AI-Driven Optimization of Fly Ash-Based Geopolymer Concrete for Sustainable High Strength and CO₂ Reduction: An Application of Hybrid Taguchi–Grey–ANN Approach

Abstract

The construction industry urgently requires sustainable alternatives to conventional concrete to reduce its environmental impact. This study addresses this challenge by developing machine learning-optimized geopolymer concrete (GPC) using industrial waste fly ash as cement replacement. An integrated Taguchi–Grey relational analysis (GRA) and artificial neural network (ANN) approach was developed to simultaneously optimize mechanical properties and environmental performance. The methodology analyzes over 1000 data points from 83 studies to identify key mix parameters including fly ash content, NaOH/Na₂SiO₃ ratio, and curing conditions. Results indicate that the optimized FA-GPC formulation achieves a 78% reduction in CO₂ emissions, decreasing from 252.09 kg/m³ (GRC rank 1) to 55.0 kg/m³, while maintaining a compressive strength of 90.9 MPa. The ANN model demonstrates strong predictive capability, with R² > 0.95 for strength and environmental impact. Life cycle assessment reveals potential savings of 3941 tons of CO₂ over 20 years for projects using 1000 m³ annually. This research provides a data-driven framework for sustainable concrete design, offering practical mix design guidelines and demonstrating the viability of fly ash-based GPC as high-performance, low-carbon construction material.

1. Introduction

Concrete is an elemental building material that mainly consists of ordinary Portland cement (OPC), which is estimated to account for approximately 8% of total global carbon emissions [1,2,3]. Owing to the environmental impacts of concrete manufacturing, scholars have focused on the utilization of supplementary cementitious materials (SCMs) as a massive part of OPC replacement. The use of ground granulated blast furnace slag (GGBFS) and fly ash (FA) decreases CO₂ emissions while improving the mechanical characteristics and workability of concrete [4,5]. Fly ash-based geopolymer concrete (FA-GPC) uses large quantities of fly ash, a by-product of coal combustion, as an ideal waste precursor for geopolymer synthesis due to its rich content of amorphous SiO₂ and Al₂O₃. Simultaneously, it constitutes the most prominent industrial waste product in the world. Globally, burned coal produces one billion tons of FA each year, with only 30% being utilized [3], indicating a significant potential for FA-GPC to reduce CO₂ emissions. In contrast to conventional concrete, FA-GPC not only decreases carbon dioxide emissions [4,5] but also provides similar mechanical properties [6,7], enhanced resistance to acid and sulfate attacks [8], and superior fire protection [9,10].

It is important to recognize that, while SCMs offer environmental benefits, they may also come with their own environmental implications that need to be considered. A new type of concrete that is synthesized when OPC is completely replaced by SCMs is called geopolymer composite [6,7,8]. Similar findings indicate that GPC improves the possibility of concrete production as a sustainable option for OPC concrete. However, there are limitations to the exact determination of the contributions of SCMs to geopolymer concrete mixes and the resulting CO₂ emissions. This lack of clarity has prevented the complete application of such sustainable materials in construction lines and, therefore, can hinder the reduction of impacts on the environment adequately.

The lack of accurate methods for forecasting the carbon footprint of GPC materials persists as a problem for enhancing the sustainability of this novel concrete innovation. The use of various SCMs derived from industrial waste has been shown to significantly reduce CO₂ emissions. Specifically, GPC can generate up to 26–80% less CO₂ than OPC-based concrete even with equivalent binder loadings [9,10]. This is because the manufacture of geopolymer concrete triggers the reaction of SCMs rich in silica and alumina using alkali activator solutions. Unlike OPC-based concrete, geopolymerization yields tetrahedral aluminosilicate shapes with sufficient mechanical stiffness for construction [11,12]. Thus, GPC is an ideal alternative to OPC-based concrete, potentially resulting in a significant reduction in CO₂ emissions in the construction sector [13,14,15]. Moreover, several production parameters, such as the type of alkaline solutions [16,17,18] and curing conditions [19,20], interact very closely with SCMs, affecting the GPC geopolymerization process. These factors affect the carbon emissions associated with GPC synthesis. There has already been significant variability in the predicted CO₂ emission reductions for GPC compared with OPC concrete. Published data are widely inconsistent, with some estimates of CO₂ emission cutting as high as 80% [21], whereas others estimate cuts between 26% and 45% [22,23]. In some cases, cutbacks have been as low as 9% [24]. The different estimates of CO₂ emissions from GPC can be explained by factors such as the composition of elements in the raw SCMs, types and concentrations of alkaline activators, and curing temperature and time. For instance, the addition of FA to GPC yields a secondary binder, but requires activation using high-pH alkaline activators [25]. Activation increases CO₂ emissions, making FA-based GPC less environmentally sound compared with concrete, but this does not apply to all geopolymers [26]. Testing is still the most commonly used method for identifying optimal parameter values and learning how they affect the GPC attributes and performance. However, these studies generally illustrate the immobility and sensitivity of CO₂ measurements. Machine learning (ML) techniques can help solve this problem by delivering better adaptive estimates with extremely precise estimation algorithm [27].

ML algorithms have become common in various fields over the past few decades, owing to their simplicity and accuracy. These researchers used several regression and ML algorithms to evaluate its features, such as a support vector machines (SVMs) [28], artificial neural networks (ANNs) [29,30], decision trees (DTs) [27,31], random forest (RFs) [32,33], k-nearest neighbors (kNNs) [6,34], and extreme learning machines (ELMs) [35,36] for prediction modeling. Some studies have successfully predicted the strength performance of FA-GPC using these methods. However, challenges persist in optimizing the mixture design to ensure both strength and sustainability.

This study integrates Taguchi–Grey relational analysis (GRA) and artificial neural networks (ANNs) to optimize fly ash-based geopolymer concrete (FA-GPC) for strength and CO₂ reduction. While Taguchi-GRA efficiently optimizes multiple parameters [37,38], few studies have applied it to FA-GPC mix design [39]. Additionally, predictive models for eco-efficiency and performance are lacking. This study’s findings provide practical guidelines for mix design and a scalable computational model bridging materials science with decarbonized construction practices. This study is the first to combine Taguchi-GRA with ANNs for FA-GPC optimization, utilizing an extensive dataset (>1000 entries). The approach simultaneously enhances strength, reduces emissions, and provides a predictive carbon footprint model by bridging sustainability and engineering needs in geopolymer concrete.

This study aims to propose and predict the CO₂ footprint and compressive strength of FA-GPC using a data-driven approach that integrates Grey relational analysis (GRA) and artificial neural networks (ANNs). By doing so, it contributes to the advancement of construction sustainability. A novel predictive framework is developed that combines the multi-criteria decision-making capabilities of GRA with the nonlinear modeling strengths of ANN. The main objectives of the study are as follows: First, a reference dataset consisting of over 1000 entries is collected to support robust model development. Second, the integration of GRA with ANN is employed to optimize FA-GPC mixture designs that balance strength and minimal CO₂ emissions. Third, a life cycle assessment (LCA) is conducted to evaluate the CO₂ emissions of various FA-GPC mixtures, with validation through ANN model predictions. Lastly, sensitivity analysis is performed to investigate the influence of specific parameters on the environmental impacts of optimized FA-GPC.

