LLM-Guided Digital Twin for Network Experimentation: Fidelity, Robustness, and Optimization

Evaluating communication networks traditionally requires costly simulations or physical testbeds. This work presents a closed-loop framework that combines a Large Language Model (LLM) for scenario generation, a Random Forest (RF) surrogate for rapid prediction of key performance indicators (KPI), and an NS-3 digital twin (DT) for validation.The LLM produces schema-constrained scenarios, the RF filters scenarios by predicted throughput, latency, and loss, and high-value cases are executed in the DT with results fed back to refine the RF model and guide subsequent exploration.Experiments demonstrate three outcomes: fidelity, with the digital twin consistently reproducing baseline KPIs across batches (e.g., Cubic-FQ-CoDel maintains 71.13±0.01 Mbps); robustness, with accurate capture of degradation and recovery under failures and identification of resilient protocol-queue combinations (NewReno-CoDel: 95%+ retention); and optimization, with Pareto analysis highlighting Westwood+RED for lowest latency ( 13 ms) and Cubic+RED as overall best (performance score: 23.48).The framework achieves near-perfect prediction accuracy (R2 ≈ 0.99), reduces simulation overhead, and offers interpretable insights for multi-objective optimization. These results position LLM-driven DTs as a practical tool for efficient, data-driven network experimentation.