Causal Digital Twin-Enabled Deep Reinforcement Learning for Energy-Efficient and Explainable Resource Allocation in 6G Networks

The proliferation of sixth-generation (6G) networks has intensified the need for intelligent resource allocation strategies that simultaneously satisfy stringent quality-of-service requirements and minimize energy consumption. However, conventional deep reinforcement learning approaches often suffer from poor generalization under distribution shifts and provide limited explainability regarding their decision-making processes. We propose a Causal Digital Twin framework that constructs a structural virtual replica of the physical 6G network by encoding causal dependencies among base stations, user equipment, and edge servers through a causal graph neural network. This twin captures the underlying causal mechanisms governing network behavior rather than merely fitting statistical correlations. Furthermore, invariant risk minimization and disentangled representation learning are employed to ensure that the learned representations remain stable across diverse traffic regimes, hence improving robustness under varying operational conditions. The resulting differentiable simulator enables gradient-based actor-critic optimization within a constrained Markov decision process formulation, where energy efficiency is maximized subject to latency and rate constraints. Counterfactual rollouts are then generated to assess the causal impact of specific allocation decisions, thereby providing explainability and facilitating sim-to-real transfer. The proposed method is expected to achieve substantial energy savings while maintaining service reliability, and the causal structure of the digital twin offers interpretable insights into the factors driving each decision.