2. Development of Reference Datasets

In the initial phase, the factors affecting the concrete strength and CO₂ emissions of FA-GPC and the critical factor for a predictive model were explored. Material optimization in civil engineering aims to achieve an optimal balance between maximizing strength and minimizing CO₂ emissions. This process can be categorized into four primary objectives: cost reduction, enhancement of structural performance, reduction of carbon emissions, and multi-objective optimization, which combines two or more of these goals. The methodology for structural optimization involves using optimization techniques and practices. Computational software, such as ANNs and GRA models, iteratively determines the best material mix design. Python (version 3.9) is used for ANN-based calculations, while Excel is employed for CO₂ and GRA calculations. Once optimized, the material mix data can be seamlessly transferred to structural design software like ETABS, or ABAQUS for further application.

Table 1. Descriptive analysis of the influential factors for FA-GPC composite.

No.	Influential Factors		Unit	Mean	Std.	Min.	Q1	Q2	Q3	Max.
X1	SiO2	Chemical composition	%	51.4	11.1	17.6	45.2	51.3	59.7	77.2
X2	Al2O3		%	24.9	5.9	6.4	21.4	26.5	28.4	45.9
X3	CaO		%	6.84	7.66	0.00	2.10	3.42	10.7	37
X4	Fe2O3		%	9.24	6.45	0.90	4.57	7.12	13.2	40.5
X5	MgO		%	1.77	1.51	0	0.77	1.26	2.20	9.39
X6	Fly ash	Mix proportions	kg/m3	415.7	94.3	254.5	388	408	450	1050
X7	C-Agg		kg/m3	1167.5	174.2	594.3	1080	1201	1276.8	2291
X8	F-Agg		kg/m3	635	134	318	548	630	663	1728
X9	NaOH		kg/m3	379	667.5	0	44.9	58.4	108	1967
X10	Na2SiO3		kg/m3	133	55.1	0	103	121.7	143	386
X11	Molarity		kg/m3	11.7	3.1	2	10	12	14	20
X12	Extra. H2O		kg/m3	15.1	26.6	0	0	0	20.70	175
X13	Na2O	Alkaline activator factors	kg/m3	35.1	13.1	14.7	27.2	31.8	37.9	128.6
X14	SiO2		kg/m3	40.5	15.8	0	30.3	38.4	44.9	113.5
X15	H2O		kg/m3	137.5	50.9	63.8	100	123.0	167.8	358.9
X16	Density	-	kg/m3	2431	298	1991	2371	2400	2442	5564
X17	Na/Al	Formulation ratios	-	0.61	0.28	0.18	0.42	0.54	0.68	2.16
X18	Si/Al		-	2.25	0.89	0.62	1.84	2.03	2.46	7.76
X19	Temp	Curing conditions	°C	62.8	24.8	20	60	60	80	120
X20	HC days		days	0.89	0.68	0	1	1	1	4
X21	ages		days	28.5	67.4	1	7	7	28	540
X22	Specs.	-	-	0.39	0.49	0	0	0	1	1
Y1	fc	Output responses	MPa	38.50	15.02	1.10	28.61	39.52	48.38	90.90
Y2	CO2		kg/m3	186.80	61.50	55.01	151.11	172.60	194.70	594.33
Y3	GRG		0–1	0.58	0.06	0.35	0.55	0.59	0.61	0.79

lists the factors studied in related research, including the FA properties, mix design proportions, curing conditions, and concrete specimens. Previous studies [40,41,42] have focused on selective variables, and this study included 22 input variables, such as FA characteristics, mix design proportions, alkali activator chemical contents, formulation ratios, curing parameters, and specimen morphology for a robust strength and CO₂ reduction model of FA-GPC. The impact of the specimen geometry was considered by converting the cylinder strengths to equivalent cube strengths for analysis [43]. The Chinese and European standards prefer cubes, while the US standard uses cylindrical specimen with a length-to-diameter (l/d) ratio of 2.

A dataset of 1126 FA-GPC mixture observations from 82 existing studies was compiled. This dataset includes variables such as the fly ash dosage, alkaline activator concentration, curing conditions, compressive strength, and CO₂ emissions. Table 1 presents a statistical assessment of the data variability for 22 input components (X1 to X22) and FA-GPC output parameters. Factors X1–X5 represent the key chemical constituents of the FA binder, indicating the FA class. Variables X6 to X12 are related to the FA-GPC mix design proportions. Variables X13–X15 pertain to the composition of water glass or sodium silicate (Na₂SiO₃), including Na₂O, SiO₂, and water. Sodium hydroxide (NaOH) is a part of the alkali activation process. X16 relates to concrete density, X17 and X18 to formulation ratios, X19–X21 to curing conditions, and X22 to specimen type such as cube or cylinder. Table 1 presents the key statistical metrics, including the maximum (max), minimum (min), mean, first quartile (Q1, 25%), second quartile (Q2, 50%), third quartile (Q3, 75%), and standard deviation (std), highlighting data trends and variability for significant insights into data characteristics. Q1, Q2, and Q3 provide key insights into the factors influencing FA-GPC composite by identifying and quantifying the relationships between specific parameters and their impact on the composite’s properties, supporting the descriptive analysis of these influential factors.

Figure 1 displays the correlation matrix of various material properties and their relationships with different variables, including SiO₂, Al₂O₃, CaO, and others. The color-coded heatmap indicates the strength and direction of these correlations, with blue and red hues representing negative and positive correlations, respectively. The correlation graph in Figure 2 illustrates the influence of 22 attributes on the Grey rational analysis (GRA) grade, with molarity showing the highest positive correlation at 0.36, indicating that higher molarity significantly enhances the GRA grade and potentially strengthens the concrete structure. Other positive factors include temperature (0.19), SiO₂ content (0.16), heat-curing days (0.14), and coarse aggregate (C-Agg, 0.13), suggesting that increased curing time, silica content, and coarse aggregate contribute positively to the GRA grade, likely improving durability and strength. In contrast, several attributes showed strong negative correlations with the GRA grade. Na₂O in AS has a significant negative contribution with a correlation of −0.57, Na/Al ratio of −0.49, and Na₂SiO₃ in AS with −0.36, indicating that higher concentrations of these components degrade the performance. CaO has a correlation of −0.31 and density −0.2, but it is less significant. Attributes such as Fe₂O₃ (0.07) and specifications (0) have negligible to no relationship with the GRA grade. Overall, the data suggest that increasing the molarity, temperature, SiO₂, and coarse aggregate may positively impact the GRA grade, while controlling Na₂O, Na/Al, and other silicates is crucial for preventing performance degradation. These insights guide the optimization of concrete formulations, balancing key attributes for improved structural properties with environmental considerations.

