A comprehensive review published in the journal AI in Civil Engineering examines the critical balance between the predictive power of artificial intelligence and the interpretability required for safe engineering decisions. The study, authored by Monjurul Hasan and Ming Lu, specifically focuses on construction engineering and management, where complex input-output relationships demand both accurate predictions and clear explanations. The authors argue that explainable AI is not merely a technical add-on but a fundamental requirement for safety and compliance-sensitive applications, ensuring transparency in how AI systems reach their conclusions.
The paper identifies a persistent challenge in the field: while modern machine learning algorithms excel at processing high-dimensional data and emulating human intuition, they often operate as opaque black boxes. This opacity creates a significant barrier to trust, as engineering professionals hesitate to rely on predictions without understanding the underlying decision rules or mathematical frameworks. The authors note that current AI models frequently lack clear causal mechanisms, leading to inconsistencies and errors that can be difficult to diagnose or correct in professional practice.
To address this gap, the review distinguishes between two key concepts often conflated in the literature: explainability and interpretability. Explainability refers to the generation of audience-specific explanations that address particular needs, such as validating a model’s logic for a domain expert. Interpretability, on the other hand, measures how well these explanations align with the user’s existing domain knowledge and make contextual sense. This distinction is crucial for ensuring that AI outputs are not only technically accurate but also logically resonant with the experience and expertise of civil engineers.
The authors propose a tailored assessment framework for evaluating machine learning algorithms in civil engineering, focusing on problem context, data constraints, and model explainability. This framework categorizes XAI methods into post-hoc and ante-hoc approaches. Post-hoc methods, such as LIME and SHAP, provide explanations after a decision has been made, often approximating the model locally. Ante-hoc methods, like decision trees and linear regression, are built with inherent transparency, allowing users to understand the model’s logic from the outset. The review highlights that while post-hoc methods are useful for opaque models, ante-hoc approaches offer greater reliability for critical engineering decisions.
A central contribution of the paper is the evaluation of the Enhanced Model Tree (EMT), a novel method that integrates binary data classification with variance analysis. Unlike conventional machine learning models that typically provide only point-value predictions, the EMT offers analytical prediction of variance estimates alongside point values. This dual capability is particularly valuable in safety-sensitive applications, such as concrete strength prediction and structural health monitoring, where understanding the uncertainty of a prediction is as important as the prediction itself. The EMT architecture uses a divide-and-conquer approach, recursively splitting data into subsets and fitting linear regression models to each, thereby maintaining a clear and traceable logic from input to output.
The study compares the EMT against several commonly applied machine learning algorithms, including linear regression, k-nearest neighbors, Bayesian models, decision trees, support vector machines, and artificial neural networks. While artificial neural networks often deliver high predictive performance, they are difficult to interpret without specialized tools. In contrast, the EMT maintains explainability through its clear prediction logic and provides a self-documenting model architecture. This transparency facilitates communication between technical teams, project stakeholders, and regulatory bodies, which is essential for establishing trust in AI-assisted solutions for cost control, structural integrity, and environmental impact assessments.
The review also explores the broader landscape of XAI applications in civil engineering, citing recent studies that have employed techniques like SHAP and LIME to interpret models for road transport impacts on groundwater, seismic retrofitting performance, and building energy forecasting. These examples demonstrate the growing adoption of interpretability methods across diverse sub-fields, from geotechnical assessments to sustainable design. However, the authors point out that the field still grapples with balancing interpretability and performance, with many current solutions relying on post-hoc explanations that may not fully capture the model’s internal logic.
One of the key limitations identified in the review is the susceptibility of AI models to noise and inconsistencies in real-world construction data. The authors emphasize the importance of data preprocessing and feature selection to ensure that models are not overfitting to specific datasets or capturing spurious correlations. They recommend using techniques such as stepwise regression and principal component analysis to identify genuinely predictive features and distinguish signals from noise. This rigorous approach to data handling is essential for ensuring that AI systems remain reliable and generalizable across different engineering contexts.
The paper concludes by drawing an analogy to the Yin-Yang philosophy, suggesting that AI and XAI must be balanced to provide effective solutions. AI advances in efficiency and optimization, while XAI provides the logical reasoning behind conclusions. This synergy fosters trust in AI systems, enabling safer and more socially responsible outcomes. The authors argue that by formalizing this balance, the research can help bridge the gap between complex AI models and end-user accountability, ensuring that the full potential of AI is realized in the field of civil engineering.
Looking ahead, the study suggests that future research should focus on developing ante-hoc solutions that maintain both transparency and predictive power. The EMT framework offers a promising direction, as it inherently integrates accuracy and transparency without the need for additional post-hoc explanation tools. By providing a robust analytical framework that satisfies both technical requirements and organizational needs for verifiable decision support, the EMT could become a standard tool for construction engineering predictive analytics, facilitating more informed decision-making and efficient resource allocation.
