Researchers at National Taiwan University (NTU) have unveiled an artificial intelligence workflow that merges multimodal GPT models with system dynamics, offering a faster way to assess how circular economy policies could reshape resource flows.
Using Taiwan’s material flow data from 2013 to 2022, the team simulated nine scenarios to track progress toward the country’s 2030 circularity goals.
The framework allows GPT to read and interpret stock-and-flow diagrams that map industrial activities, emissions, and resource pathways, then translate them into executable system-dynamics models. A six-stage pipeline—labeled GPT1 through GPT6—was designed to automate tasks from diagram parsing to policy sensitivity testing, cutting weeks from the time typically needed to configure large-scale simulations.
Results suggest that Taiwan’s circular material use rate, currently 22%, could climb to 29% by 2030 under the most ambitious mix of measures, including stronger recycling incentives, industrial symbiosis programs, and expanded eco-design mandates. Resource productivity was projected to rise from NT$65 to NT$88 per kilogram of material input, signaling a potential decoupling of economic growth from raw material consumption.
Lead author Hwong-wen Ma noted that the approach does more than speed up calculations: by embedding language models in systems modeling, researchers can refine assumptions and iterate scenarios with less manual coding. That agility, he said, is essential as governments attempt to balance climate, waste-reduction, and industrial policy goals in the face of volatile commodity markets.
While the NTU study focuses on Taiwan, its authors argue the method is transferable to other regions where material flow accounts and policy data are available. Scaling the approach could help policymakers stress-test circularity strategies against economic or technological shocks, aligning investments with targets such as the EU’s Circular Economy Action Plan or Japan’s resource-efficiency roadmap.