Figure 3 illustrates the impact of various materials, ratios, and curing conditions on the GRA grade, measured from 0 to 1. For SiO₂ (%) and Al₂O₃ (%), with levels up to 90% and 50%, respectively, the trend remained flat, indicating no influence on the GRA grade. CaO (%) shows a gentle positive trend up to 40%, and Fe₂O₃ (%) up to 30%, slightly increasing the GRA grade. NaOH (kg/m³) had a clear negative correlation; higher concentrations of up to 350 kg/m³ significantly reduced the GRA grade. Similarly, Na₂SiO₃ (kg/m³) up to 400 kg/m³ also showed a negative trend, further reducing the GRA grade. Fly ash up to 1200 kg/m³ exhibited a slight, almost negligible, negative trend. Coarse aggregates up to 2000 kg/m³ show a soft positive trend, indicating a slight improvement in the GRA grade. Fine aggregates presented a weak negative effect on kg/m³. While the study did not examine the type, gradation, or size of aggregates, it is important to note that these factors are known to significantly influence the strength and performance of FA-GPC. Molarity and extra water showed soft positives for up to 20 M and 100 kg/m³, respectively, indicating slight GRA grade improvements at these levels. Formulation ratios, such as those of Na/Al and Si/Al, of up to 2.5 and 10, respectively, showed slight negative trends, indicating minor reductions at higher ratios. Heat curing temperatures from 20 °C to 120 °C and heat curing periods of up to 5 days have mild positive effects on the GRA grade, with higher temperatures and prolonged times slightly improving the material quality. Additionally, a small upward trend was observed for the total curing age up to 600 days, suggesting that prolonged curing may favor the GRA grade. In summary, while high concentrations of NaOH and Na₂SiO₃ reduce GRA grade, optimized curing conditions and slight increases in molarity and water content offer marginal improvements.

3. Life Cycle Assessment and CO₂ Calculations

This work introduces the “gate to cradle CO₂ emission” technique to assess the life cycle and predict the carbon footprint for 1 m³ of FA-based geopolymer concrete composite, illustrated in Figure 4. This method accounts for all materials used in FA-GPC manufacturing and their components, providing a carbon impact. A comprehensive database of 1126 data points from 83 previous studies was employed to achieve the aims of this study, forming the basis for developing computational machine learning techniques. The dataset includes detailed information on key mix design elements, such as the percentages of the main chemical elements in the FA binder, the quantities of FA and aggregates (kg/m³), and the ingredients and composition of the alkali activation. Data collection and validation followed the initial conclusions of [44,45], in which identical blends across studies were identified and excluded to ensure dataset integrity. In this study, the dataset was further refined by selecting only articles related to FA-based GPC, while maintaining consistency in feature presentation using uniform metric measures.

The Equation (1) for CO₂ emissions includes sources from raw material production, transportation, and curing processes, integrated via an algorithm documented in several studies [46,47]:

$$\begin{array}{ll}{\mathrm{C}\mathrm{O}}_2 \mathrm{emissions} & =E_{\mathrm{RawMatProd} }+E_{\mathrm{RawMatTrans} }+E_{\mathrm{Curing} }\\ & =\sum \left(F_i\times W_i\right)+\sum \left(F_{\mathrm{Trans} }\times W_i\times D_i\right)+F_{\mathrm{Eng} }\times {Eng}_{\mathrm{Curing} }\end{array}$$

(1)

Equation (1) is the sum of CO₂ emissions from the raw material production ($E_{\mathrm{RawMatProd}}$), transportation ($E_{\mathrm{RawMatTrans}}$), and curing ($E_{\mathrm{Curing}}$) processes. Each component is weighted by factors representing their contributions to overall emissions. The emission factor (kgCO₂/kg) represents the amount of CO₂ emissions produced per kilogram of each material. The emission factor (kgCO₂/kg) represents the amount of CO₂ emissions produced per kilogram of each material. Transportation emissions are influenced by emission factors (kgCO₂/kg-km) and transportation distances (kms). The curing contributions are captured by reflecting the energy consumption during curing (kgCO₂/kWh). This structure integrates multiple emission sources for a detailed environmental impact analysis of the production.

Equation (2) evaluates the energy consumption for concrete curing [24,47,48,49], based on the energy conservation. Concrete is brought to the appropriate curing temperature and shielded from thermal loss through the application of energy.

$${Eng}_{\mathrm{Curing} }={ \mathrm{Eng} }_{\mathrm{Heat} }+{ \mathrm{Eng} }_{\mathrm{Loss} }=M\times \mathcal{C}\times \left(T-T_0\right)+P\times t_1/24$$

(2)

In this context, “${\mathrm{Eng}}_{\mathrm{Heat}}$” denotes the energy required to elevate curing heat temperature, “${\mathrm{Eng}}_{\mathrm{Loss}}$” represents the heat loss energy expended to counteract, “M” signifies the total mass (kg) of the concrete, $\mathcal{“}\mathcal{C}\mathcal{”}$ indicates the specific heat capacity (i.e., C = 700 unit, J/kg-°C) [47], “T” refers to the curing temperature (°C), $“T_0”$ is the room curing temperature (°C), $“P”$ is the heating power utilized to mitigate heat loss (unit, kW), and ${“t}_1”$ denotes the curing heat duration in days. The heating power $“P”$ required to counteract heat loss is precisely related to the temperature differential between “T” and $“T_0”,$ as per the heat transfer principle [49]. It is evident that $“P”$ at $“T_0”$ is equal to 0.

Designating $“P”$ at an oven temperature of 80 °C as $P_0$, Equation (3) is formulated to compute $“P”$ under various curing heating conditions.

$$P=\left(T-T_0\right)/\left({80 }^{\circ }\mathrm{C}-T_0\right)\times P_0$$

(3)

This study sourced energy consumption and emissions data from the literature, noting that many CO₂ prediction models overlook factors like material transport impacts and differences between mass manufacturing and controlled testing, often due to insufficient data [44,50]. This study addresses these gaps by integrating critical factors into CO₂ emission predictions using reference data and incorporating average material transport distances from the literature. The emission factors used for the calculation of CO₂ emissions are presented in

Table 2. Parameters used for CO₂ calculation.

Parameters	Range/Average	Transport Distances (Km)	References
Emission factors (kgCO2/kg)
Portland cement	0.73	50–300 = 75	[51,52]
Fly ash	Varies with price	50–300 = 75	[52]
Aggregates (fine and coarse)	0.0039	50	[51,52]
NaOH pellets	1.04–1.59 = 1.32	50–300 = 75	[52,53]
Na2SiO3 solution	0.376	50–300 = 75	[48,52]
H2O	0.00015	0	[51]
Admixture (superplasticizer)	0.72	50–300 = 75	[51,52]
Transportation details
Mode of transport	By road	/	/
Unit price (USD/ton. km)	0.069	/	[54]
Emission factors (kgCO2/kg⋅ km)	0.000137	/	[51]
Curing details
Energy source	Electricity	/
Unit price (USD/kWh)	0.0809	/	[55]
Emission factors (kgCO2/kWh)	0.55	/	[55]

.

