2020. doi: https://doi.org/10.3390/app10217838.
  » https://doi.org/10.3390/app10217838
• ^[4] MAHESWARAN, M., CHRISTY, C.F., MUTHUKANNAN, M., “Synergistic effect of agro-industry by-product and silica nanoparticle on the fresh, mechanical, durability and microstructure enhancement of fly ash and GGBS-based geopolymer concrete”, European Journal of Environmental and Civil Engineering, v. 29, n. 12, pp. 2514–2544, Sep. 2025. doi: https://doi.org/10.1080/19648189.2025.2492074.
  » https://doi.org/10.1080/19648189.2025.2492074
• ^[5] PARUTHI, S., RAHMAN, I., KHAN, A.H., et al., “Strength, durability, and economic analysis of GGBS-based geopolymer concrete with silica fume under harsh conditions”, Scientific Reports, v. 14, n. 1, pp. 31572, Dec. 2024. doi: https://doi.org/10.1038/s41598-024-77801-z. PubMed PMID: 39738079.
  » https://doi.org/10.1038/s41598-024-77801-z
• ^[6] DAS, S.K., SHRIVASTAVA, S., “Siliceous fly ash and blast furnace slag based geopolymer concrete under ambient temperature curing condition”, Structural Concrete, v. 22, n. S1, pp. E341–E351, Jan. 2021. doi: https://doi.org/10.1002/suco.201900201.
  » https://doi.org/10.1002/suco.201900201
• ^[7] SAHA, A.K., “Effect of class F fly ash on the durability properties of concrete”, Sustainable Environment Research, v. 28, n. 1, pp. 25–31, Jan. 2018. doi: https://doi.org/10.1016/j.serj.2017.09.001.
  » https://doi.org/10.1016/j.serj.2017.09.001
• ^[8] RASHAD, A.M., “A brief on high-volume Class F fly ash as cement replacement - A guide for Civil Engineer”, International Journal of Sustainable Built Environment, v. 4, n. 2, pp. 278–306, 2015. doi: https://doi.org/10.1016/j.ijsbe.2015.10.002.
  » https://doi.org/10.1016/j.ijsbe.2015.10.002
• ^[9] AISHEH, Y.I.A., ATRUSHI, D.S., AKEED, M.H., et al., “Influence of steel fibers and microsilica on the mechanical properties of ultra-high-performance geopolymer concrete (UHP-GPC)”, Case Studies in Construction Materials, v. 17, pp. e01245, Dec. 2022. doi: https://doi.org/10.1016/j.cscm.2022.e01245.
  » https://doi.org/10.1016/j.cscm.2022.e01245
• ^[10] SINGH, R.P., VANAPALLI, K.R., CHEELA, V.R.S., et al., “Fly ash, GGBS, and silica fume based geopolymer concrete with recycled aggregates: Properties and environmental impacts”, Construction & Building Materials, v. 378, pp. 131168, May. 2023. doi: https://doi.org/10.1016/j.conbuildmat.2023.131168.
  » https://doi.org/10.1016/j.conbuildmat.2023.131168
• ^[11] SINGH, R.P., VANAPALLI, K.R., JADDA, K., et al., “Durability assessment of fly ash, GGBS, and silica fume based geopolymer concrete with recycled aggregates against acid and sulfate attack”, Journal of Building Engineering, v. 82, pp. 108354, Apr. 2024. doi: https://doi.org/10.1016/j.jobe.2023.108354.
  » https://doi.org/10.1016/j.jobe.2023.108354
• ^[12] LING, Y., WANG, K., WANG, X., et al., “Prediction of engineering properties of fly ash-based geopolymer using artificial neural networks”, Neural Computing & Applications, v. 33, n. 1, pp. 85–105, Jan. 2021. doi: https://doi.org/10.1007/s00521-019-04662-3.
  » https://doi.org/10.1007/s00521-019-04662-3
• ^[13] AMIN, M.N., KHAN, K., AHMAD, W., et al., “Compressive strength estimation of geopolymer composites through novel computational approaches”, Polymers, v. 14, n. 10, pp. 2128, May. 2022. doi: https://doi.org/10.3390/polym14102128. PubMed PMID: 35632011.
  » https://doi.org/10.3390/polym14102128
• ^[14] RATHNAYAKA, M., KARUNASINGHE, D., GUNASEKARA, C., et al., “Machine learning approaches to predict compressive strength of fly ash-based geopolymer concrete: A comprehensive review”, Construction & Building Materials, v. 419, pp. 135519, Mar. 2024. doi: https://doi.org/10.1016/j.conbuildmat.2024.135519.
  » https://doi.org/10.1016/j.conbuildmat.2024.135519
• ^[15] DONG, J.F., XU, Y., GUAN, Z.W., et al., “Freeze-thaw behaviour of basalt fibre reinforced recycled aggregate concrete filled CFRP tube specimens”, Engineering Structures, v. 273, pp. 115088, Dec. 2022. doi: https://doi.org/10.1016/j.engstruct.2022.115088.
