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

The paper studies why topology optimization surrogate models generalize unevenly under load and boundary-condition shifts.

Using a causal Markov chain view and the data processing inequality, it argues that adjoint sensitivity is the strongest conditioning signal.

The authors introduce pseudo-sensitivities to separate informative physical fields from weaker ones.

A sensitivity-conditioned Bernoulli flow-matching model achieves the best out-of-distribution performance on structural and CFD benchmarks.