Physical Simulators as Do-Operators: Causal Discovery under Latent Confounders for AI-for-Science

Researchers introduce CFM-SD, a causal discovery method that leverages physical simulators to identify cause-and-effect relationships in scientific domains while handling latent confounders—a common problem in molecular design and materials science. The approach achieves significantly higher accuracy than existing methods and demonstrates practical improvements in real-world applications like toxicity prediction and battery optimization.

CFM-SD addresses a fundamental limitation in causal discovery: existing methods assume perfect causal sufficiency (no hidden variables) and rely on cheap virtual interventions, assumptions that break down in real scientific experimentation. This research reframes physics-based simulators as causal operators within Pearl's interventional framework, enabling researchers to work with expensive, real-world simulation data while accounting for unmeasured confounders that plague materials and molecular science.

The theoretical contribution proves that causal structure becomes identifiable with only O(d) single-variable interventions—matching the theoretical minimum under physical constraints. This efficiency matters significantly because running molecular dynamics simulations or quantum chemistry calculations costs hours or days per data point, making sample efficiency crucial for practitioners.

The empirical validation spans both controlled benchmarks and genuine scientific problems. Synthetic experiments show F1 scores of 0.800 versus 0.127-0.562 for competing methods, while real-world evaluations demonstrate 57-58% bias reduction in toxicity prediction and electrolyte optimization—metrics that translate directly to faster drug discovery cycles and better battery materials.

This work bridges a gap between causal inference theory and computational science practice. Rather than forcing scientists to choose between statistical rigor and computational realism, CFM-SD enables both. The ability to incorporate domain knowledge through physics-based simulations while maintaining causal validity opens new applications in drug discovery, materials design, and climate modeling where both computational expense and hidden variables present genuine obstacles.

• →CFM-SD integrates physical simulators into causal discovery, handling latent confounders that plague scientific research.
• →Theoretical framework achieves identifiability with O(d) interventions, matching the physical realizability minimum.
• →Achieves 0.800 F1 score on synthetic benchmarks versus 0.127-0.562 for baseline methods.
• →Real-world validation demonstrates 57-58% bias reduction in molecular toxicity and battery electrolyte optimization.
• →Enables efficient causal inference when interventions are computationally expensive, directly applicable to drug and materials discovery.