CFM-SD: Causal Discovery with Physical Simulators as Do-Operators

The recently introduced CFM-SD (Causal Flow Matching with Simulation Data) method employs first-principles physical simulators as do-operators within Pearl's interventional calculus, addressing latent confounders and authentic interventional data in AI-for-Science applications, such as molecular design and materials science. Current approaches (IGSP, DCDI, ENCO) presuppose causal sufficiency and depend on virtual interventions, which is impractical given the prevalence of latent confounders and the high cost of real interventions like physics-based simulations. Theoretically, CFM-SD can identify d-variable causal structures with O(d) single-variable interventions, the least required under physical realizability constraints. In tests on synthetic data (γ=0.2–0.8), CFM-SD recorded an average F1 score of 0.800, compared to 0.127–0.562 for existing methods. This research is available on arXiv under ID 2605.07467.