Causal-GNN SupplyNets Enabling Resilient Semiconductor Supply Chains with Causal World Models and Lyapunov-Safe Control

Moon-Kuen Mak, Tiejian Luo

Keywords: Causality, Graph Time Series, Safe Reinforcement Learning, Federated Learning, Digital Twin, Re-entrant Queueing Networks, Semiconductor Supply Chain

TL;DR: A causal GNN world model and Lyapunov-guided RL jointly deliver robust, constraint-respecting control for semiconductor supply chains.

Abstract: The inherent cyclicality of semiconductor supply chains and the associated severe volatility pose a significant challenge to the global electronics ecosystem. During periods of tight capacity, micro-level disruptions (e.g., tool failures, yield fluctuations) are rapidly amplified through the complex network structure, leading to protracted order delivery delays and system-wide disruptions. The core problem for achieving resilience lies in making decisions based on partial, incomplete information while providing high-probability guarantees that critical operational constraints (e.g., capacity, work-in-process inventory) are satisfied. Existing approaches often decouple forecasting and decision-making, lacking either a causal understanding of intervention effects or the ability to provide provable safety guarantees, resulting in suboptimal performance in turbulent environments. To overcome these challenges, we present **Causal-GNN SupplyNets**, a framework that unifies causal reasoning with *safe constrained optimization*. Our approach introduces three key innovations: (1) We learn a graph neural network-based "world model" that incorporates macro-level causal structural priors, enabling accurate prediction of the causal effects of sudden shocks and local interventions (e.g., adjusting dispatch policies) throughout the supply network; (2) We design a Lyapunov-based safe reinforcement learning controller that *provably* optimizes material dispatch and replenishment policies while satisfying safety constraints with high probability; (3) We introduce a privacy-preserving federated distillation mechanism, allowing different organizations to collaboratively improve their interventional knowledge without sharing raw sensitive data. Extensive experiments in simulated environments and on anonymized real-world manufacturing data demonstrate that our method significantly outperforms baseline models across various load and shock scenarios. It consistently improves on-time delivery rate (**up to 17 percentage points at peak load**), shortens cycle times, and accelerates post-shock recovery. Ablation studies further confirm that the causal constraints are crucial for accurate counterfactual prediction, and the Lyapunov safety guard is necessary for ensuring *near-zero* constraint violations. Our work provides a new pathway for achieving provable resilient control in highly uncertain and dynamic complex networks.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 17660