Currently, sustainability in civil construction is a recurring theme and drives research into innovative and/or environmentally friendly materials. There is a need for sustainable techniques, focusing on the feasibility of additives or the use of materials considered innovative, Queiroz et al. (2025), and one of the objectives is the partial or total replacement of cement. However, according to Rebello et al. (2023), as a material with cementitious properties, cement remains almost unbeatable, as it exhibits the desired strength characteristics.

Soil-cement bricks stand out as an alternative, since, according to the Brazilian standard ABNT NBR 8491 (2012), at least 85% of their apparent volume consists of a cohesive mixture of soil, Portland cement, water, and, occasionally, additives. Silva et al. (2021) highlight that ecological bricks represent a sustainable solution, as they reduce dependence on limited natural resources and energy-intensive production processes. Grande (2003) indicates that soil presents itself as a widely usable and low-cost material, with a long history of application in construction and good performance in terms of mechanical strength. The Brazilian standard ABNT NBR 10833:2013 and the Technical Bulletin of the Brazilian Portland Cement Association (ABCP, 2000) provide technical guidelines, dosage recommendations, and parameters for the production of soil-cement elements, aiming to reduce cement content without compromising strength. Along these lines, the inclusion of additives is a common practice, and focusing on sustainability, the reuse of waste stands out as a relevant strategy. As indicated by Queiroz et al. (2025), which has been the subject of constant studies, some materials generated in the food industry are being tested for use in soil stabilization, among other applications. In this regard, there are studies on the use of rice husk ash, sugarcane bagasse ash (also used in concrete), açaí seed ash, peanut shell ash, and biomass ash from industrial boilers (Rocha et al., 2021; Garcez et al., 2024; Jordan et al., 2019; Sathiparan et al., 2023; Santos et al., 2022), and, among the residues generated in large quantities in Brazil, eggshells and spent coffee grounds are noteworthy. According to the Brazilian Coffee Industry Association (ABIC), in 2021 alone, more than 21.5 million coffee bags were consumed in Brazil, consequently generating a large volume of waste that is mostly disposed of directly. Mussatto et al. (2011) point out that this amount could be substantially reduced by incorporating such material into ecological construction elements.

Regarding egg consumption, waste generation is also abundant. According to the Brazilian Institute of Geography and Statistics (IBGE), in 2021 alone approximately 3.98 billion dozen eggs were consumed, following the same trend observed with coffee consumption, that is, producing large amounts of discarded material. As noted by Ladu & Morone (2021), incorporating eggshells into construction not only contributes to addressing environmental issues related to waste generation, but also has indirect benefits, since soil-cement bricks are manufactured without kiln firing processes, being produced solely through pressing.

In general terms, an Artificial Neural Network (ANN) is designed to work similarly to the human brain, and to achieve adequate performance, so-called artificial neurons are employed—namely, processing units that, via synaptic weights, are used to store acquired knowledge. Thus, they possess a high capacity to capture complex nonlinear relationships, exhibit flexibility with pre-processed data, and are adaptable to regression problems (Haykin, 2007).

Research shows the ability of ANNs to establish relationships between various input and output parameters. Tavares et al. (2020) developed an ANN model to estimate the compressive strength of concrete, using 1030 samples with strengths ranging from 2 MPa to 80 MPa, and the model achieved the best results, reaching a mean squared error (MSE) of 15.80 after 23 training epochs. In Brazil, Lorenzi et al. (2017) and Felix et al. (2018) also reported positive results when using ANNs to estimate mechanical properties and analyze the durability of reinforced concrete structures.

Based on these considerations and the absence of studies using machine learning techniques with eco-friendly bricks, this work aims to explore and evaluate the application of Artificial Neural Networks (ANNs) to optimize the dosage of components and predict the mechanical performance of reusing coffee grounds and eggshell in soil-cement mixtures, with an emphasis on compressive strength.

The tests were conducted at the Construction Materials and Structures Laboratory – LabMATE, at the Federal University of Paraná (UFPR) – Jandaia do Sul Campus, consisting of the production of soil-cement bricks and the development of an ANN algorithm for prediction and data analysis. The soil was collected from an experimental area belonging to the UFPR Jandaia do Sul Campus, with 350 kg of soil collected to determine particle size distribution according to ABNT NBR 10833:2012. Soil analyses were performed at the Agricultural Engineering Soil Laboratory on the Jandaia do Sul campus, following procedures outlined in the EMBRAPA Soil Analysis Manual (1997).

