Concrete compressive strength prediction represents a fundamental challenge in modern construction engineering, directly impacting structural safety, material optimization, and sustainable infrastructure development [1]. Accurate prediction of concrete strength enables engineers to optimize mix designs, reduce material waste, ensure structural reliability, and minimize construction costs while maintaining desired performance characteristics [2]. Traditional empirical methods for strength estimation, while historically valuable, often prove inadequate for contemporary concrete formulations that incorporate supplementary cementitious materials, chemical admixtures, and recycled aggregates [3], which exhibit complex non-linear relationships between constituent materials and resulting mechanical properties.

The increasing complexity of modern concrete technology necessitates advanced computational approaches capable of modeling intricate interactions between multiple mix design parameters [4]. Machine learning algorithms have emerged as powerful alternatives to conventional methods, offering superior capability to capture non-linear relationships and complex dependencies inherent in concrete strength development [5]. Gradient boosting algorithms, particularly CatBoost, have demonstrated exceptional performance in regression tasks involving heterogeneous datasets with mixed feature types [6], making them highly suitable for concrete strength prediction applications [7,8].

However, the effectiveness of machine learning models critically depends on optimal hyperparameter configuration, which significantly influences predictive accuracy and generalization capability [9]. Traditional hyperparameter optimization methods, such as grid search and random search, often prove computationally expensive and may fail to identify global optima in complex search spaces [10]. This limitation has motivated the development and application of metaheuristic optimization algorithms that can efficiently explore large parameter spaces while balancing exploration and exploitation mechanisms [11].

In hyperparameter tuning, the application of bio-inspired optimization algorithms has been of particular interest as it can offer the advantage of multi-dimensional, non-convex optimization problems that are useful to address high-dimensional hyperparameter tuning problems [12]. It follows that these algorithms are inspired by nature and biological processes and develop superior search strategies capable of identifying the best solutions in challenging landscapes [13]. The Somersaulting Spider Optimizer (SSO) is a new bio-inspired metaheuristic algorithm that is inspired by the locomotion behavior of a unique type of spider known as Cebrennus rechenbergi that is capable of combining high-energy exploration behavior with low-energy exploitation behavior to produce balanced search behavior for complex optimization problems.

The prediction of concrete strength can be precise when using metaheuristic optimization with machine learning, which has a significant potential to enhance the power of concrete strength prediction tolerance, and thus, it can lead to the elimination of prediction errors in concrete strength prediction [14,15]. CatBoost is particularly efficient in work with concrete datasets due to the flexibility of the software with numerical data, the quality of automatic preprocessing, and anti-overfitting features of the used algorithm, but SSO also includes an effective search mechanism and adaptive energy management, which guarantee the optimal optimization process of the hyperparameters as possible [16].

This study presents a comprehensive framework for concrete compressive strength prediction that combines the Somersaulting Spider Optimizer with CatBoost algorithm to achieve superior predictive performance. The research makes several significant contributions to the field of computational concrete engineering:

Primary Contributions 1.

Novel Bio-Inspired Optimization Algorithm: Introduction and detailed mathematical formulation of the Somersaulting Spider Optimizer (SSO), a novel metaheuristic algorithm inspired by the unique locomotion behavior of Cebrennus rechenbergi. The algorithm incorporates adaptive energy management and dual-phase movement mechanisms (somersaulting for exploration and rolling for exploitation) specifically designed for hyperparameter optimization in machine learning applications.

The remainder of this paper is organized as follows: Section 2 presents a comprehensive literature review of machine learning approaches in concrete strength prediction, Section 3 details the materials and methods including dataset description, SSO algorithm formulation, and CatBoost implementation, Section 4 presents experimental results and comparative analysis, Section 5 provides discussion of findings and implications, and Section 6 concludes with key contributions and future research directions.

The ensemble learning techniques have also demonstrated specific potential in challenging the complexity of prediction of concrete strength. A recent study by way of illustration of this trend [17] where pure physical testing is replaced by the ISSA-BPNN-AdaBoost model, which has shown itself to be superior to the physical testing model when compared to other traditional physical testing models. The technique does not suffer time consumption and cost constraints that characterizes conventional strength testing yet high degree of accuracy is achieved. The effectiveness of optimizing the base learners in ensemble structures through excellent results is further proved by the superiority between ensemble learners in different datasets.

