Evaluating 28-Days Performance of Rice Husk Ash Green Concrete under Compression Gleaned from Neural Networks

Cement manufacturing and utilization is one of the majorly responsible factors for global CO 2 emissions. In light of sustainability and climate change concerns, it is essential to fnd alternative solutions to reduce the carbon footprint of cement. Secondary cementitious materials (SCMs) are helpful in reducing carbon emissions from concrete. One such solution is the use of agricultural waste as SCMs to reduce carbon emissions from concrete. Especially rice husk ash (RHA) is a silica-rich, globally available agricultural waste material. Te compressive strength (CS) of concrete is important and is used to evaluate the material’s strength and durability. Predicting CS using a laboratory method is a costly, time-consuming


Introduction
Concrete is the second most used material in the world after water. Cement is used as a binding material in concrete because of its benefts in various properties such as durability, stifness, strength, fre resistance, density, porosity, and thermal resistance. Among these properties, compressive strength (CS) is the foremost essential property of the concrete, as it strongly afects the safety and serviceability of the structural members. Te main constituents of concrete are binding material (cement), fne aggregate, coarse aggregate, water, and admixtures.
Tese substances and their mixes have an impact on the CS of concrete, including size of aggregate, water-tobinder ratio, composition of waste, and variety of binder [1].
For the past few years, India has been booming its construction industry by introducing mega-construction projects. Increasing construction development also has increased the production and usage of cement [2]. Cement is a substance which uses nonrenewable sources and is responsible for carbon emissions [3]. Around 50% of greenhouse gas emissions are done by the construction industry around the world [4]. Te construction sector in India raised its CO 2 emissions by 99% from year 1990 to year 2020 and China by 53%, whereas the United States has reported a reduction in carbon emissions from the construction industry by 4% from year 1990 to 2020, according to a report of the European Union [5]. It is understandable that India is undertaking massive construction projects because it is a developing nation but given that it also emits twice as much carbon today as it did thirty years ago, this data does not support the notion that the globe is going toward carbon emission. As the second-largest country by population [6], India is doubling its carbon emissions. To solve this problem, researchers are fnding ways to use waste material in the production of cement as a secondary cementitious material (SCM) to reduce carbon footprint. Fly ash (FAH), silica fume (SF), blast-furnace slag, limestone powder (LP) [7], and metakaolin [8] are some industrial waste materials, and bagasse ash (BA), bamboo leaf ash (BLA), corn cob ash (CCA), rice husk ash (RHA), palm oil fuel ash (POFA), wood waste ash (WWA) [9], and wheat straw (WS) [10] are the agricultural waste materials that are being used in the cement industry as SCM.
In the past 20 years, more research works have been done on the waste material related to the farming industry. Te waste generated from the farming industry has an overwhelming amount of minerals present in agricultural waste [8]. RHA has high pozzolanic activity due to the availability of high silica (SiO 2 ) content which is around 80%-95% [11]. Tat is why it becomes a promising waste material used in concrete as SCM. A visual representation of raw rice husk is shown in Figure 1(a), where Figure 1(b) shows the RHA before and after grinding. Te pozzolanic activity of RHA depends on the silica crystallization phase, silica content, and size and surface area of ash particles [13]. Due to its high activity, RHA furthermore provides mechanical benefts to concrete such as an increase in CS, tensile strength, fexural strength [14], durability [15], corrosion resistance [13], a reduction in water absorption [16], and better resistance to sulphate attracts [17]. Te use of RHA in concrete also improves the environment by lowering CO 2 emissions from the concrete [18] and by lowering the energy required to create SCM, as cement requires 31 times more energy to produce [19]. Tus, using RHA concrete has environmental and economic benefts, as it helps reduce the carbon footprint [18], requires 31 times less energy for production [19], and consumes less quantity of conventional cement for production. Te use of RHA-based concrete in larger construction projects may present some difculties.Among these challenges are ensuring a consistent and reliable supply of RHA, as well as potential issues with quality control and material variability. Furthermore, there may be concerns about RHA-based concrete's long-term durability and performance. Figure 2 illustrates the diferent advantages of using RHA in concrete.
RHA-based concrete can also be used in combination with diferent agricultural and industrial waste materials. According to an article published by Jayanthi et al. [20], the incorporation of micronized biomass silica (MBS) from rice husk, an agricultural waste material, and ground granulated blast-furnace slag (GGBS), an industrial waste material, into geopolymer concrete (GPC) enables the development of high-strength concrete with reduced environmental impact.
According to the existing research, RHA primarily infuences concrete CS in a favourable way. Te experiment work was done by Siddique et al. [21] for an optimum concrete CS of 32.8 MPa, and the results show that the CS of concrete increases by 8.7%, 10%, and 13.4% with respect to control concrete at the ages of 7, 28, and 56 days. Te results depict that a 10% replacement of RHA is optimum for concrete CS at all ages. Another study was conducted by Ganesan et al. [22] and the results show that 30% is the optimum RHA content for concrete in M20 mix. By comparing these two results from the literature, it can be concluded that for a higher concrete mix, less RHA content should be used, but 30% RHA content for residential structural concrete of M20 mix can be used. In research done by Kishore et al. [23], their results show that 10% is the optimum RHA content for M40 and M50 concrete mix. Tis proves that less amount of RHA should be used for higher grades of concrete.