The review situates the growing demand for explainable AI within a broader historical context, noting that academic and industry interest has surged in recent years. According to Google Trends data cited in the paper, searches for “explainable artificial intelligence” increased by 80% since 2023 compared to the previous decade. This trend parallels the wider adoption of artificial intelligence and reflects an increasing recognition that interpretability is a prerequisite for responsible deployment across technical domains. The authors trace the formalization of XAI back to 2004, when Van Lent et al. introduced the concept in the context of game simulations, designing systems to articulate their internal decision-making logic. Later, Gunning et al. refined the definition within computing research, characterizing XAI as the development of secondary models capable of explaining primary machine learning model behavior while maintaining sufficient predictive fidelity. This historical progression underscores that the push for transparency is not merely a recent reaction to black-box models but a long-standing effort to reconcile accuracy with understanding.
Within the XAI framework, the paper distinguishes between post-hoc and ante-hoc explainability, each offering different trade-offs. Post-hoc methods provide local explanations after a decision is made, focusing on specific outputs rather than the entire system. Key techniques include LIME, which approximates the model locally with interpretable representations such as linear models or decision trees, and BETA, which balances fidelity and interpretability when explaining black-box classifiers. In contrast, ante-hoc explainability is built into the model architecture, ensuring inherent transparency. Examples include traditional interpretable models like linear regression and decision trees, as well as Generalized Additive Models, which maintain interpretability while detecting patterns in high-dimensional data. The review notes that post-hoc methods excel at explaining opaque models, whereas ante-hoc approaches prioritize transparency by design, presenting a fundamental trade-off between flexibility and interpretability that engineers must navigate based on application requirements.
The paper also outlines three critical requirements for reliable explainability in AI systems. First, explainable data requires that the data used to train the model be clearly identified, including the rationale for selection, methods for assessing fairness, and steps taken to mitigate bias. Second, explainable predictions demand that the features activated by the model to produce specific outputs be transparent, with the reliability of the output comprehensible to users. Third, an explainable model requires that individual components or layers be described, detailing their roles and contributions to the final output. These requirements ensure that AI systems do not merely produce accurate results but also provide the logical underpinning necessary for professional validation. The authors emphasize that explainability is inherently subjective, as its effectiveness depends on end-users’ ability to evaluate the plausibility and trustworthiness of the explanations provided.
Interpretability, distinct from explainability, refers to an AI model’s built-in capacity to produce human-understandable outputs based on established knowledge. Effective interpretability requires explanations that match users’ cognitive frameworks, expertise, and potential biases. The paper identifies three key dimensions of transparency that support interpretability: simulatability, decomposability, and algorithmic transparency. Simulatability ensures that a model is simple enough for users to understand and accurately replicates the behavior it is designed to represent. Decomposability involves breaking down a model into its fundamental components—inputs, parameters, and computations—for individual analysis. Algorithmic transparency focuses on understanding the procedures a model employs to generate its output. These dimensions are essential for ensuring that predictive models remain interpretable despite their complexity, allowing stakeholders to reliably predict system behavior across various scenarios.
The review further examines the perspectives from which XAI necessity is assessed, including regulatory, scientific, industrial, developmental, and end-user viewpoints. From a regulatory perspective, domains where AI decisions carry substantial legal and ethical consequences, such as credit approvals under the EU’s GDPR, require explanations to ensure accountability. The scientific perspective highlights that while black-box models can solve complex problems, XAI techniques help unpack this knowledge, potentially leading to novel scientific insights. The industrial perspective notes that business adoption of black-box AI faces challenges from regulatory constraints and user distrust, often leading industries to prioritize interpretable models for compliance, sometimes at the cost of reduced accuracy. The developmental perspective emphasizes that XAI provides tools for diagnosing failure modes such as data biases and overfitting, improving system reliability. Finally, the end-user perspective addresses concerns about reliability and fairness, as black-box systems can produce counterintuitive results that raise questions about the alignment of AI decisions with intended goals.
Subject of Research: Explainable AI frameworks for predictive modeling in civil engineering
Article Title: Bridging AI and explainability in civil engineering: the Yin-Yang of predictive power and interpretability
Article References: Hasan, M., & Lu, M. (2025). Bridging AI and explainability in civil engineering: the Yin-Yang of predictive power and interpretability. AI in Civil Engineering, 4(1), Article 21. https://doi.org/10.1007/s43503-025-00066-6
Image Credits: AI Generated
DOI: 10.1007/s43503-025-00066-6
Keywords: Explainable AI, Civil Engineering, Machine Learning, Construction Management, Predictive Modeling, Enhanced Model Tree, Interpretability, Data Science, Bridging, explainability, civil, engineering
News Source: Denise Maddox. (November 10, 2026). Review Links AI Predictive Power and Explainability in Civil Engineering. Scienmag.



