On the Generalization in Topology Optimization via Sensitivity-Conditioned Bernoulli Flow Matching

Researchers introduce sensitivity-conditioned Bernoulli flow matching to improve out-of-distribution generalization in topology optimization surrogate models. By conditioning on adjoint sensitivities—the gradient information that drives classical optimization—the approach achieves state-of-the-art performance across structural and computational fluid dynamics benchmarks under distribution shifts like changing loads and boundary conditions.

This research addresses a fundamental challenge in physics-informed machine learning: why surrogate models trained on one set of conditions fail dramatically when conditions change. The authors propose that the root cause is information loss in the conditioning signal, establishing an information-theoretic framework through the Data Processing Inequality to explain generalization failures.

The work bridges classical optimization theory with modern generative modeling. Traditional topology optimization relies on adjoint sensitivity fields to guide design updates. The authors hypothesize that reproducing this information structure in neural surrogate models should preserve the same generalization properties. Rather than using expensive exact sensitivities, they introduce the concept of pseudo-sensitivities—approximations derived from physical fields that monotonically relate to true sensitivities—making the approach practical.

The empirical validation is compelling. Testing on structural optimization tasks with load variations and a new CFD-TO dataset with boundary condition shifts demonstrates that sensitivity conditioning substantially outperforms naive parameter conditioning. The framework quantifies information content across different physical fields, providing theoretical justification for architectural choices in surrogate design.

For the engineering and scientific computing community, this represents progress toward more robust AI-assisted design tools. The methodology could extend beyond topology optimization to other inverse design problems where sensitivity information is naturally available. The release of datasets and code amplifies impact by enabling reproducibility and follow-up work. However, the practical adoption depends on computational costs of sensitivity computation relative to the improvement gains.

• →Information-theoretic analysis shows adjoint sensitivities are optimal conditioning signals for topology optimization generalization
• →Pseudo-sensitivities enable practical approximation of expensive adjoint computations while preserving generalization benefits
• →Sensitivity-conditioned flow matching achieves state-of-the-art out-of-distribution performance under load and boundary condition shifts
• →New CFD-TO dataset provides benchmark for evaluating generalization in computational fluid dynamics topology optimization
• →Framework offers general design principle for physics-informed surrogate models across inverse design applications