Ground Vibration and Air Overpressure Prediction Caused by Blasting Operations in a Gneiss Quarry Using a Gene Expression Programming Approach

Abstract Introduction Rock blasting is a fundamental operation in mining and civil engineering, but it generates ground vibrations and air overpressure that can affect nearby structures and communities. Among these effects, peak particle velocity (PPV) and air overpressure (AOp) are commonly used indicators for environmental control and regulatory compliance. Traditional empirical predictors, such as scaled-distance relationships, have been widely used for their simplicity and ease of interpretation. However, these approaches often rely on strong assumptions about the functional form of the relationship between blasting parameters and vibration response, which may not adequately capture complex site-specific conditions. In recent years, data-driven techniques have been increasingly explored to improve predictive performance. Linear statistical models, such as Multiple Linear Regression (MLR), offer interpretability but may suffer from multicollinearity and limited flexibility for modeling nonlinear interactions. Alternatively, evolutionary algorithms such as Gene Expression Programming (GEP) can generate explicit nonlinear equations without requiring predefined model structures, potentially capturing more complex relationships. Despite these advances, comparative studies of interpretable linear models and symbolic regression approaches remain limited, particularly when both PPV and AOp are considered simultaneously using field data from real blasting operations. This study aims to compare the performance of MLR and GEP in predicting PPV and AOp using a limited dataset of 31 detonations at a gneiss open-pit mine in southern Brazil. The objective is not to develop generalized predictive models but to evaluate how different modeling paradigms perform under data-constrained conditions and to assess their respective advantages and limitations in practical applications. The dataset showed notable dispersion in the response variables, as indicated by statistical measures such as standard deviation and the observed spread of the data distribution. Materials and Methods A methodology was proposed to predict the physical parameters associated with PPV and AOp. An engineering seismograph monitored shockwaves over a three-year period. The seismograph reports obtained were evaluated using SPSS Statistics IBM and Microsoft Excel to develop MLR models, and via GeneXproTools 5.0 to construct a shock wave prediction model using GEP. The MLR approach initially included multiple predictors but required variable reduction because of multicollinearity, resulting in simplified models primarily based on scaled distance relationships. In contrast, GEP was used to derive explicit nonlinear equations without prior assumptions about the functional form. Results Model performance was evaluated using the coefficient of determination (R²) and error metrics. For PPV prediction, the MLR and GEP models achieved R² values of 0.795 and 0.903, respectively. For AOp, the corresponding R² values were 0.557 and 0.824. The results indicate that GEP provided improved fitting capability for the available dataset, particularly for AOp. Conclusions GEP produced nonlinear explicit equations that achieved a higher goodness-of-fit than MLR, particularly for AOp prediction. Although GEP demonstrated improved performance, the results must be interpreted cautiously. The limited number of observations, the absence of external validation, and the site-specific nature of the data limit the robustness and transferability of the proposed models. Therefore, the main contribution of this study is an exploratory comparison of linear and symbolic regression approaches under constrained data conditions. Future work should focus on larger datasets, independent validation, and incorporating additional physically meaningful variables to improve model reliability and generalization.