4. Research Methodology

4.1. Taguchi–Grey Rational Analysis (GRA) Calculation Model

The Taguchi–Grey relational analysis (Taguchi-GRA) method was selected for its efficiency in solving multi-objective optimization problems. By combining the robustness of the Taguchi method with GRA, which ranks multiple conflicting objectives, Taguchi-GRA proves effective in optimizing complex systems. For single-objective optimization, the Taguchi method is recognized for its statistical power, while GRA is ideal for optimizing multiple response functions [39,56,57]. This study applied Taguchi-GRA to assess the effects of various factors on compressive strength and CO₂ emissions. The method identifies key parameters that enhance the FA-GPC performance by increasing the strength and reducing CO₂ emissions. The research methodology adopted in this study is presented in Figure 5.

GRA analysis involves three main steps. The first step was to normalize the measured output function individually. Normalization is a data scaling technique used to bring all variables to a comparable scale, facilitating the integration of multiple conflicting objectives in optimization. To minimize CO₂ emissions, the “smaller is better” normalization approach is applied, as shown in Equation (4).

$$Y_{ij}={\displaystyle \frac{\left(max\left(z_{ij}\right)-\left(z_{ij}\right)\right)}{max\left(z_{ij}\right)-min\left(z_{ij}\right)}}$$

(4)

where $Y_{ij}$ is the normalized value of the i-th output function for the j-th experiment, $z_{ij}$ is the measured value for the i-th output in the j-th experiment, and $min\left(z_{ij}\right)$ and $max\left(z_{ij}\right)$ are the minimum and maximum measured values for the i-th output across all experiments.

The goal for CO₂ emissions is minimization, where smaller values are preferred. In contrast, compressive strength is maximized, as higher values signify better material performance, contributing to improved structural integrity and durability. Therefore, the criterion for compressive strength (CS) is that higher values are desirable, as reflected in Equation (5).

$$Y_{ij}={\displaystyle \frac{Z_{ij}-min\left(z_{ij}\right)}{max\left(z_{ij}\right)-min\left(z_{ij}\right)}}$$

(5)

Here, “Z_ij” indicates the values derived from the experimental dataset, with min (Z_ij) representing the lowest experimental value and max (Z_ij) representing the highest value from the literature reference data. The normalized data were determined using Equation (6):

$$GR{C}_{ij}={\displaystyle \frac{\left({\delta }_{min}-\gamma {\delta }_{max}\right)}{\left({\delta }_{ij}-\gamma {\delta }_{max}\right)}}$$

(6)

where $“GR{C}_{ij}”$ represents the Grey relational coefficient for the i-th output function in the j-th experiment. The parameters $“{\delta }_{min}”$ and ${“\delta }_{max}”$ correspond to the minimum and maximum values of the deviation sequence, respectively. Additionally, $“\gamma ”$ is a distinguishing coefficient, typically taking values between 0 and 1, which helps to differentiate the importance of each factor in the optimization process.

A superior Grey relational grade (GRG) suggests an ideal combination of input parameters in the FA-GPC. The GRG was calculated using the following:

$$GR{G}_i={\displaystyle \frac{1}{n}}\epsilon GR{C}_{ij}$$

(7)

where $“GR{G}_i”$ represents the Grey relational grade for the i-th output, $“GR{C}_{ij}”$ is the Grey relational coefficient for the i-th output and the j-th experiment, and $“n”$ refers to the total number of experiments conducted in the analysis.

Table 3. Grey rational grade analysis for the best combination based on higher strength and low CO₂ emission criteria for FA-GPC.

