Invariant Structural Causal Modeling for Counterfactual Prediction and Intervention Planning in Multiphase Thermal-Hydraulic Networks

Abstract

Multiphase thermal-hydraulic networks are governed by conservation laws but are operated and monitored through sparse, heterogeneous measurements. In gas--liquid systems, regime-dependent interfacial organization and equipment interactions induce distribution shifts across operating campaigns, so predictors that fit one condition can fail under another even when the underlying physics is unchanged. This paper develops an invariant structural causal modeling framework for counterfactual prediction and intervention planning in networked two-fluid systems. The core idea is to treat uncertain closures, sensor transformations, and unobserved disturbances as causal variables that mediate between control actions, boundary conditions, and measured signals, and to identify a representation whose residual mechanisms remain invariant across environments defined by operating policies and facility states. The approach couples a constrained mechanistic backbone with a data-driven causal layer that separates stable causal effects from spurious correlations created by changes in excitation, telemetry filtering, and latent regime mixtures. Counterfactual operators are derived by intervening on inputs and policies within the causal model while preserving invariants, enabling robust prediction of pressure and holdup outcomes under planned actuation changes. Identification is performed using a joint objective that combines likelihood, conservation-consistent residuals, and an invariance penalty enforcing independence between environment labels and structural noise. A risk-aware planning formulation then selects interventions that are robust to remaining causal ambiguity. Computational studies on annular segments and branched networks demonstrate improved out-of-environment calibration and more stable intervention ranking compared to non-invariant baselines, particularly under sensor drift and closure drift scenarios common in practice