With the continuous development of infrastructure in China, the demand for concrete has increased significantly. As a critical raw material in concrete production, sand has also seen a sustained rise in consumption. It is estimated that global annual sand usage ranges between 40 and 50 billion tons. Due to the non-renewable and non-recyclable nature of river sand, its excessive extraction has led to the rapid depletion of existing reserves. To enhance resource efficiency, the construction industry has gradually implemented a variety of mitigation strategies, such as incorporating recycled coarse aggregates into concrete and employing manufactured sand as a substitute for river sand. The use of manufactured sand not only alleviates the environmental degradation caused by overexploitation of river sand but also offers a practical solution to the growing scarcity of this natural resource. Manufactured sand is produced by mechanically crushing and screening rock particles. Compared with river sand, it is characterized by a rougher surface texture and a higher content of stone powder, which can significantly influence the long-term mechanical performance of concrete. Creep, defined as the time-dependent deformation of concrete under sustained loading, plays a crucial role in determining the structural stability and durability. However, existing creep models for concrete have proven inadequate in accurately predicting the creep behavior of manufactured sand concrete. In recent years, machine learning techniques have been increasingly applied to the prediction of creep behavior in recycled aggregate concrete, ultra-high-performance concrete, and concrete incorporating mineral admixtures, demonstrating excellent fitting accuracy and generalization capabilities. Compared with traditional predictive models, machine learning approaches are better suited to handling complex issues such as multi-factor coupling and nonlinear relationships. However, their limitations are also notable: these models often function as "black boxes," lacking physical interpretability, and are highly dependent on the quality and quantity of input data, making it difficult to uncover the underlying physical mechanisms of creep. To address these challenges, some researchers have introduced explainable algorithms to improve model transparency and robustness Simultaneously, other studies have explored explicit modeling techniques to directly generate predictive formulas. Nonetheless, the resulting expressions are often overly complex and lengthy, limiting their applicability in practical engineering scenarios.

To tackle these challenges, a novel modeling framework that integrates both implicit and explicit techniques is proposed. First, a dataset of manufactured sand concrete creep was constructed. Advanced implicit machine learning models, including the Feature Token Transformer, Physics-Informed Neural Networks (PINNs), TabNet, and XGBoost, were employed to predict the creep behavior of manufactured sand concrete. Feature importance analysis was subsequently conducted based on model outputs to identify the most influential variables. Next, AutoFeat was used to automatically generate and select polynomial features from the identified key variables, thereby constructing a high-order explicit model capable of achieving accurate creep prediction. Furthermore, neural symbolic regression (NSR) was introduced to automatically discover nonlinear relationships among variables through symbolic function composition, yielding concise, physically meaningful analytical expressions.

Existing concrete creep prediction models, when applied to manufactured sand concrete, exhibit substantial errors due to their inability to effectively capture the material-specific characteristics and time-dependent behavior. This highlights an urgent need for the development of more targeted and robust predictive models.

The constructed implicit machine learning models—TabNet, FT-Transformer, PINNs, and XGBoost—demonstrated excellent predictive performance. These models effectively capture the complex and nonlinear relationships between the creep behavior of manufactured sand concrete and various influencing factors.

Explicit prediction models for concrete creep compliance were developed based on two approaches: AutoFeat and Neural Symbolic Regression (NSR). Both methods successfully generated functional expressions that map input parameters to the target output, thereby enhancing model interpretability. The AutoFeat-based model combined SHAP-based feature importance analysis with Pearson correlation analysis during feature selection, ultimately yielding an interpretable model primarily composed of first-order linear terms, complemented by selected second- and third-order interaction terms. In contrast, the NSR method exhibited a superior capacity for capturing nonlinear dependencies. It automatically identified key variables and constructed complex mathematical formulations characterized by nested structures and nonlinear operator combinations, providing deeper insight into the underlying mechanisms governing creep development.

Under comparable levels of predictive accuracy, both AutoFeat and Neural Symbolic Regression are capable of generating explicit expressions with clear mathematical structures, significantly enhancing model interpretability. Despite differences in feature selection strategies and model architectures between implicit and explicit modeling approaches, both demonstrate strong fitting performance with respect to real-world data, thereby validating the feasibility of modeling concrete creep through multiple methodological pathways. The integrated application of both approaches facilitates an optimized balance between predictive accuracy and interpretability.