Ultra-high-performance concrete, or UHPC, has emerged as a critical material for modern engineering due to its exceptional strength, toughness, and durability under various loading conditions. While its static properties are well-documented, the material’s behavior under high-strain-rate scenarios—such as those encountered in vehicle collisions, blast events, or high-velocity fragment impacts—remains complex to predict. The dynamic compressive strength and the dynamic increase factor are vital parameters for assessing the load-carrying capacity and safety of protective structures, yet obtaining these values experimentally is both costly and resource-intensive. Current experimental protocols rely on specialized facilities like split Hopkinson pressure bars, which often result in heterogeneous datasets that are difficult to consolidate. To address these challenges, researchers have developed a physics-guided two-stage XGBoost framework designed to predict the dynamic compressive strength of UHPC by effectively capturing the dominant physical trends and material-system-dependent deviations from diverse, cross-literature data sources.
The study behind this framework involved the construction of a comprehensive cross-literature database comprising 227 experimental records gathered from 32 independent research studies. By systematically organizing data on mix proportions, fiber characteristics, and strength indices, the researchers established a unified variable framework that allows for more robust statistical analysis. A crucial aspect of this database is the retention of unreinforced reference matrices from systematic material series, ensuring that fiber-reinforced mixtures can be evaluated against their corresponding baseline responses. To further improve the model’s coverage of sparse regions in the experimental database, the team employed numerical simulations conducted with the LS-DYNA software using the Karagozian and Case concrete model. These simulated records were only retained if they met strict relative deviation criteria compared to experimental values, with the final training set consisting of 181 experimental records and 619 validated numerical records, while all experimental test observations were kept separate to ensure rigorous evaluation.
The central innovation of the proposed model is its two-stage trend-residual decomposition strategy, which deviates from conventional black-box data-driven approaches that often struggle with generalization across heterogeneous sources. In the first stage, the framework captures the dominant global trend governing dynamic compressive strength by using logarithmic strain rate and static compressive strength as inputs. By focusing on these two physically meaningful variables, the Stage I model establishes a baseline rate-dependent response. To ensure physical plausibility, the researchers imposed a monotonic non-decreasing constraint on the logarithmic strain-rate feature, which prevents the model from generating irregular, physically inconsistent predictions in data-sparse regions of the feature space. This structural prior helps align the model with established engineering expectations for how concrete materials typically behave under increasing strain rates.
In the second stage of the framework, the model learns the residual corrections associated with specific material systems, using variables such as cement, supplementary cementitious materials, filler, fine aggregate, water, superplasticizer, and steel-fiber volume fraction. Unlike the Stage I trend model, Stage II specifically targets the deviations from the global rate-dependent behavior induced by the unique composition and fiber characteristics of different UHPC mixtures. To improve robustness in the presence of irregular error patterns common in cross-literature datasets, the team utilized a Pseudo-Huber loss function for Stage II. This objective function maintains efficiency similar to standard mean squared error for small deviations while reducing the penalty for large outliers, effectively mitigating the influence of experimental noise and systematic biases present in the compiled literature.
The training process employed a grouped splitting strategy where each independent source study was treated as a single grouping unit, ensuring that samples originating from the same experiment were never split between the training and test partitions. This paper-wise evaluation methodology provides a more realistic assessment of the model’s ability to generalize to previously unseen sources, preventing optimistic performance estimates caused by data leakage. Following the training phase, the model’s accuracy was benchmarked against a wide variety of machine learning methods, including decision trees, random forests, and other boosting-based algorithms. Under the fixed paper-wise split, the physics-guided two-stage framework demonstrated superior performance, achieving a coefficient of determination of 0.9816, a root mean squared error of 6.593 MPa, a mean absolute error of 4.880 MPa, and a mean absolute percentage error of 2.71 percent.