  » https://doi.org/10.1016/j.engstruct.2022.115088
• ^[16] ANWAR, M.K., QURASHI, M.A., ZHU, X., et al., “A comparative performance analysis of machine learning models for compressive strength prediction in fly ash-based geopolymers concrete using reference data”, Case Studies in Construction Materials, v. 22, pp. e04207, Jul. 2025. doi: https://doi.org/10.1016/j.cscm.2025.e04207.
  » https://doi.org/10.1016/j.cscm.2025.e04207
• ^[17] ABDELLATIEF, M., HASSAN, Y.M., ELNABWY, M.T., et al., “Investigation of machine learning models in predicting compressive strength for ultra-high-performance geopolymer concrete: A comparative study”, Construction & Building Materials, v. 436, pp. 136884, Jul. 2024. doi: https://doi.org/ 10.1016/j.conbuildmat.2024.136884.
  » https://doi.org/10.1016/j.conbuildmat.2024.136884
• ^[18] ABDELLATIEF, M., ALANAZI, H., RADWAN, M.K.H., et al., “Multiscale characterization at early ages of ultra-high performance geopolymer concrete”, Polymers, v. 14, n. 24, pp. 5504, Dec. 2022. doi: https://doi.org/10.3390/polym14245504. PubMed PMID: 36559870.
  » https://doi.org/10.3390/polym14245504
• ^[19] LOPES, A., LOPES, S., PINTO, I., “Influence of curing temperature on the strength of a metakaolin-based geopolymer”, Materials (Basel), v. 16, n. 23, pp. 7460, Nov. 2023. doi: https://doi.org/10.3390/ma16237460. PubMed PMID: 38068203.
  » https://doi.org/10.3390/ma16237460
• ^[20] LE, Q.-H., NGUYEN, D.-H., SANG-TO, T., et al., “Machine learning based models for predicting compressive strength of geopolymer concrete”, Frontiers of Structural and Civil Engineering, v. 18, n. 7, pp. 1028–1049, Jul. 2024. doi: https://doi.org/10.1007/s11709-024-1039-5.
  » https://doi.org/10.1007/s11709-024-1039-5
• ^[21] TANYILDIZI, H., “Deep learning-based prediction of compressive strength of eco-friendly geopolymer concrete”, Environmental Science and Pollution Research International, v. 31, n. 28, pp. 41246–41266, 2024. doi: https://doi.org/10.1007/s11356-024-33853-2. PubMed PMID: 38844634.
  » https://doi.org/10.1007/s11356-024-33853-2
• ^[22] GAMIL, Y., “Machine learning in concrete technology: a review of current researches, trends, and applications”, Frontiers in Built Environment, v. 9, pp. 1145591, Feb. 2023. doi: https://doi.org/10.3389/fbuil.2023.1145591.
  » https://doi.org/10.3389/fbuil.2023.1145591
• ^[23] MANDAL, A., “Predicting compressive strength of concrete using advanced machine learning techniques: a combined dataset approach”, Asian Journal of Civil Engineering, v. 26, n. 3, pp. 1225–1241, Mar. 2025. doi: https://doi.org/10.1007/s42107-024-01247-x.
  » https://doi.org/10.1007/s42107-024-01247-x
• ^[24] VARALAKSHMI, K., KUMAR, J., “Optimized predictive maintenance for streaming data in industrial IoT networks using deep reinforcement learning and ensemble techniques”, Scientific Reports, v. 15, n. 1, pp. 27201, Jul. 2025. doi: https://doi.org/10.1038/s41598-025-10268-8. PubMed PMID: 40715212.
  » https://doi.org/10.1038/s41598-025-10268-8
• ^[25] ÖZBEY, M.F., TURHAN, C., “A novel comfort temperature determination model based on psychology of the participants for educational buildings in a temperate climate zone”, Journal of Building Engineering, v. 76, pp. 107415, Oct. 2023. doi: https://doi.org/10.1016/j.jobe.2023.107415.
  » https://doi.org/10.1016/j.jobe.2023.107415
• ^[26] ARUNVIVEK, G.K., ANANDARAJ, S., KUMAR, P., et al., “Compressive strength modelling of cenosphere and copper slag-based geopolymer concrete using deep learning model”, Scientific Reports, v. 15, n. 1, pp. 27849, Jul. 2025. doi: https://doi.org/10.1038/s41598-025-13176-z. PubMed PMID: 40738956.
  » https://doi.org/10.1038/s41598-025-13176-z
• ^[27] LI, H., WANG, Q., WEI, D., “A novel hybrid model combining BPNN neural network and ensemble empirical mode decomposition”, International Journal of Computational Intelligence Systems, v. 17, n. 1, pp. 77, 2024. doi: https://doi.org/10.1007/s44196-024-00446-3.
  » https://doi.org/10.1007/s44196-024-00446-3
• ^[28] ABDUL HAMID, N., MOHD NAWI, N., GHAZALI, R., et al., “Learning efficiency improvement of back propagation algorithm by adaptively changing gain parameter together with momentum and learning rate,” In: ICSECS 2011: Communications in Computer and Information Science, 2011. doi: https://doi.org/10.1007/978-3-642-22203-0_68.