In 2021-2022, 509.26 million metric tons (MMT) of milled rice was produced all around the world. China contributed around 148.3 MMT and India contributed around 124.37 MMT to the global production of rice, which is around 53.54% of the world rice production only from these two countries whereas Bangladesh, Indonesia, and Vietnam produced around 34.6, 34.5, and 27.38 MMT, respectively [24]. Figure 3 shows a chart of top 15 riceproducing countries [24]. In India, Punjab produced around 12.5 MMT (10%) in 2021-2022 [25]. While producing milled rice from rice paddy (rough rice), most rice varieties are composed of around 20% Rice Husk, 11% bran layers, and 69% starchy endosperm (total milled rice) [26]. Tus, it can be assumed that around 20% of rice husk (147.67 MMT of world rice) can be used in the construction industry as SCM, which can reduce around 20% of carbon emissions done by cement around the world. Tus, converting rice husk to rice husk ash and using it for the construction industry is benefcial for both the environment and the construction industry.
RHA can be manufactured by combustion and grinding of rice husk in a simple furnace with gasoline placed under the air ducts inside the furnace. Researchers make their own furnaces to produce RHA, but Zain et al. [13] gave a generalized method and procedure to produce RHA. Tey concluded that 500-600°C is the average burning temperature for RHA with 30 minutes of combustion with fre duration and an air supply of 60 minutes and chilling duration of 2 days. After the burning, grinding is also important, as particle size has a major infuence on the pozzolanic activity [13]. According to BS 3892-1 [27], particles of RHA retained on a 45 µm sieve should not be more than 12%; therefore, it is important to fnely grind RHA after burning. Details about the furnace and the detailed procedure can be found in the research article by Zain et al. [13].
Concrete is a diverse, heterogeneous material. To accurately predict the CS of concrete is a very challenging task. Te CS of the concrete is often assessed within the laboratory during the tests by crushing the cubes and cylinders, after the desired time of casting the sample. However, laboratory tests nowadays are uneconomical and inefcient because they are time-consuming and expensive process. Recently, with the advancement of artifcial intelligence (AI) and machine learning (ML) algorithms have been utilized to solve various complex problems in diferent sectors [28][29][30]. ML techniques can be useful in predicting the mechanical properties of concrete at diferent efciency levels. ML techniques can also be used to optimize concrete mix designs, so one can save time and resources when doing trial mixes [31]. Different ML techniques like clustering, classifcation, and regression are currently in use in diferent sectors of engineering [1,[32][33][34][35].
Amlashi et al. [3] developed three ML models (multivariate adaptive regression spline (MARS), artifcial neural network (ANN), and M5P model tree) with 909 datasets samples of 1 to 365 days age of CS of concrete containing RHA. In this study, cement, coarse aggregate, fne aggregate, water, superplasticizer, and RHA were used as input parameters, and CS was the output parameter. Te parametric analysis shows that coarse aggregates had the most infuence on the CS of concrete after water. Te coefcient of determination (R 2 ) of the ANN, MARS, and M5P model trees was 0.9665, 0.9105, and 0.8785, respectively. ANN model had the best accuracy and less error than MARS and M5P model tree ML techniques.
Iqtidar et al. [36] utilized ML algorithms to predict the CS of concrete containing RHA. Four ML models, adaptive neuro-fuzzy inference system (ANFIS), linear regression (LR), multiple nonlinear regression (MNLR), and ANN were used to predict the CS of concrete, and it was found that ANN model has the best accuracy, with an R 2 value of 0.98. Islam et al. [37] used regression analysis (RA) to predict the      [39] applied GEP, MNLR, and LR machine learning algorithms on a dataset of 250 samples and found that the GEP model was the most efective ML algorithm to predict 1-90 days CS of RHA concrete, while LR was the least accurate ML algorithm. Sarıdemir et al. [40] developed two GEP models for estimating the CS of concretes containing RHA at the ages of 1, 3, 7, 14, 28, 56, and 90 days. Both GEP-I and GEP-II models performed well, as their R 2 values were 0.9629 and 0.9437, but GEP-I had slightly more error than GEP-II. Te outcome of the results shows that the GEP-I model was more accurate than the GEP-II model because of its higher R 2 value.
Asteris et al. [7] developed an ANN model to predict that 28 days CS of concrete contains fy ash (FAH), limestone powder (LP), granulated blast-furnace slag (GBFS), rice husk ash (RHA), and silica fume (SF) as chemical admixtures. Te R value of the developed ANN model comes out to be 0.9825 by using 2 hidden layers with 5 and 4 neurons. Another ANN model was developed by Asteris and Kolovos [41] with the same input parameters and waste materials as Asteris and Kolovos [41], and the sensitivity analysis shows that RHA was the most infuenced waster material among the others [42]. Diferent ML models developed by various researchers to estimate the CS of concrete with diferent waste materials are shown in Table 1.
Te use of RHA-based concrete in construction projects can provide several potential environmental and economic benefts. Firstly, it can reduce the environmental impact of the construction industry by reducing the need for Portland cement, which is a major source of CO 2 emissions. Secondly, RHA-based concrete can use waste materials that would otherwise be disposed of in landflls, reducing the environmental burden associated with waste disposal. Tirdly, the use of RHA-based concrete can lead to cost savings due to the lower cost of RHA compared to Portland cement, as well as improved durability and reduced maintenance requirements. Overall, the use of RHA-based concrete can ofer a sustainable and cost-efective solution for construction projects.
Te motive behind this study is to reduce the carbon footprint of the construction industry and this study will help the researchers estimate CS of the RHA-based concrete is good enough to be used in their project, as they can estimate the CS before doing any laboratory work. Te ML-based ANN model is robust, cost-efective, and accurate in predicting the compressive strength of RHA-based green concrete.