Sources		Inputs Variables																					Output Parameters		Grey Rational Analysis (GRA)
		Chemical Composition of Fly Ash				Mixture Proportion							Alkaline Activator				Formulation Ratios		Curing Conditions
No.	Ref.	SiO2	Al2O3	CaO	Fe2O3	FA	CAgg	FAgg	NaOH	Na2SiO3	M	Ex.H2O	Na2O	SiO2	H2O	ρ	Na/Al	Si/Al	CTem	Cper	Ages	Spec.	fc	CO2	Grade	Rank
	Unit	%	%	%	%	Kg/m3	Kg/m3	Kg/m3	Kg/m3	Kg/m3	/	Kg/m3	Kg/m3	Kg/m3	Kg/m3	Kg/m3	/	/	C	Days	Days		MPa	Kg/m3	0–1	-
1	[58]	47.9	28.0	3.8	14.1	410	1155	587	41	164	10	2.6	33.7	48.2	125.7	2359	0.48	1.81	80	1	28	Cu	55.5	169.90	0.63	162
2	[59]	50.2	26.4	4.3	10.0	408	1110	739	41	103	14	0	27.6	30.3	86.1	2401	0.42	1.86	75	2	7	Cy	63.1	154.72	0.67	50
3	[60]	51.5	23.6	1.7	15.3	408	1168	660	41	103	16	0	29.0	30.3	84.7	2380	0.50	2.12	60	1	28	Cy	74.9	141.28	0.75	5
4	[61]	32.6	31.2	17.1	8.5	300	1372	696	43	107	14	0	28.8	31.5	89.7	2518	0.51	1.17	20	0	28	Cu	17.1	129.83	0.58	614
5	[62]	53.6	28.5	4.2	8.7	350	1282	690	45	113	15	0	32.7	39.1	85.7	2479	0.54	1.93	70	1	3	Cu	46.8	154.37	0.62	232
6	[63]	37.6	14.8	19.6	18.6	310	1204	649	49	122	10	0	31.2	42.4	97.4	2334	1.12	2.95	80	1	365	Cu	47.3	162.11	0.61	284
7	[64]	42.0	33.6	12.7	4.0	427	1277	547	31	76	14	0	20.5	22.4	63.8	2358	0.24	1.20	80	1	28	Cu	42.4	122.15	0.64	130
8	[65]	50.8	27.3	5.4	4.6	460	1201	539	49	151	14	0	28.2	41.8	130.0	2400	0.37	1.86	80	1	480	Cu	80.7	176.65	0.75	4
9		50.8	27.3	5.4	4.6	460	1174	527	80	160	14	0	38.4	44.1	156.7	2400	0.50	1.88	80	1	480	Cu	84.5	220.88	0.75	6
10		50.8	27.3	5.4	4.6	460	1186	533	74	147	12	0	32.9	40.7	147.2	2400	0.43	1.85	80	1	480	Cu	83.3	207.99	0.75	7
11	[66]	32.3	16.4	19.1	19.0	475	1060	707	64	96	14	74	28.4	27.6	177.9	2476	0.60	1.98	27	0	180	Cu	42.0	156.99	0.60	362
12	[67]	54.1	26.7	3.4	6.5	425	1201	647	36	91	12	42.3	22.6	25.5	121.7	2442	0.33	1.92	75	1	28	Cu	49.1	137.02	0.64	127
13	[68]	27.9	14.4	27.9	15.6	400	970	795	187	93	2	0	25.9	31.8	222.3	2445	0.74	2.12	27.5	0	90	Cy	32.5	318.49	0.47	1060
14	[69]	49.5	29.6	3.5	10.7	1050	2291	1728	209	198	10	88	77.9	58.2	358.9	5564	0.41	1.58	80	1	7	Cy	46.1	451.01	0.45	1085
15	[70]	49.5	29.6	3.5	10.7	1016	2291	1728	123	324	10	79	76.3	95.3	354.4	5561	0.42	1.69	80	1	90	Cy	59.6	385.24	0.52	991
16	[71]	49.4	31.3	4.8	4.5	416	927	699	65	292	15	8	62.4	85.8	216.8	2407	0.79	1.90	80	1	28	Cy	43.5	252.09	0.53	958
17	[72]	47.9	28.0	3.8	14.1	416	927	699	65	292	15	8	62.4	85.8	216.8	2407	0.88	2.08	80	1	365	Cy	90.9	252.09	0.79	1
18	[73]	45.2	20.0	15.5	13.2	414	1091	588	69	138	20	0	49.1	45.4	112.6	2300	0.98	2.39	23	0	28	Cy	55.1	176.77	0.62	193
19	[74]	37.7	20.0	23.4	5.6	570	780	658	N/A	228	N/A	0	28.5	42.8	156.8	2236	0.41	1.92	50	2	28	Cy	56.3	136.22	0.67	67
20	[75]	35.9	15.1	17.2	17.3	450	1150	500	108	162	12	0	49.4	48.9	171.7	2370	1.20	2.63	60	2	7	Cy	48.9	258.85	0.54	918
21	[76]	62.2	27.5	2.3	3.9	480	1140	625	56	112	14	35	33.5	32.9	136.5	2448	0.42	2.13	80	1	28	Cy	67.9	171.43	0.68	44
22	[77]	51.7	31.9	1.2	3.5	500	950	760	100	100	10	0	31.7	27.6	140.7	2410	0.33	1.52	60	0.167	7	Cy	30.7	216.18	0.53	971
23	[78]	48.0	29.0	1.8	12.7	408	1202	647	41	103	14	15	27.6	30.3	101.1	2416	0.38	1.62	60	1	28	Cy	46.6	141.51	0.63	161
24	[79]	53.7	27.2	1.9	11.2	400	1222	658	40	100	14	0	23.7	30.0	86.3	2420	0.36	1.91	20	0	180	Cy	41.3	123.77	0.64	143
25	[80]	50.0	28.3	1.8	13.5	400	1209	651	46	114	14	0	27.0	35.1	97.9	2420	0.39	1.77	21.5	0	56	Cy	33.0	137.29	0.60	368
26	[81]	49.6	45.9	0.0	4.5	400	1100	720	N/A	160	10	40	58.8	52.0	249.2	2580	0.53	1.16	23	0	28	Cu	12.0	308.55	0.44	1094
27	[82]	50.7	28.8	2.4	8.8	400	1293	554	45	113	14	0	28.8	30.1	99.1	2405	0.41	1.72	100	3	28	Cu	45.0	188.24	0.58	600
28	[83]	53.7	27.2	1.9	11.2	400	1209	651	46	114	14	0	27.0	34.4	98.6	2420	0.41	1.95	21.5	0	90	Cy	38.4	137.29	0.61	266
29	[84]	63.3	26.8	2.5	5.6	380	1189	660	49	122	8	0	27.5	35.9	107.6	2400	0.44	2.31	25	0	28	Cu	30.0	144.87	0.59	546
30	[85]	37.6	14.8	19.6	18.6	310	1204	649	57	114	10	0	31.8	39.6	99.7	2334	1.14	2.89	80	1	90	Cu	45.5	169.76	0.60	390
31	[86]	53.0	27.9	4.2	8.7	310	1204	649	57	114	10	0	31.8	39.6	99.7	2334	0.60	2.00	80	1	365	Cu	47.5	169.76	0.60	330
32	[87]	32.1	19.9	18.8	16.9	390	1092	585	67	167	16	0	37.7	50.1	146.2	2301	0.80	1.92	30.5	0	28	Cy	38.0	186.91	0.57	750
33	[88]	47.5	17.3	2.3	6.0	279	956	721	21	363	15	0	60.1	106.7	217.2	2340	2.05	4.22	120	1	7	Cu	30.9	223.48	0.52	983
34	[89]	51.3	30.1	8.7	4.6	400	950	850	57	143	10	48	33.7	42.1	172.2	2448	0.46	1.75	70	2	28	Cu	51.4	188.65	0.60	378
35	[90]	43.9	20.1	21.2	5.0	450	972	810	68	68	10	78.3	25.7	19.8	167.8	2445	0.47	2.04	70	1	28	Cy	48.4	167.27	0.61	296
36	[91]	51.2	24.0	5.6	6.6	400	950	850	57	143	12	40	36.5	42.5	161.0	2440	0.62	2.19	70	1	28	Cu	53.0	182.17	0.61	290
37	[92]	58.5	21.0	3.8	8.3	420	1090	630	60	150	12	31	38.2	44.1	158.7	2381	0.71	2.79	80	1	7	Cy	54.5	189.80	0.61	298
38	[93]	50.5	26.6	2.1	13.8	499	1168	584	43	107	14	18.8	28.7	31.4	108.1	2419	0.36	1.82	75	1	91	Cy	65.8	152.75	0.69	33
39	[94]	45.8	21.4	13.7	12.6	405	1269	683	81	81	13	81	31.4	21.9	189.7	2600	0.60	2.03	80	1	56	Cu	29.0	192.45	0.54	928
40	[95]	49.1	26.4	5.2	4.6	459	1172	539	57	143	10	0	35.3	55.0	109.7	2370	0.48	1.97	20	0	90	Cu	50.5	163.67	0.62	218
41	[96]	74.2	9.4	5.5	4.2	437	1318	647	62	156	10	0	28.2	43.1	146.6	2620	1.13	7.57	80	1	28	Cu	32.0	198.00	0.54	917
42	[97]	56.7	24.9	5.2	6.9	400	1200	600	60	120	8	0	29.3	35.3	115.4	2380	0.48	2.24	23	0	28	Cy	15.9	158.45	0.55	890
