A High-Performance Monte Carlo Simulation Toolbox for Uncertainty Quantification of Closed-loop Systems

Abstract

We apply Monte Carlo simulation for performance quantification and tuning of controllers in nonlinear closed-loop systems. Computational feasibility of large-scale Monte Carlo simulation is achieved by implementation of a parallelized high-performance Monte Carlo simulation toolbox for closed-loop systems in C for shared memory architectures. The toolbox shows almost linear scale-up on 16 CPU cores on a single NUMA node, and a scale-up of 27.3 on two NUMA nodes with a total of 32 CPU cores. We demonstrate performance quantification and tuning of a PID controller for a bioreactor in fed-batch operation. We perform 30,000 closed-loop simulations of the fed-batch reactor within 1 second. This is approximately a 2300 times computational performance increase compared to a serial reference implementation in Matlab. Additionally, we apply Monte Carlo simulation to perform automatic tuning of the PID controller based on maximizing average produced biomass within 8 seconds.