Research Significance
Te construction sector is responsible for 50% of global carbon emissions [4]. SCM helps reduce the carbon footprint by partially replacing cement. RHA is an agricultural waste material that contains a good amount of silica for pozzolanic reactivity with cement, so RHA can be used as SCM. Te CS of concrete is a characteristic that evaluates the strength and durability of concrete. However, predicting CS by laboratory method is a tedious and time-consuming process. Te solution to this problem is based on MLbased predictive models. In this study, an ANN-based ML prediction model has been developed to predict the CS of concrete cubes as well as cylinder specimens. Tis study developed a precise ML model for estimating the 28-day CS of RHA concrete. Tis study will help reduce the time needed to fnd the CS of RHA concrete and also help save the environment as RHA-based concrete produces less CO 2 than conventional concrete.

Data Bank.
In the present study, experimental data from 407 samples were collected from the available literature. After collecting the data from the literature, fltration of the data was conducted to remove the outliers from the databank. After the fltration process, only 212 samples were left, as shown in Table 2. It has seven input parameters, namely, cement (C) (kg/m 3  Te summary of the database and frequency distribution histogram obtained from the literature are shown in Table 3 and Figure 4, respectively. Te minimum, maximum, average, and standard deviation values of the collected database are shown in Table 3.
Te frequency distribution gives information about the minimum and maximum range of the samples and the frequency of datasets. It can be seen from Figure 4      Advances in Materials Science and Engineering ype of sample is also considered as an input parameter to increase the accuracy of the model. Two types of concrete samples, cube and cylinder, were considered, represented by 1 and 2, respectively.