43	[98]	48.7	16.6	18.7	6.9	494	858	691	99	99	14	0	44.9	29.6	123.0	2241	0.90	2.80	60	3	4	Cy	83.6	226.46	0.74	10
44	[99]	37.3	6.4	10.7	40.5	513	1477	409	60	120	10	0	23.4	23.8	132.5	2579	1.18	5.60	24	0	56	Cy	55.8	162.60	0.64	134
45	[100]	70.3	23.1	0.2	1.4	467	1039	784	147	234	10	10	68.7	68.8	253.5	2681	1.05	3.13	80	1	360	Cy	38.9	342.27	0.47	1054
46	[101]	57.9	31.1	1.3	5.1	300	1131	754	39	96	10	68.81	24.7	33.4	145.7	2388	0.44	1.89	100	1	28	Cu	26.9	149.03	0.58	649
47	[102]	74.2	9.4	5.5	4.2	459	1172	539	57	143	10	0	35.3	55.0	109.7	2370	1.34	7.76	20	0	56	Cu	17.0	163.67	0.55	906
48	[103]	54.0	34.7	2.0	5.3	426	1213	643	55	137	12	0	26.4	38.3	126.9	2472	0.29	1.54	60	1	28	Cu	53.0	173.98	0.62	228
49	[104]	57.3	27.1	0.0	8.1	428	1170	630	57	114	14	43	36.6	39.9	137.5	2442	0.52	2.09	90	1	1	Cu	49.0	176.50	0.60	350
50	[105]	65.3	14.5	5.6	6.3	460	1201	539	50	150	14	0	28.5	41.5	130.1	2400	0.70	4.36	80	1	28	Cu	72.8	177.74	0.70	27
51	[106]	51.2	24.0	5.6	6.6	400	950	850	57	143	12	40	35.8	42.1	162.1	2440	0.61	2.19	70	1	2	Cu	53.5	182.17	0.61	278
52	[107]	51.9	9.2	2.3	34.0	410	1044	531	67	117	10	79.2	33.7	38.6	191.5	2249	1.48	5.69	20	0	180	Cu	29.0	164.00	0.57	745
53	[108]	51.3	30.1	8.7	4.6	400	950	850	57	143	12	48	35.8	42.1	170.1	2448	0.49	1.75	70	2	30	Cu	53.8	188.65	0.61	308
54	[109]	47.8	24.4	2.4	17.4	408	1294	554	41	103	8	0	23.1	30.3	90.6	2400	0.38	1.92	60	1	7	Cy	69.7	141.49	0.72	16
55	[110]	50.5	26.6	2.1	13.8	408	1201	647	68	103	14	0	30.1	29.8	111.1	2427	0.46	1.85	60	1	28	Cy	56.6	177.65	0.63	177
56	[111]	51.1	25.6	4.3	12.5	408	1233	554	47	118	8	0	26.5	34.7	103.8	2360	0.42	1.98	60	1	7	Cy	48.5	154.86	0.62	199
57	[112]	77.1	17.7	0.6	1.2	334	1329	633	58	58	13	41.23	26.4	20.1	111.7	2454	0.73	3.99	90	0.33	7	Cu	61.3	151.05	0.67	60
58	[113]	50.5	26.6	2.1	13.8	408	1201	647	41	103	14	16.5	27.6	30.3	102.6	2417	0.42	1.85	60	1	28	Cy	70.1	141.50	0.72	15
59	[114]	50.3	26.3	2.3	13.6	408	1201	647	41	103	14	16.5	27.6	30.3	102.6	2417	0.42	1.87	60	1	28	Cy	70.2	141.50	0.72	14
60	[115]	38.7	20.8	26.6	5.3	460	976	496	77	153	10	75.5	40.4	45.1	220.0	2237	0.70	1.98	60	1	28	Cy	39.1	206.78	0.55	868
61	[116]	51.3	30.1	8.7	4.6	400	950	850	57	143	12	48	36.5	42.5	169.0	2448	0.50	1.75	70	2	28	Cu	44.7	188.65	0.58	612
62	[117]	59.2	24.4	2.2	7.1	500	1150	575	64	161	14	0	43.2	47.2	134.5	2450	0.58	2.39	70	2	28	Cu	51.6	207.47	0.59	555
63	[118]	50.7	28.8	2.4	8.8	388	1293	554	45	113	14	0	28.8	30.1	99.1	2393	0.42	1.73	100	3	28	Cu	37.0	187.90	0.56	785
64	[119]	53.4	21.6	3.3	6.9	410	1100	590	N/A	161	N/A	0	40.9	47.4	137.2	2326	0.76	2.56	70	1	28	Cu	38.0	198.41	0.56	843
65	[120]	49.4	32.1	4.8	5.2	564	733	599	1332	125	N/A	0	44.6	32.2	224.6	2197	0.41	1.46	75	0.67	28	Cu	33.5	334.49	0.46	1071
66	[121]	58.2	25.1	2.9	4.6	338	1013	506	39	96	8	0	14.7	26.6	93.7	1991	0.29	2.24	20	0	28	Cu	27.8	114.77	0.62	233
67	[122]	17.6	36.4	10.6	12.4	350	1250	650	41	103	10	0	24.7	30.6	88.6	2394	0.32	0.62	20	0	28	Cu	19.0	125.13	0.59	521
68	[123]	48.3	30.5	2.8	12.1	381	1294	554	49	140	8	0	30.1	41.2	117.7	2418	0.43	1.65	60	3	28	Cy	75.0	176.31	0.71	19
69	[124]	51.7	29.1	8.8	4.8	350	1200	645	41	103	10	35	24.7	30.6	123.6	2374	0.40	1.77	65	1	28	Cy	58.9	141.85	0.67	61
70	[125]	59.7	28.4	2.1	4.6	314	1204	648	49	123	10	0	31.6	42.7	98.4	2339	0.58	2.20	100	1	7	Cu	60.0	174.29	0.64	124
71	[126]	60.1	26.5	4.0	4.2	500	935	765	N/A	300	N/A	0	29.8	86.9	183.3	2500	0.37	2.48	27	0	28	Cu	16.2	152.75	0.55	849
72	[127]	37.3	14.9	17.9	16.5	528	1155	495	N/A	72	N/A	150	36.7	33.1	152.2	2400	0.77	2.49	23	0	90	Cu	60.8	60.91	0.79	2
73		37.3	14.9	17.9	16.5	616	1068	458	0	84	N/A	175	42.8	38.6	177.5	2400	0.77	2.49	23	0	90	Cu	54.3	66.82	0.75	3
74		37.3	14.9	17.9	16.5	528	1155	495	0	72	N/A	150	36.7	33.1	152.2	2400	0.77	2.49	23	0	28	Cu	46.1	60.91	0.74	9
75	[128]	49.3	22.7	3.1	16.0	400	1209	651	11	129	N/A	0	27.2	38.1	74.7	2400	0.49	2.20	60	1	90	Cy	61.9	111.21	0.72	17
76	[129]	46.2	26.4	0.7	3.2	408	1346	612	41	103	N/A	0	24.2	30.3	89.5	2510	0.37	1.72	60	1	28	Cu	52.4	142.69	0.65	115
77	[130]	70.3	23.1	0.2	1.4	409	909	686	1595	204	N/A	10	68.0	60.0	215.1	2347	1.18	3.13	80	1	540	Cy	39.2	302.09	0.49	1036
78	[131]	62.8	16.7	6.4	7.4	500	1185	532	1718	122	N/A	0	28.8	33.6	120.1	2400	0.57	3.54	80	1	28	Cu	79.8	181.80	0.74	8
79	[132]	55.8	30.3	4.1	3.9	500	1113	603	1716	N/A	N/A	0	45.0	0.0	163.1	2424	0.49	1.57	60	1	182	Cu	59.8	325.10	0.55	907
80	[133]	39.6	17.5	16.6	18.1	560	936	624	1560	200	N/A	0	40.1	57.4	182.5	2400	0.67	2.42	20	0	28	Cu	71.9	218.29	0.66	75
81	[134]	49.0	31.0	5.0	3.0	475	1253	539	1792	78	N/A	0	16.1	22.5	80.4	2386	0.18	1.47	25	0	28	Cy	35.9	119.07	0.63	169
82	[135]	49.5	27.5	3.7	7.4	719	729	583	1312	N/A	N/A	0	128.7	0.0	275.4	2435	1.07	1.53	60	2	14	Cu	17.5	594.33	0.36	1135
83	[136]	65.6	26.5	0.3	5.5	298	1377	590	1967	96	N/A	0	25.6	33.7	74.8	2399	0.53	2.46	60	1	28	Cu	32.4	132.89	0.61	329
84	[137]	43.4	26.2	5.4	17.4	300	1321	622	60	90	12	11	29.4	26.5	105.1	2404	0.62	1.69	70	1	7	Cy	53.1	163.76	0.63	175
85	[138]	47.8	24.4	2.4	17.4	408	1201	647	41	103	8	0	23.1	30.3	90.6	2400	0.38	1.92	90	1	90	Cy	70.7	150.63	0.71	18
86	[139]	66.8	15.5	5.0	6.8	460	1200	540	1740	133	N/A	0	24.3	36.9	138.9	2400	0.56	4.11	80	1	3	Cu	37.7	193.47	0.56	819
87	[139]	66.8	15.5	5.0	6.8	460	1200	540	1740	133	N/A	0	28.5	36.9	134.7	2400	0.66	4.11	80	1	28	Cu	50.6	193.47	0.59	453