The results showed a high percentage of clay in the soil used, classifying it as type 3 soil (clayey texture soils, with a clay content greater than or equal to 35%), which is characteristic of the region. The soil composition was 18.60% sand, 67.97% clay, and 13.43% silt, contributing to greater cohesion between the soil particles. This can be considered positive, as it results in bricks with higher compressive strength. However, it may increase the water absorption capacity of the bricks, leading to greater susceptibility to moisture, which could negatively affect their long-term durability, especially in humid environments.

The waste materials used (eggshells and coffee grounds) were collected from locations on campus, the university restaurant, and the academic community. They were then taken to the Construction Materials and Structures Laboratory – LabMATE and stored in plastic trays exposed to the sun for drying. To produce the bricks, the solid components of the mixture —soil, cement, and residue— were properly homogenized, and water was gradually added until reaching approximately 0.44 kg, the amount necessary to achieve optimal moisture content, as indicated in the specific literature.

Four different mixes were tested for evaluation, varying the proportion of waste addition at 0%, 10%, 30%, and 50% soil replacement. Specimens were molded using concrete analysis molds (Figure 1), and after molding, curing was performed for 7 days, with water sprayed every two hours.

The data were subjected to analysis of variance and Tukey’s mean comparison test at a 5% significance level, based on a completely randomized experimental design, comprising a 3x3 factorial arrangement and processed using the SISVAR software.

A model was developed in the Spyder IDE environment (Scientific Python Development Environment), using Python 3.11.11 (64-bit) with graphical support provided by Qt 5.15.8 and PyQt5 5.15.9. After preprocessing the data, which included one-hot encoding for categorical variables and standardization for numerical variables to predict resistance, a multilayer neural network model (MLPRegressor) with two hidden layers (150 and 100 neurons), ReLU activation function, Adam optimizer, and a maximum number of iterations set to 1000 was trained, as presented in Equation 1 and represented in Figure 2.

Where y^ is the compressive strength estimated by the neural network; x is the input vector, consisting of 4 components (3 binary variables resulting from the one-hot encoding of the sample type (SC, SCBC, and SCCO), and the value of the additive concentration, previously normalized using the z-score method); w(¹) and b⁽¹⁾ represent the weight matrix and bias vector of the first hidden layer, respectively, which consists of 150 neurons; w(²) and b(²) correspond to the weight matrix and bias vector of the second hidden layer, which has 100 neurons; w(³) and b⁽³⁾ indicate the weight vector and bias of the output layer, which generates a single continuous estimate, the compressive strength in MPa; σ(.) is the ReLU activation function, applied individually to each neuron in the hidden layers.

The performance of the model was evaluated using the metrics Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). Based on the results of the descriptive analysis and the model’s predictions, the optimal combination of type and concentration that maximizes resistance was identified.

The Brazilian Standard NBR 10836 of 2013 states that the minimum compressive strength must be 2.0 MPa, and the absolute individual value must be at least 1.7 MPa.

The results of the analysis of variance (Table 1) showing significant differences between the variables analyzed: Soil-Cement (SC), Soil-Cement with Eggshell (SCCO), and Soil-Cement with Coffee Grounds (SCBC), as well as their interactions.

Figure 3 presents the meaning results considering the variables “Type” and “Concentration.”

Since this study utilized an Artificial Neural Network (ANN) developed to analyze mechanical strength data and predict such resistance based on two perspectives—sample type and concentration, or the variables “Type” and “Concentration”—predictive modeling is crucial for relating independent elements to the objective. Analyzing the results generated by the model is of great importance in terms of process efficacy. Thus, the Learning Curve is essential, and Figure 4 indicates the performance of the generated model considering different dataset sizes for training. Both curves converge relatively early (at low values), suggesting that the model was efficient in “learning” from the data used.

When comparing the actual and predicted values (Figure 5), one can observe how the model analyzes the data structure. Additionally, the predicted values follow the same trend as the actual values used to create the model. However, discrepancies are present in the graph, specifically near the higher resistance values, which can be explained by the lower density of samples in these value ranges, leading the model to adjust with less precision.

Figure 6 shows the behavior of the predictions along the order of data collection. Although the model can capture the temporal and sequential trends of the actual values, it fails to predict sharp peaks and sudden drops in resistance. This limitation can be attributed to the characteristics of the developed model (MLP network), and the results of the Artificial Neural Network are presented in Table 2.

