Additive-feature-attribution methods: a review on explainable artificial intelligence for fluid dynamics and heat transfer (2409.11992v1)

Abstract: The use of data-driven methods in fluid mechanics has surged dramatically in recent years due to their capacity to adapt to the complex and multi-scale nature of turbulent flows, as well as to detect patterns in large-scale simulations or experimental tests. In order to interpret the relationships generated in the models during the training process, numerical attributions need to be assigned to the input features. One important example are the additive-feature-attribution methods. These explainability methods link the input features with the model prediction, providing an interpretation based on a linear formulation of the models. The SHapley Additive exPlanations (SHAP values) are formulated as the only possible interpretation that offers a unique solution for understanding the model. In this manuscript, the additive-feature-attribution methods are presented, showing four common implementations in the literature: kernel SHAP, tree SHAP, gradient SHAP, and deep SHAP. Then, the main applications of the additive-feature-attribution methods are introduced, dividing them into three main groups: turbulence modeling, fluid-mechanics fundamentals, and applied problems in fluid dynamics and heat transfer. This review shows thatexplainability techniques, and in particular additive-feature-attribution methods, are crucial for implementing interpretable and physics-compliant deep-learning models in the fluid-mechanics field.

• The paper presents additive-feature-attribution methods that utilize SHAP values to explain complex machine-learning models in turbulent flow simulations.
• It details four implementations—in kernel, tree, gradient, and deep SHAP—highlighting their role in enhancing RANS models for fluid dynamics.
• The review emphasizes practical applications, from turbulence modeling to industrial heat-transfer optimization, paving the way for more interpretable AI-driven designs.

Overview of Additive-Feature-Attribution Methods in Explainable AI for Fluid Dynamics and Heat Transfer

The paper by Cremades et al. provides a comprehensive review of additive-feature-attribution methods, a subset of explainable artificial intelligence (XAI) techniques, and their applicability in the complex fields of fluid dynamics and heat transfer. The use of data-driven methods remains a pivotal advancement in fluid mechanics, made evident by the adoption of machine-learning techniques for predicting the evolution of turbulent flows where solutions to governing equations, like the Navier-Stokes, remain computationally intensive and at times infeasible for certain scenarios.

Additive-Feature-Attribution Methods

The core of the paper presents additive-feature-attribution methods centered around Shapley values, a concept derived from cooperative game theory. These methods provide a linear explanation model to interpret the impact of each input feature on the prediction. SHAP (SHapley Additive exPlanations) values stand out as unique solutions adhering to properties of local accuracy, missingness, and consistency—essential for elucidating the relationships captured by machine-learning models. The paper meticulously details four popular implementations: kernel SHAP, tree SHAP, gradient SHAP, and deep SHAP, each tailored to accommodate different model architectures and computational constraints.

Applications in Fluid Dynamics

1. Turbulence Modeling: The authors describe the application of additive-feature-attribution methods to enhance Reynolds-averaged Navier-Stokes (RANS) models. By linking physical flow properties with machine-learning-derived corrections, these methods yield interpretable, physics-informed models that capture the complexities inherent in turbulence. The ability to visualize the importance of parameters across the flow domain enhances trust and transparency, making data-driven turbulence models more robust and acceptable in engineering practices.
2. Understanding Fundamental Phenomena: The paper highlights applications for deciphering complex turbulent structures and modes. Using SHAP values, researchers can now rank the contributions of different flow features or turbulence events, providing new insights into their relative influence on flow dynamics. This aids in refining models for predicting flow behavior or for developing interventions to control turbulent flows in practical applications.
3. Applied Problems: Beyond fundamental research, the paper examines various industrial applications. These include using explainable AI to assess the impact of atmospheric conditions on airport operations, optimizing heat-transfer systems, and analyzing aerodynamic loads on structures. By quantifying the influence of design and environmental parameters, engineers can design more efficient and resilient systems.

Implications and Future Work

The implications of these methodologies are profound. They not only enhance the interpretability of machine-learning models in fluid mechanics but also lay the groundwork for integrating AI into industrial applications more broadly. The potential for energy savings and improved operational efficiency in sectors such as aerospace, automotive, and energy production is significant.

Furthermore, this review points towards an exciting trajectory for future research. Advancements in the structural understanding of multi-phase flows, optimization of rotating machinery, and noise reduction strategies could benefit vastly from the ongoing refinement of XAI methods. As computational resources and AI models continue to evolve, the intersection of these technologies with traditional fluid mechanics promises new avenues for innovation, ultimately leading to the development of more sustainable and efficient systems.

The successful application of additive-feature-attribution methods across various scenarios underscores their utility as a versatile tool in the AI-enhanced toolkit of engineers and researchers in fluid dynamics and heat transfer.