Notes: Cu = cube, Cy = cylinder specimen. In this table, the combination with rank 1 is optimal, indicating that it surpasses other combinations based on ranking criteria. The top 1–10 ranks using the GRA grade are highlighted in green, showing the best output, a combination of FA-GPC with higher strength and lower CO₂ emissions. The most efficient rank values or the best GRA grade values from each source are highlighted in green pink and blue color shows output response values. Yellow color represents cited references and orange color shows input variables.

presents the most efficient FA-GPC combinations, achieving the maximum strength while minimizing CO₂ emissions. It includes the best combinations from each reference, with green indicating excellent combinations and pink indicating good conditions below a value of 1. GRG was calculated by averaging the sum of the Grey relational coefficients (GRC). Higher GRG values indicate optimal FA-GPC parameters. The top 1–10 optimal GRG values across 83 references are highlighted in green, whereas the best values for each combination aiming for higher strength and lower CO₂ footprint are in pink. After determining the GRG, an ANN model was applied to further improve the GRC values and optimize the mix design proportion for FA-GPC to achieve higher strength and lower CO₂ emissions.

4.2. Artificial Neural Network (ANN) Model

Artificial neural networks (ANNs) [42] activate information processing. These consist of interconnected neurons that work simultaneously to solve specific problems. Figure 6 shows a three-layer ANN with 22 input variables (X1 to X22) and an output neuron representing the Grey rational coefficient grade. A backpropagation algorithm was employed to train the input and output values using an experimental dataset comprising more than 1000 samples. ANNs model the relationship between the input variables and the Grey rational grade to increase the compressive strength and reduce CO₂ emissions. The architecture includes multiple layers, using optimization techniques to enhance learning and generalization. The correlation coefficient (R²) was primarily utilized to assess model performance [140]. Consequently, this study also incorporates the mean absolute deviation (MAD), root mean squared error (RMSE), root absolute squared error (RASE), and relative squared error (RSE). The model’s optimal calibration is evidenced by its elevated R-value and low values for RMSE, MAD, RSE, and RRMSE. To have a strong relationship between the two sets of data, the R-value needs to be higher than 0.8 (the optimum value (R = 1) [141].

5. Results and Discussions

5.1. Results of Artificial Neural Networks

Figure 7 shows the model’s predictive performance for “GRC-grade” using scatter plots and key metrics. The top scatter plots show actual versus predicted values closely aligned along the 45-degree line, indicating a strong correlation and high R² value. This suggests that the model captures a significant variance in GRC grade values, with minor prediction errors around 0.6 to 0.8. The bottom plots show residuals against predicted values, mostly within −0.05 to 0.05, indicating minimal bias, although residuals spread slightly for higher predictions (above 0.7), up to 0.15, indicating reduced accuracy in this range.

The performance metrics of the ANN model are presented in

Table 4. Performance parameters of the ANN model.

Statistical Measures	Training	Validation
R-square	0.984	0.918
Root absolute square error (RASE)	0.011	0.021
Mean absolute deviation	0.005	0.012
−Log likelihood	−2790.66	−930.14
Sum of squared error	0.100	0.1526
Sum frequency	788	338

. In the training set, the model had an R² value of 0.984 with low errors (RASE, 0.0112; MAD, 0.0055; SSE, 0.1001), indicating a strong fit. In the validation set, R² decreased to 0.919 with higher errors (RASE: 0.0212, MAD: 0.012, SSE: 0.1526), suggesting overfitting and reduced accuracy of the new data. The reported R² value of 0.919 corresponds to the validation set, while the training R² is 0.984, indicating overfitting. To mitigate this, dropout regularization and 10 k cross validation were employed, although further refinements are required to enhance generalization. These observations align with the residual analysis, showing variability for higher GRC grade predictions. Overall, the model performs well, but requires refinement to improve generalization across all GRC-grade ranges.

5.2. Prediction Profiler Using ANNs

The prediction profiler graph illustrates the effects of various factors on the GRC grade, highlighting the relationships between the material components and the predicted values (Figure 8). For SiO₂, an initial increase decreased the GRC grade, indicating an inverse relationship. Al₂O₃ had a mild positive effect, stabilizing at higher levels. CaO strongly increases the GRC grade before leveling off, while Fe₂O₃ contributes positively, but with diminishing returns. MgO showed a slight negative effect, reducing the GRC grade as its content increased. Fly ash increased the GRC grade up to a certain point and then stabilized, indicating a range-dependent positive contribution.

C-Agg and density have a minimal impact, whereas F-Agg shows a small positive effect. NaOH significantly increased the GRC grade, whereas Na₂SiO₃ and molarity negatively affected it. The added H₂O and Na₂O were positively correlated with the GRC grade, while the Si/Al ratio and temperature showed slight negative effects. The Na/Al ratio had a slight positive impact, and age-related factors, such as HC days and curing, stabilized or slightly improved the grade. Lastly, ‘Specs’ initially increases the GRC grade but tapers off, indicating a diminishing positive influence. Key contributors include CaO, NaOH, and fly ash, whereas molarity and Na₂SiO₃ require careful management because of their negative effects.

5.3. Comparison Results of GRA vs. ANNs with Optimal Parameters

Figure 9 illustrates the factor values influencing the GRG in both the GRA and ANN models. The x-axis represents key factors such as chemical composition (SiO₂, Al₂O₃), mix proportions (fly ash, NaOH), and formulation ratios (Na/Al, Si/Al), while the y-axis displays the corresponding numerical values. The solid black line represents the GRA results, while the red dashed line indicates the predicted values from the ANN model. The ANN model optimized F-Agg and NaOH for a higher GRG, slightly reduced SiO₂, and maintained similar density values to the GRA model.