The Mean Squared Error (MSE) was 1.75%, indicating that for the prediction of compressive strength, the model exhibits a low error. The Root Mean Squared Error (RMSE) of 13.21% suggests that the predicted values are aligned with the actual values, while the Mean Absolute Error (MAE) of 10.78% further demonstrates the good performance of the model. Additionally, the Coefficient of Determination (R²) of 88.21% confirms the model’s reliability.

These results corroborate Rocha et al. (2021), who pointed out that the replacement of cement with eggshell in appropriate proportions favors the increase in compressive mechanical strength.

The results obtained below the values indicated by the standard are consistent with the findings of Santos et al. (2022) in their study on the addition of biomass ash from industrial boilers. By testing the addition of 30%, 45%, and 65% ash, they found compressive strength values ranging from 0.68 MPa (30%) to 0.18 MPa (65%). Similarly, Jordan et al. (2019), in their study on the inclusion of sugarcane bagasse ash, tested the addition of 30% and 40% ash as a replacement for soil and obtained better compressive strength results, although still below the standard requirements, with values of 1.30 e 1.27 MPa for the proportions studied (30% and 40%, respectively). These results align with those found in this study. Additionally, the authors noted that the results may have been affected by the lack of ash treatment, a fact that also occurred in this experiment since the studied residues did not undergo any treatment.

In contrast to the findings of this study, Garcez et al. (2024) obtained promising results when testing the addition of açaí seed ash in soil-cement bricks, achieving compressive strengths ranging from 4 – 6.2 MPa for additions of 5% to 20% ash. On the other hand, Rocha et al. (2021), when analysing the influence of rice husk ash addition in soil-cement bricks, found compressive strength values lower than those of tests without ash addition, with reductions ranging from 12.64 – 26.78%, corroborating the findings of this experiment that the greater the ash addition, the lower the compressive strength of the soil-cement.

In general, studies typically use powdered materials rather than raw (“in natura”) materials, as was the case in the present study. This factor may have influenced the compressive strength results obtained.

Asimakopoulou et al. (2015) evaluated the use of Artificial Neural Networks (ANNs) to estimate soil resistance using experimental data, achieving a correlation coefficient (R²) of 99.89% between measured and estimated data. This highlights the effectiveness of ANNs in predicting resistance values, even though the R² value obtained in this study was 88.21%. In a similar vein, Agbemenou et al. (2024), studying the prediction of lateral pile resistance in predominantly cohesive soils using ANNs, found correlations of 85% between the data used for training the ANN, which is comparable to the results obtained in this study.

In applications related to concrete, Naderpour et al. (2018), analysing the prediction of shear strength in FRP-reinforced concrete beams using ANNs, observed an average model error of 9.72% in predicting the results, which is considered low and aligns with the findings of this experiment, where the mean error was 1.75%. Similarly, Yadollahi et al. (2016), studying the use of ANNs to predict optimal mixtures for concrete radiation shields, reported that, akin to the findings of this study, the ANN predictions showed a good fit with experimental results, with an error of approximately 1.7%, closely matching the 1.75% error obtained here.

Since the results are influenced by the evaluated parameters —in this case, the type of additive and its concentration— they align with the findings of Bal & Buyle-Bodin (2013), who studied the use of ANNs to predict drying shrinkage in concrete. They suggested that parametric studies could allow quantification of the effects of certain parameters using an ANN model.

Based on the results presented in this work, it is evident that neither of the waste incorporations meet the standard requirement of 2.0 MPa, contradicting the findings of Mussatto et al. (2011), who indicated that coffee grounds can improve the mechanical properties of bricks, making them lighter and more insulating. According to the ANN created, considering the data and mix proportions.

Based on the analyzed results, the following conclusions can be drawn: i) The addition of elements to soil-cement did not achieve the minimum value recommended by the Brazilian standard; ii) The ANN created identified a concentration of 30% eggshell-soil-cement (SCCO) as providing the best resistance, with a value of 1.186 MPa; iii) Given the values of MSE (1.75%), RMSE (13.21%), MAE (10.78%), and the coefficient of determination (R² = 88.21%), the results of the analysis using the ANN can be considered to have a good fit; iv) For the data analyzed, the ANN was unable to find a solution (mix/concentration) that met the minimum resistance specified by the standard.