  » https://doi.org/10.1007/978-3-642-22203-0_68
• ^[29] KHODAPARASTI, M., ALIJAMAAT, A., POURAMINIAN, M., “Prediction of the concrete compressive strength using improved random forest algorithm”, Journal of Building Pathology and Rehabilitation, v. 8, n. 2, pp. 92, Dec. 2023. doi: https://doi.org/10.1007/s41024-023-00337-8.
  » https://doi.org/10.1007/s41024-023-00337-8
• ^[30] WANG, Y., SHUI, Z., HUANG, Y., et al., “Improving the pozzolanic activity of metakaolin by urea intercalation technique”, Construction & Building Materials, v. 172, pp. 19–28, May. 2018. doi: https://doi.org/10.1016/j.conbuildmat.2018.03.189.
  » https://doi.org/10.1016/j.conbuildmat.2018.03.189
• ^[31] FARZINPOUR, A., DEHCHESHMEH, E.M., BROUJERDIAN, V., et al., “Efficient boosting-based algorithms for shear strength prediction of squat RC walls”, Case Studies in Construction Materials, v. 18, pp. e01928, 2023. doi: https://doi.org/10.1016/j.cscm.2023.e01928.
  » https://doi.org/10.1016/j.cscm.2023.e01928
• ^[32] CHICCO, D., WARRENS, M.J., JURMAN, G., “The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation”, PeerJ. Computer Science, v. 7, pp. e623, Jul. 2021. doi: https://doi.org/10.7717/peerj-cs.623. PubMed PMID: 34307865.
  » https://doi.org/10.7717/peerj-cs.623
• ^[33] VERMA, M., UPRETI, K., KHAN, M.R., et al., “Prediction of compressive strength of geopolymer concrete by using random forest algorithm”, In: International Conference on Advanced Communication and Intelligent Systems, 2022.
• ^[34] KHALAF, A.A., KOPECSKÓ, K., MERTA, I., “Prediction of the compressive strength of fly ash geopolymer concrete by an optimised neural network model”, Polymers, v. 14, n. 7, pp. 1423, Mar. 2022. doi: https://doi.org/10.3390/polym14071423. PubMed PMID: 35406295.
  » https://doi.org/10.3390/polym14071423
• ^[35] AHMED, H.U., MOHAMMED, A.A., MOHAMMED, A., “Soft computing models to predict the compressive strength of GGBS/FA- geopolymer concrete”, PLoS One, v. 17, n. 5, pp. e0265846, May. 2022. doi: https://doi.org/10.1371/journal.pone.0265846. PubMed PMID: 35613110.
  » https://doi.org/10.1371/journal.pone.0265846
• ^[36] GUPTA, P., GUPTA, N., SAXENA, K.K., et al., “Random forest modeling for fly ash-calcined clay geopolymer composite strength detection”, Journal of Composites Science, v. 5, n. 10, pp. 271, Oct. 2021. doi: https://doi.org/10.3390/jcs5100271.
  » https://doi.org/10.3390/jcs5100271
• ^[37] TRAN, N.T., NGUYEN, D.H., TRAN, Q.T., et al., “Experimental and machine learning based study of compressive strength of geopolymer concrete”, Magazine of Concrete Research, v. 76, n. 13, pp. 723–737, Jul. 2024. doi: https://doi.org/10.1680/jmacr.23.00144.
  » https://doi.org/10.1680/jmacr.23.00144
• ^[38] WANG, Y., IQTIDAR, A., AMIN, M.N., et al., “Predictive modelling of compressive strength of fly ash and ground granulated blast furnace slag based geopolymer concrete using machine learning techniques”, Case Studies in Construction Materials, v. 20, pp. e03130, Jul. 2024. doi: https://doi.org/10.1016/j.cscm.2024.e03130.
  » https://doi.org/10.1016/j.cscm.2024.e03130
• ^[39] LATIF, S.D., “Concrete compressive strength prediction modeling utilizing deep learning long short-term memory algorithm for a sustainable environment”, Environmental Science and Pollution Research International, v. 28, n. 23, pp. 30294–30302, Jun. 2021. doi: https://doi.org/10.1007/s11356-021-12877-y. PubMed PMID: 33590396.
  » https://doi.org/10.1007/s11356-021-12877-y
• ^[40] KINA, C., TANYILDIZI, H., AL BAKRI ABDULLAH, M.M., et al., “Comparison of deep LSTM and machine learning models for predicting compressive strength of fly ash/slag-based geopolymer concrete”, Scientific Reports, v. 15, n. 1, pp. 32871, Sep. 2025. doi: https://doi.org/10.1038/s41598-025-14365-6. PubMed PMID: 40998878.
  » https://doi.org/10.1038/s41598-025-14365-6

Publication Dates

• Publication in this collection
  24 July 2026
• Date of issue
  2026

History

• Received
  17 Jan 2026
• Accepted
  03 June 2026