Preparation of Data.
Prior to handling the data in ML algorithms, the data should be normalized within a certain range so that it becomes unitless. Te goal of normalisation is to transform the datasets into a common unit without altering fuctuations in the value ranges. Te commonly used normalisation ranges are 0 to 1, −1 to 1, etc. However, in this study, −1 to +1 normalisation range was used, and the expression of the same is shown in the following equation [107]: where x normalized is referred to normalized output, x is the value to be normalized, and x min and x max are the minimum and maximum values in the selected dataset, respectively. After that, the dataset was divided into three parts. Tis dataset was randomly divided into three parts, namely, training (80%), testing (10%), and validation (10%) and the same has been used to assess the performance of the network. Te methodology adopted to develop this model is shown in Figure 5.

Performance Criteria.
To assess the performance of the neural network model, performance criteria, such as MAE (mean absolute error), MAPE (mean absolute percentage error), NS, a-20 index, R (correlation coefcient), and RMSE (root mean square error), were calculated. An R value closer to 1 indicates a superior ftting result, while an R value greater than 0.85 indicates good ftting results [107]. When the values of MAE, MAPE, and RMSE are 0, the performance of the ML models will be greater [108]. Te numerical equations of R, RMSE, MAE, MAPE, NS, and a-20 index are shown in equations (2) to (7) [109][110][111].
where N is the number of experimental datasets, E i is the experimental value of the CS at the i th level, E is the mean of the experimental values, P i is the predicted value of CS of concrete values at the i th level, and P is the mean of the predicted values.
where the value of m20 is the number of samples with the ratio of experiential by predicted values of CS of concrete between the ranges of 0.8 to 1.2.

Artificial Neural Network
Artifcial neurons represent the structure of the human brain and mimic the properties of biological neurons composed of inputs. Each input represents the output of a diferent neuron. Te inputs used to solve the problem are multiplied by appropriate weights, and these weighted inputs are summed with the bias value [53]. Te input sums are processed in hidden layers using transfer functions such as linear, tan-sigmoid, and log-sigmoid [112]. Information processed by the transfer function is sent along the output layer as the desired result. A structure in which neurons are arranged in layers and a pattern of connections exists between neurons in each layer is called network architecture [53].  Advances in Materials Science and Engineering Figure 6 shows the structure of an artifcial neuron model. Feed-forward network (FFN) is usually used with the backpropagation technique for training networks where the error obtained at the output layer is sent back to the input layer and hidden weights and biases are updated accordingly to reduce the errors at the output layer. Te main aim of the FFN process is to reduce the overall error between the observed and estimated values by adjusting the weights, and these weights are combined and processed through an activation function and released to the output layer [53].

Development of the ANN Model.
To develop an ANN model for predicting the 28-day CS, the input parameters are already defned in Section 3.1. In this study, seven input parameters are considered in the input layer with one hidden layer and one output layer. Previous research has demonstrated that one hidden layer is adequate to achieve the optimal ANN structure [113]. Te dataset was randomly divided into three parts: 80%, 10%, and 10% for training, validation, and testing, respectively. In this study, an ANN algorithm named Levenberg-Marquardt's (LM) backpropagation train the ML model. Te trial and error method was utilized to determine the ideal number of neurons in the hidden layer. Based on the R and MSE values, the fnal neuron was selected where the errors were less and the R value was the maximum. As seen in Figures 7(a) and 7(b), the hidden layer's neurons were set to range from 3 to 21.

Advances in Materials Science and Engineering
Every neuron is connected to the input and hidden layer using nodes. Te hidden layer receives inputs from the input layer in normalized form to calculate the weight values. Te output and hidden layers' transfer functions were the linear transfer function (purelin) and the hyperbolic tangent sigmoid (tansig), respectively. [1]. All the values of input are multiplied by the weight value, and after that biases are added as expressed in the following equation: where W I-H is the weight from input to hidden, I norm is the normalized input value, and B I-H is the biases from input to hidden layer. An activation function (tansig) (equation (9)) is applied to start the process.
After that, the output is obtained during the output process by using equations (10) and (11) to calculate the CS of the inputs provided.
In this study, the neuron with the highest R value and lowest MSE is number 14. Neuron 14 has been selected as the fnal neuron and used to predict the CS of RHA-based concrete.

