Bayesian optimization with executable digital twins: Fast controller tuning with multi-source information

High-performance controller tuning in industrial systems is still constrained by costly and time-consuming on-hardware experimentation. Although simulation models are widely available, they are often used only for offline verification and not as active information sources during optimization. This paper proposes the Adaptive Twin-in-the-Loop architecture, a framework that runs an ensemble of real-time digital twins in parallel with the physical plant to accelerate controller tuning in real applications. We develop a multi-source Bayesian Optimization method that jointly models real and virtual data through a multi-output Gaussian Process. A dedicated batch acquisition function then selects virtual evaluations that maximally reduce uncertainty in the most promising region of the real-system objective. Statistical validation on standard benchmarks and experimental testing on a brushless direct current motor speed controller show that the proposed framework accelerates convergence, reaching near-optimal closed-loop performance, on average, with only one-third of the on-hardware trials required by single-source Bayesian Optimization.