A symbiotic generative adversarial network framework for supply chain stress-testing

This paper introduces a symbiotic conditional generative adversarial network (c-GAN) framework to address the critical challenge of generating disruption scenarios that are novel relative to simulated known risks (model-relative unknowns) in complex supply networks. Traditional predictive models fail against Black Swan–like events (within the modelled system) due to their reliance on historical data and inductive inference limitations. Our framework explores the tail-risk space of plausible disruption scenarios beyond the simulated known-unknown boundary by conditioning on network topology metrics through a specialised c-GAN architecture. These synthetic disruption vectors integrate with an adaptive inoperability input-output model (AIIM) that simulates cascading failures using tri-load coordination mechanisms. The framework incorporates rigorous evaluation metrics to quantify the novelty of the situation and its systemic impact. A symbiotic human-AI interface enables domain experts to validate plausibility and to steer scenario generation through explainable AI and latent-space regulation. Experimental validation in a 127-node automotive supply network demonstrates the framework's capability to systematically generate scenarios with 60.5% higher simulated cascading failure impact and 64.3% greater relative simulated economic losses than the simulated known-unknown scenarios used for training.