Results and Discussion
Tis section describes the results of the developed ANN model and the relative importance of the input parameters.
Te frequency distribution of the training, testing, and validation dataset is also explained in this section. Te mathematical expression to predict the CS of RHA-based concrete is also expressed in this section.

Results of the ANN Model.
ANN trains data in three groups, training, validation, and testing. Figure 8 represents the accuracy and error distribution of the ANN prediction model with training (Figure 8(a)), validation (Figure 8(b)), and testing (Figure 8(c)) datasets. In Figure 8(a), the training dataset contains 170 samples that are lying in the error range of +20% to −20% and the value of R is 0.9928, as shown in Table 4. In Figure 8(b), the validation set has the R value of 0.9865 and all the values lie in the range of +20% to −20%.
Te R value of the developed ANN model is close to one, and the RMSE value is less than 5 MPa. Figure 9 shows the frequency versus error distribution of training, validation, testing, and all datasets. Tis indicates how many datasets fall inside the minimum or maximum error range. Figure 9(a) gives an idea of the training datasets, where 55.3% of the datasets lie in the −2 MPa to +2 MPa error range and only a handful of data have errors greater than −4 MPa to +4 MPa. Figure 9(b) shows the frequency-error distribution of validation datasets, where 71.4% of the datasets  Advances in Materials Science and Engineering also lie in the range of −2 MPa to +2 MPa. Figure 9(c) shows the frequency-error distribution of testing datasets, whose maximum error range is highest among others, about 85.7% of the datasets lie in the error range of −5 MPa to +5 MPa. Te frequency-error distribution of all datasets is shown in Figure 9(d), where 76.4% of the datasets have an error between −2 MPa and +2 MPa. Table 5 shows the statistical characteristics of the datasets used to develop the model, including the mean, standard deviation, minimum, and maximum.
Te infuence of the individual parameters on the CS of concrete is shown in Figure 10. Based on ANN predictions, the Cement (C) and coarse aggregates (CA) have a 16.99% and 16.94% efect on the CS of the concrete, respectively. Similarly, the efect on other input parameters such as fne aggregates (FA), water (W), rice husk ash (RHA), superplasticizer (SP), and type of concrete sample (S ty ) is 14.44%, 14.66%, 12.19%, 10.40%, and 14.35%, respectively. Te impact of C and CA on the CS of concrete is the greatest among all the inputs. Tis information can therefore be used to adjust the cement and coarse aggregate values in a more targeted manner when optimizing the mix design of RHAbased concrete to achieve the desired CS.
Te fnal formulation to predict the CS of RHA-based concrete at 28 days is expressed in the following equation:

Conclusions
In this study, an ANN-ML model has been developed to predict the CS of RHA-based concrete. Te model developed in this study is based on a diverse set of data, including seven input parameters demonstrated by experimental studies available in the literature. Te considered input parameters to develop the machine learning model are: C, FA, CA, W, RHA, SP, and S ty . Te results of this study are more consistent with experimental research. Te R values of the developed ANN-ML model for training, testing and validation datasets are 0.9928, 0.9545 and 0.9864, respectively, which is close to one. Te RMSE values of the training, testing and validation datasets are 1.65 MPa, 4.43 MPa, and 2.71 MPa. Te relative importance analysis shows that cement and coarse aggregate have a greater impact on the CS of RHA-based concrete. When the data for the essential input parameters is available, the ANN model can be utilized to reliably predict the CS of concrete. Tis will save time and eforts.
Whereas this study limits only focus on predicting the CS of RHA-based concrete at 28 days, further studies are needed to investigate the long-term behaviour of this material. Another potential limitation is the fact that this study only considers seven input parameters, and there may be other parameters or factors that could impact the CS of RHA-based concrete. For instance, environmental factors such as temperature and relative humidity could infuence the strength of the material over time.
Te potential direction for future research could be to investigate the impact of diferent types of agricultural waste on the mechanical properties of concrete under diferent environmental conditions. Experimental validation of the predicted results can be carried out in order to establish the reliability and accuracy of the developed ANN model.

Data Availability
Te data (rice husk ash-based concrete) used to support the fndings of this study are available from the authors upon reasonable request.

Conflicts of Interest
Te authors declare that there are no conficts of interest.