There has been a growing role of model interpretability because machine learning models are now being incorporated into significant engineering choices. Especially, one of the recent studies established a framework of interpretable machine learning applied to high-performance concrete (HPC) compressive strength prediction [18] indicates the use of random forest, AdaBoost, XGBoost, and LightGBM models, hyperparameter optimization (by cross-validation methods). SHAP analysis indicated that age, water-to-cement ratio, slag and content of water, compose crucial elements that affect forecasted compressive strength and superplasticizers ratios had lower impacts. This decision supportable method improves precision and visibility of prediction such that engineers gather the associations between mix elements and connections of strength properties.

Extensive dataset generation has brought many achievements in the field and researchers identify the significance of an effective data base to develop and validate the model. As an example, ConcreteXAI dataset [19] is an important contribution as it is a ten year lab study consisting of 18,480 measurements each of 12 different concrete formulations. This is holistic data consisting of both mechanical and non-destructive tests, which makes a proper development of a predictive model quite feasible and decision-making processes to stay transparent.

The use of concrete has been studied beyond the regular concrete to consider more specific formulations and uses according to the various needs of the construction projects judged by the modern needs. As an illustration, one study [20] examining preplaced aggregate concrete (PAC) discovered that the complexity of loading cementitious mixtures into existing form of pre-placed coarse aggregates led to complicated methods of prediction because of production peculiarities and the material properties the final products developed. Likewise, one study [21] also examined hybrid-fiber-reinforced recycle aggregate concrete (HFRRAC) where eleven machine learning models were used to overcome non-linearity between constituents. The results revealed that light gradient boosting machine algorithms were the most precise with regards to this complicated material system and they obtained high R2 values and low error measures.

In modern engineering practices, high-performance concrete has been given special focus because of enhanced mechanical characteristics and multi-faceted engineering practise in most modern construction works. According to a recent study [22], the compressive strengths that can be reached using HPC are higher than 200 MPa, the average strengths of 100–120 Mpa can be used in 28 days in a variety of engineering projects. The problems related to predictability of such high-strength material assume use of advanced modeling methods, which are capable of modeling complex correlations between advanced admixtures and final properties.

The combination of optimization algorithms with machine learning has also become a potent method of predicting concrete strength that combines the pattern matching abilities of machine learning with the methodical search capacity of optimization algorithms. Indeed an example of this was noted by research study carried out by [23] using a hybrid combination of a random forest with a particle swarm optimization (PSO) to predict concrete strength using blast furnace slag and fly ash. The model had a correlation coefficient of 0.954 on the testing set, which proved the great generalization capacity to sustainable forms of concrete.

Progress in hybrid methods are occurring, and scientists are creating more powers of prediction of concrete strength through computer models. Among others, more recent research has proposed an advanced hybrid algorithm incorporating deep learning with reinforcement learning that utilizes Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) based feature extraction and a Dueling Double Deep Q-Network agent that optimises mix ratios by repeatedly repeating the optimization of features and mix ratios [24]. The methods are a future of computational intelligence in a real world design.

Concrete strength prediction research has come to include an increased focus on environmental sustainability (sustainable construction practices) as the prevailing attitude of the construction industry. One case in point consists of the recent introduction of the SmartMix Web3 model, which is an overview of ensemble machine learning integrated with blockchain technology to lend more strength to low-carbon concrete design [25]. This is a novel way of approaching things to show that, even to the technical challenges that are on the table, the environment is being challenged by modern computational methods.

It may also be an important line of research to resolve the data limitations because not all types of concrete and all its applications are covered with extensive datasets. Here, the framework offered by recent research [26], combining Model-Agnostic Meta-Learning (MAML) and Shapley Additive Explanations (SHAP) aids the engineering of solutions to get a better predictive accuracy, as well as to improve an interpretation in data-limited scenarios, where both methodologies are essential ways to achieve higher predictive performance using fewer explanations grounded on prediction equations.

Table 1 provides a systematic comparison of the various machine learning approaches discussed, highlighting the diversity of methodologies and their specific applications in concrete strength prediction.