The environmental impact analysis shown in Figure 10 compares CO₂ emissions with compressive strength. This study’s findings align with and extend existing research on optimizing geopolymer concrete (GPC) for both strength and sustainability. Specifically, the ANN model achieved a significant reduction in CO₂ emissions, decreasing from 252.09 kg/m³ (GRA grade rank 1 [72]) to 55.0 kg/m³, representing a 78% reduction, while maintaining a strength of 90.9 MPa. This result is significantly better than previous studies. For example, Turner & Collins [24] reported that conventional GPC emits around 80% less CO₂ than ordinary Portland cement (OPC) concrete, though with varying strengths, ranging from 30–70 MPa, depending on the mix design. Similarly, Provis & van Deventer [142] have suggested that optimizing alkali activators could reduce CO₂ emissions by 60–70%, but balancing strength with emissions remains a challenge. Davidovits [143] found that high-calcium fly ash-based GPC could achieve 50–60 MPa with around 150 kg/m³ CO₂, but further optimization was required for ultra-high-performance applications.

In contrast, the ANN model in this study not only achieved higher strength but also drastically reduced CO₂ emissions, surpassing the benchmarks set by these earlier studies. This superior optimization leads to improved sustainability and enhanced structural performance. For instance, in a 1000 m³ project, the ANN model could save 197 tons of CO₂ emissions compared with the GRA model. By adjusting key factors such as NaOH, molarity, and fly ash, the ANN model supports greater resource efficiency and contributes to a circular economy. This provides a scalable and sustainable solution for reducing the environmental impact of geopolymer concrete without compromising its performance.

Regarding the strength vs. sustainability trade-off, the ANN model in this study broke this typical trade-off by achieving a compressive strength of 90.9 MPa while drastically cutting emissions, which outperforms most previous studies. For instance, Rangan [144] achieved around 70 MPa strength with 180 kg/m³ CO₂ but noted that increasing strength often required more alkali activators, which in turn raised emissions. Nath & Sarker [145] reported GPC strengths between 60–80 MPa with emissions of around 200 kg/m³ CO₂, showing the trade-off between strength and sustainability. Liu et al. [146] optimized GPC using machine learning, reaching 86.51 MPa with 251 kg/m³ CO₂, (59.87% reductions) but did not explore further reductions in emissions. In contrast, the ANN model in this study intelligently adjusted key mix proportions such as NaOH, fly ash, and molarity, achieving ultra-high strength while maintaining a significant reduction in CO₂ emissions, thus overcoming the common strength–sustainability trade-off.

In terms of machine learning applications in GPC optimization, the ANN model outperformed the Grey rational analysis (GRA), achieving higher accuracy (R² = 0.984 in training, 0.918 in validation) in predicting the optimal mixes. Previous studies have used machine learning for GPC optimization, such as Arifeen et al. [34], who applied different machine learning models, including ANNs, for strength prediction with an R² of 0.91 but did not integrate CO₂ emissions into their model. Ansari et al. [147] utilized GRA for sustainable GPC but lacked AI-driven optimization, which limited their ability to reduce emissions. Jiang et al. [148] combined ANNs with genetic algorithms to achieve around 50.70 MPa strength with 134 kg/m³ CO₂, but their model was less efficient than the approach presented in this study. This study’s integration of GRA for multi-objective ranking and ANNs for precise prediction offers a more robust framework for sustainable GPC design, providing a significant step forward in the field.

5.4. Life-Cycle Assessment (LCA) Comparison of GRA vs. ANNs

Figure 11 presents a 20-year life-cycle CO₂ emission study comparing the GRA and ANN models. In the ANN model, CO₂ emissions are reduced from 252.09 kg/m³ (the best GRA grade rank 1 [72]) to 55.0 kg/m³, representing a substantial 78% reduction. Over a 20-year period, the ANN model saves approximately 3941.80 tons of CO₂ per 1000 m³ of annual usage, surpassing the GRA model and aligning with global sustainability goals. The ANN model utilizes industrial by-products such as fly ash, which reduces the use of high-carbon cement, decreases waste, and helps preserve resources. This approach maximizes resource efficiency, reduces costs, and maintains strong structural performance while providing a scalable, environmentally friendly solution for sustainable construction. By improving supply chain resilience and reducing environmental footprints, the ANN model represents a significant advancement in low-carbon concrete technologies.

Previous life-cycle assessment (LCA) studies have shown that GPC can reduce CO₂ emissions by 50–60% compared with ordinary Portland cement (OPC) over its lifecycle, as noted by Habert et al. [149]. Furthermore, Habert & Ouellet-Plamondon [48] have reported that high-strength GPC (~80 MPa) still emitted around 200 kg/m³ CO₂, indicating the need for improved mix designs. In contrast, the ANN model not only exceeds these benchmarks but also offers a scalable, low-carbon solution that aligns with international sustainability targets, such as the United Nations Sustainable Development Goals (SDGs) and the Net-Zero 2050 initiative [150].

In summary, the findings of this study demonstrate how the ANN model optimizes both strength and sustainability. By achieving a remarkable reduction in CO₂ emissions while maintaining ultra-high strength, this study sets a new benchmark in geopolymer concrete design, offering an environmentally sustainable and performance-efficient solution for the construction industry.

6. Limitations and Future Research Directions

This study demonstrates the feasibility of fly ash-based geopolymer concrete (FA-GPC) as a sustainable alternative to conventional concrete by integrating Taguchi–Grey relational analysis (GRA) with artificial neural networks (ANNs). However, there are several limitations and avenues for future research. Future research will focus on conducting a comparative analysis of various machine learning models to evaluate their generalization capabilities and identify the most effective techniques. Additionally, further exploration of advanced regularization methods and hyperparameter tuning will be pursued to enhance the model’s robustness and predictive performance across diverse datasets. In addition, this study focused on the relationship between aggregate content (kg/m³) and GRA grade, as a result, it did not investigate the impact of aggregate type, gradation, or size, which are critical for strength development in FA-GPC. Future research should explore these factors in greater detail to enhance the understanding of how aggregate characteristics contribute to the performance of FA-GPC and optimize its mix design for better structural outcomes. Furthermore, sample scenarios and experimental validation have not been conducted in this study, but future research will include real-world application scenarios and experimental validation to further enhance the practical applicability and engineering contribution of the model.

7. Conclusions

This study demonstrates the feasibility of fly ash-based geopolymer concrete (GPC) as a sustainable alternative to conventional concrete through an integrated Taguchi–Grey relational analysis (GRA) and artificial neural network (ANN) approach. Analysis of 1126 datasets from 82 studies reveals that ANN-optimized mixtures achieve a 78% reduction in CO₂ emissions (55.0 kg/m³ vs. 252.09 kg/m³ for GRA grade rank 1) while maintaining high compressive strength (90.9 MPa). The optimized formulation maximizes fly ash utilization, reducing cement dependence and aligning with global decarbonization goals. Life cycle assessment indicates potential savings of 3941 tons of CO₂ over 20 years for projects utilizing 1000 m³ annually. The hybrid GRA-ANN framework provides a data-driven alternative to trial-and-error mix design, offering a scalable solution for sustainable construction. While limited to specific material combinations, this work establishes a methodological foundation for future research to expand material variability and explore additional machine learning techniques. These findings advance both academic knowledge and practical applications in low-carbon building materials, supporting the construction industry’s transition toward carbon neutrality.