As machine learning (ML) increasingly dominates the analysis of complex flows, the field of fluid mechanics faces a critical challenge: moving beyond "black box" predictions toward a deeper, physics-based understanding of why these models work. This review by Andrés Cremades, Sergio Hoyas, and Ricardo Vinuesa explores the intersection of fluid mechanics and Explainable Artificial Intelligence (XAI), focusing on how additive-feature-attribution methods can bridge the gap between high-performance data-driven models and fundamental physical principles.
The Need for Explainability in Fluid Dynamics
Fluid dynamics is governed by the Navier–Stokes equations, which are notoriously non-linear and multi-scale. While deep learning models are exceptionally good at predicting these complex behaviors, their lack of transparency is a major hurdle for scientific discovery and industrial safety. If an engineer cannot determine why a model predicts a specific turbulence pattern or heat transfer rate, they cannot fully trust it for critical applications. XAI provides the necessary "explainability layer" to validate that models are learning actual physics rather than statistical noise.
Methodology: Additive Feature Attribution
The paper focuses on additive-feature-attribution methods, which explain a complex model by creating a simplified, linear local approximation. The gold standard in this domain is the Shapley value, a concept from cooperative game theory that fairly distributes the model's prediction among the input features.
The authors detail four primary SHAP (SHapley Additive exPlanations) implementations used in the field:
- Kernel SHAP: A model-agnostic method using weighted linear regression.
- Tree SHAP: An efficient algorithm designed for tree-based models like Random Forests.
- Gradient SHAP: A method for deep neural networks utilizing output gradients.
- Deep SHAP: A high-speed approximation that recursively passes importance values backward through neural network layers.
Key Insights: Physics Validation and Global Importance
A central finding of the review is that XAI allows researchers to check if a model is "physics-compliant." For instance, SHAP analysis has confirmed that models correctly identify "sweeps" and "ejections"—key turbulent structures—as the most influential factors for skin friction in channel flows.
Furthermore, the authors highlight that physical magnitude does not always equate to importance. SHAP can reveal subtle, low-magnitude features that have a disproportionately large impact on the outcome. While SHAP is a "local" explanation for specific predictions, aggregating these values provides a "global" map of feature importance, identifying sensitivities across entire datasets.
Applications and Results
The synthesis of various studies reveals significant breakthroughs across multiple domains:
- Turbulence Modeling: In RANS (Reynolds-averaged Navier–Stokes) closures, XAI identified the viscosity ratio ($q_9$) and strain-rotation ratio ($q_4$) as the most influential parameters for correcting model errors, as shown in the feature importance analysis below.
- Aviation Safety: In predicting low-level turbulence during aircraft landing, terrain effects were found to be more influential than distance from the runway or wind direction.
- Heat Transfer: For solar air heaters, SHAP values ranked rib height and transverse pitch as the most critical design parameters for optimizing thermal efficiency.
- Pollutant Dispersion: In urban environments, the height-width ratio of buildings was shown to be the dominant factor in $CO_2$ concentration levels.
Challenges and Future Directions
Despite its potential, XAI faces challenges such as high computational costs for calculating exact Shapley values and difficulties in handling highly correlated input features. The authors suggest that future work should move toward "interpretable-by-design" models, where physical constraints are integrated into the architecture itself, guided by the insights gained from current XAI methods. Ultimately, these tools may lead to the discovery of entirely new physical laws derived directly from data.
Summary updated