The comprehensive analysis presented in Table 1 reveals several critical research gaps that motivate the present study. First, despite the proven effectiveness of gradient boosting algorithms, CatBoost remains underexplored in concrete strength prediction compared to XGBoost and LightGBM, despite offering superior handling of categorical features, built-in regularization, and resistance to overfitting. Second, systematic comparison of CatBoost with diverse bio-inspired metaheuristic optimizers is absent from the literature, with most studies limiting investigations to single optimization algorithms or relying on computationally expensive grid search methods. Third, recent bio-inspired algorithms such as the Somersaulting Spider Optimizer (SSO), which demonstrates promising exploration-exploitation balance through adaptive energy management, have not been applied to concrete strength prediction. These gaps collectively justify the need for a comprehensive framework combining CatBoost’s gradient boosting capabilities with SSO and other metaheuristic algorithms to achieve enhanced predictive accuracy, computational efficiency, and robust generalization across diverse concrete formulations.

The Somersaulting Spider Optimizer (SSO) represents a recent advancement in bio-inspired metaheuristics, employing dual locomotion modes—high-energy somersaulting for global exploration and low-energy rolling for local exploitation—regulated by adaptive energy management that balances search diversification and intensification. SSO has demonstrated competitive performance across benchmark optimization problems and has proven effective in ensemble learning applications, particularly for earthquake prediction where it achieved superior predictive accuracy and generalization compared to conventional metaheuristics [27]. Given SSO’s proven capability in handling nonconvex, multimodal optimization landscapes and its successful application to machine learning model optimization, it presents a promising candidate for CatBoost hyperparameter tuning in concrete strength prediction, addressing the identified research gap regarding the integration of advanced bio-inspired algorithms with gradient boosting frameworks for enhanced predictive performance.

This study proposes an elaborate machine learning model to forecast concrete compressive strength that is based on the advanced gradient boosting model and bio-inspired optimization tool. The three primary components of the methodology are the data preparation and analysis done with the help of exploratory statistics, a Somersaulting Spider Optimizer (SSO) algorithm to find hyperparameters, and the use of CatBoost as the predictive model. Experimental framework addresses the system of non-linear relationships among parameters of concrete mix engineering and the compressive power of resultant products through optimization of machine learning model hyperparameters. The suggested plan applies the bio-inspired exploration-exploitation principles of the SSO algorithm to identify the most appropriate CatBoost parameters and optimize the predictive performance basing on different concrete formulas. The integrated method provides a potent foundation on realistic estimation of concrete strength, to support the decision making which is based on data in optimal mixes designs and quality control uses.

The proposed methodology follows a systematic three-phase workflow, as illustrated in Fig. 1. The data preparation phase begins with the concrete dataset containing eight input features, followed by preprocessing and exploratory data analysis to identify patterns and correlations. In the model development phase, the dataset is partitioned into training (80%) and testing (20%) sets, and the CatBoost regression model is optimized using the Somersaulting Spider Optimizer (SSO) for hyperparameter tuning. The evaluation phase assesses predictive performance through multiple metrics including MSE, R2, and MAE, validates statistical significance using ANOVA and Wilcoxon tests, and performs comparative analysis against eight alternative optimization algorithms to demonstrate the superiority of the SSO-based approach.

This section presents a comprehensive evaluation of machine learning models and optimization algorithms for concrete strength prediction through three analytical phases. First, baseline machine learning models are assessed using standard performance metrics to establish fundamental predictive capabilities. Second, metaheuristic optimization algorithms enhance the best-performing baseline model, demonstrating bio-inspired optimization effectiveness. Finally, statistical analysis validates performance differences through parametric and non-parametric testing methods. Standardized hyperparameter configurations were implemented for all optimization algorithms to ensure fair comparison, as presented in Table 3.

This study demonstrates the effectiveness of integrating the Somersaulting Spider Optimizer (SSO) with the CatBoost machine learning algorithm for predicting concrete compressive strength. The proposed hybrid framework effectively balances exploration and exploitation within the hyperparameter space, enabling the identification of optimal parameter combinations that substantially enhance predictive performance. The adaptive energy management mechanism embedded in the SSO algorithm facilitates dynamic search behavior, effectively avoiding premature convergence to local minima—a common limitation of traditional optimization approaches. Combined with CatBoost’s capabilities for handling categorical and numerical features through ordered boosting and symmetric decision tree structures, the model successfully captures complex nonlinear relationships among concrete mix proportions, curing age, and compressive strength. As shown in Table 5, the SSO-CatBoost model achieved the lowest mean squared error of 12.61 and the highest coefficient of determination (R²) of 0.951 among all optimized models, representing a 24.6% improvement over the baseline CatBoost model presented in Table 4. This performance improvement is further illustrated in Fig. 17, which visualizes the comparative advantage through color intensity mapping where SSO-CatBoost exhibits the lightest coloration for error metrics. The enhanced prediction accuracy enables more refined concrete mixture designs that minimize material consumption while maximizing structural reliability, potentially leading to significant cost reductions in large-scale concrete production.

