Abstract for: Numerical Optimization in System Dynamics Models: Calibration and Optimal-Policy Search in World3

The World3 model couples population, capital, agriculture, resources, and pollution to simulate the global socio-environmental system. Its eleven published scenarios explore different policy assumptions, but none was designed to be optimal with respect to a stated criterion. Applying optimization requires overcoming numerical barriers: non-differentiable tabulated functions, iterative delay solvers incompatible with automatic differentiation, and high dimensionality from time-discretization. World3 is re-implemented in JAX and embedded in the GEMSEO multidisciplinary optimization framework. Spline interpolations replace piecewise-linear tables, delay variables are integrated into the monolithic state vector, and automatic differentiation provides exact gradients. Two applications are demonstrated: simultaneous five-sector calibration of 55 model constants against open-access historical data, and optimal control with seven policy levers over a 100-year horizon under two objectives (a Human Welfare Index and an Ecological Footprint). Calibration improves alignment with observations across all sectors, but also reveals model limitations. The two objectives produce distinct optimal strategies traceable to different activations of World3's feedback loops. Their Pareto front shows that intermediate policies exist with moderate welfare cost and reduced environmental impact. The entire workflow runs in under two minutes on a single CPU core. The optimizer requires stating which objective is being pursued, making explicit the normative choices that scenario analysis embeds in parameter selection. The Pareto front is conditional on the chosen discount rate, which biases toward near-term outcomes. The methodology is transferable to more detailed Integrated Assessment Models. Text: spellchecking, revision. Code: application code from one-sector example