Interpreting CFD Surrogates through Sparse Autoencoders

Surrogate models need better transparency for real-world use: The paper identifies that AI-based CFD surrogates, while faster than traditional simulations, have "opaque" internal representations that prevent their adoption in "safety-critical or regulation-bound settings"—highlighting a key barrier to responsible deployment in engineering applications.

Sparse autoencoders can decode what AI models learn: The method successfully extracts interpretable features from surrogate models' internal representations, identifying specific physical phenomena like "vorticity or flow structures," demonstrating a practical technique for making black-box models more understandable to domain experts.

Interpretability is achievable after model training: By applying this "posthoc interpretability framework" to already-trained models, the approach avoids requiring changes to model architecture or training procedures, making it potentially easier to retrofit existing surrogate systems with explainability features.

Physical meaning matters for trustworthiness: The paper emphasizes that extracted features align with actual physical phenomena rather than arbitrary patterns, which is crucial for building confidence in AI systems used for engineering decisions where understanding why a model makes predictions is as important as accuracy.

Model-agnostic tools increase responsible AI adoption: The framework works across different graph-based surrogate architectures, suggesting that interpretability solutions don't require redesigning entire systems—a practical pathway for organizations to enhance trustworthiness in existing AI deployments.