Intervention-Based Time Series Causal Discovery via Simulator-Generated Interventional Distributions

Researchers introduce SVAR-FM, a framework that uses physics-based simulators to discover causal relationships in time series data by treating simulation interventions as Pearl's do operator. The method recovers correct causal directions where observational methods fail due to confounding, with theoretical guarantees and empirical validation across multiple scientific domains.

SVAR-FM represents a significant methodological advance in causal inference by bridging simulation science and statistical learning. The framework addresses a fundamental limitation of observational causal discovery: confounding bias that produces reversed causal estimates. By leveraging physics simulators as mechanical implementations of interventions—physically clamping variables to sever confounding paths—the approach generates ground-truth interventional data that purely statistical methods cannot access.

This work builds on decades of causal inference research, particularly Pearl's do-calculus framework, but innovates by recognizing that domain-specific simulators can operationalize counterfactual reasoning. The theoretical contribution is substantial: proving structural VAR identifiability under coverage conditions and deriving decomposed error bounds clarifies when and why the method succeeds. The sign-flip corollary provides practical predictive power, demonstrating that simulator fidelity directly determines whether causal estimates reverse.

For scientific research and engineering domains, this methodology enables more reliable causal discovery in complex systems where randomized experiments are infeasible or unethical. The ultrafast laser physics case study validates predictions empirically, showing R² = 0.983 accuracy when simulator accuracy is sufficient. This validates the theoretical predictions about fidelity thresholds.

The broader impact extends to any field employing mechanistic simulators—climate science, drug discovery, materials engineering, and systems biology. Organizations can now systematically improve causal estimates by improving their underlying simulators. However, real-world applicability depends on simulator availability and accuracy for specific domains.

• →SVAR-FM uses physics simulators as mechanical interventions to solve causal discovery problems observational methods cannot handle
• →Theoretical guarantees prove structural VAR identifiability when simulators can clamp decision variables
• →A sign-flip corollary predicts causal estimates reverse when simulator accuracy falls below domain-specific thresholds
• →Empirical validation across four scientific domains confirms recovery of correct causal signs where confounding previously caused reversals
• →Framework generalizes to any domain with mechanistic simulators, enabling systematic improvement of causal inference through simulator refinement