Feature importance analysis, as presented in Fig. 7, indicates that cement content and concrete age exert the strongest influence on compressive strength prediction, consistent with established knowledge regarding cement hydration kinetics and strength development mechanisms. Secondary yet significant contributions from water content, supplementary cementitious materials, and chemical admixtures underscore the multifaceted nature of mix design effects on concrete performance, as evidenced by the correlation patterns shown in Fig. 2. The statistical significance of the SSO-CatBoost model’s superiority is confirmed through ANOVA results presented in Table 6, which yielded an F-statistic of 7.17 with p-value < 0.0001, and the Wilcoxon signed-rank test outcomes in Table 7, which established that SSO-CatBoost achieved the highest actual median (0.945) with statistical significance (p < 0.001). This trend is further supported by the R² distribution boxplots in Fig. 21, which demonstrates SSO-CatBoost’s highest median R² score (0.95) with the smallest interquartile range, and the radar comparison in Fig. 18, which collectively demonstrate consistent outperformance across multiple evaluation metrics. The cumulative distribution functions presented in Fig. 15 further confirm the concentrated low-error predictions achieved by the optimization framework. Although the dataset encompasses diverse mix proportions without exhaustive material property characterization, the methodology remains highly applicable to industry-standard mix design processes. Overall, this study provides a transparent and reproducible framework while generating actionable insights for concrete mixture optimization and the development of sustainable, resilient infrastructure systems.

According to the No Free Lunch theorem, the strong performance of the SSO-CatBoost framework is specific to the concrete compressive strength dataset and cannot be universally. When considering new datasets, alternative material systems, or different prediction objectives, it is necessary to re-evaluate and benchmark the framework to ensure effectiveness in each new context.

This study demonstrated a novel concrete compressive strength prediction strategy, proposed by producing the fusion of the Somersaulting Spider Optimizer (SSO) and the CatBoost model to address the relevant and considerable issue of hyperparameter optimization in machine learning-guided solutions of concrete engineering. The experimentation could demonstrate the effectiveness of bio-inspired optimization techniques for enhancing the accuracy of prediction in compound material characteristics. The developed SSO algorithm, built on the characteristic locomotion behaviour, provided a new balancing of exploratory and exploitative strategies, employing adaptive energy management and dual-phase movement strategies. The somersaulting and rolling motions of the algorithm, along with the stagnation-based intensity adjustment, are more efficient in comparison to conventional optimization procedures. In machine learning systems, the hyperparameter optimization, which was offered by mathematical models of time-dependent adaptive factors and time-dependent energy-based behavioural switching models, had a solid background.

Since there are seven known methods of metaheuristic optimization applied (Differential Evolution, Grey Wolf Optimizer, Whale Optimization Algorithm, Genetic Algorithm, Bat Algorithm, Harris Hawks Optimization and Particle Swarm Optimization), comparative analysis demonstrated that the SSO was effective across a wide scope of performance indicators. The convergence speed of the SSO was never outperformed by the competing algorithms in the quality of solutions and robustness, demonstrating that the algorithm has a broader range of applications in engineering optimization problems.

The Analysis of Variance (ANOVA) and the Wilcoxon signed-rank tests revealed the statistical value of the alterations in performance brought about by the implementation of the proposed methodology. The pertinent and stringent statistical analysis involved in the assessment of confidence vests precision in the algorithm, not only in confidence but also in workability in real-life work in the engineering field.

It is possible to identify some of the future research directions. This paper presents a variety of possible future avenues for the current research. Continued development of the SSO algorithm would enable the optimization of multiple concrete properties concurrently throughout such measures as strength, durability, workability and sustainability. It would enhance the generalisability of the algorithm to research on the algorithm applied on larger datasets comprising various geographical settings, material provision and construction techniques.

A possible industrial route will be the real-time use of the SSO-CatBoost system in concrete production facilities to test the product quality and adjust the mix compositions in real time. In addition, an improvement of the methodology in forecasting other tangible features, such as the lifespan of concrete, permeability, and thermal properties, would be helpful to widen its scope of use. A limitation is that the study does not perform sensitivity analysis on all hyperparameters.