To assess the model’s sensitivity to parameter variations and different held-out study combinations, the researchers performed repeated paper-wise evaluations, confirming that the framework remains stable and competitive across various data splits. The ablation analysis highlighted the specific contribution of each component within the two-stage model: the trend-residual decomposition, the robust Pseudo-Huber loss, and the monotonic strain-rate constraint all played identifiable roles in improving statistical performance and physical consistency. The results showed that while the monotonic constraint might slightly reduce pure statistical fit, its inclusion is necessary to ensure the model adheres to the fundamental physics of the material’s rate-dependent response, representing a favorable trade-off for engineering applications.
For transparency and interpretability, the researchers applied the Shapley Additive Explanations method to reveal how each input variable contributes to the model’s final prediction. The analysis confirmed that static compressive strength and logarithmic strain rate serve as the primary drivers of the dynamic strength prediction, consistent with their role in the Stage I trend capture. Among the Stage II mixture variables, content related to supplementary cementitious materials, steel-fiber volume fraction, and the fiber aspect ratio emerged as the most significant contributors to the material-specific residual corrections. These SHAP values provided a global view of feature importance while also offering local explanations for individual samples, highlighting how different mixture components modulate the dynamic response of UHPC beyond the initial trend.
It is important to note that the interpretability analysis reflects model attribution within the specific context of the current database rather than demonstrating direct, isolated causal effects. For example, the SHAP analysis indicated that higher steel-fiber volume fractions are predominantly associated with negative SHAP contributions in this dataset. The authors cautioned that this pattern likely reflects confounding variables, such as coupling between fiber content, matrix composition, and source-specific experimental design choices, rather than an inherent, independent decrease in material strength. This nuance is crucial for researchers using such tools to avoid misinterpreting model attribution as definitive causal inference, especially when working with heterogeneous data consolidated from numerous independent studies.
To facilitate broader adoption of this methodology in engineering practice, the research team developed a prototype graphical user interface that allows users to input key design parameters—such as strain rate, static compressive strength, and fiber descriptors—to obtain dynamic strength predictions without requiring specialized machine learning knowledge. The GUI incorporates an automatic validity check based on the training data range, which alerts the user when inputs are likely to produce extrapolative results with reduced reliability. This tool serves as a practical bridge between sophisticated data-driven modeling and day-to-day structural analysis, enabling engineers to perform rapid parametric explorations and design comparisons efficiently.
The study concludes that the two-stage framework effectively addresses the primary challenge of cross-literature heterogeneity in UHPC dynamic compressive strength prediction. While the model shows robust and competitive performance, the authors identified several limitations for future work. The current database, for example, is concentrated around a limited range of standard fiber geometries and dosages, with insufficient information on factors like fiber orientation, dispersion, and hybridization types. Furthermore, because Stage II does not explicitly model material-rate interactions, the framework may not fully capture strength-dependent variations in rate sensitivity. Future efforts will be directed toward expanding the experimental database and refining the decomposition architecture to further improve generalization and the interpretation of material-level effects.
Subject of Research: Development of a physics-guided two-stage XGBoost framework for predicting the dynamic compressive strength of ultra-high-performance concrete using heterogeneous experimental data.
Article Title: A physics-guided two-stage XGBoost framework for predicting dynamic compressive strength of UHPC from heterogeneous data
Article References: Shi, T., Wu, C., & Zhang, Y. (2026). A physics-guided two-stage XGBoost framework for predicting dynamic compressive strength of UHPC from heterogeneous data. Case Studies in Construction Materials, 25, Article e06716. https://doi.org/10.1016/j.cscm.2026.e06716
Image Credits: AI Generated
DOI: 10.1016/j.cscm.2026.e06716
Keywords: UHPC, ultra-high-performance concrete, machine learning, XGBoost, dynamic compressive strength, high strain rate, structural engineering, physics-guided learning, SHAP, model interpretability, data-driven modeling, cross-literature heterogeneity, concrete materials, numerical simulation, LS-DYNA, predictive modeling
News Source: Teresa Odom. (November 18, 2026). Physics-Guided Machine Learning Model Predicts Ultra-High-Performance Concrete Dynamic Strength. Scienmag.



