Controller retuning now screens stability first: compressor simulation deploys zero unstable candidates

Controller retuning now carries a stability screen before deployment: a candidate parameter change must not only pass the performance criterion but also keep a sufficient fraction of posterior local-dynamics models closed-loop stable. The earlier ACE procedure treated the performance requirement only as a pass/fail classification, using a Gaussian-process classifier and an acquisition function to find a minimal parameter change, without explicitly quantifying how uncertainty in the local closed-loop dynamics affects candidate stability. On an axial compressor-plenum-throttle simulation, three fixed instances were each run 100 times; UQ-ACE produced zero invalid closed-loop trajectories, while ACE produced 0.37, 0.42, and 0.22.

The screen works by learning a family of plausible local dynamical models from closed-loop trajectories already collected during tuning, then checking whether a candidate controller stabilizes those models. The construction fits the trajectories with a Bayesian posterior over sparse identification of nonlinear dynamics (SINDy), using a sparsity prior to select a few library terms rather than committing to one deterministic model. In the simulation each posterior sample is linearized at a stable operating point to form a closed-loop matrix whose spectral radius is compared against a stability margin, and a candidate is tested only if the fraction of stabilizing posterior models exceeds the prescribed threshold.

The reported scope is retuning on an axial compressor-plenum-throttle simulation benchmark, with a PID controller including derivative filtering and a four-parameter search, where stability is judged by the posterior closed-loop spectral radius lying inside a given margin. It is intended for researchers and engineers working on data-driven tuning of safety-critical loops who want to judge whether such a stability screen fits their own tuning loop.

The conclusions rest on simulation, with no plant validation reported. The authors list improved unstable-loop detection and convergence-based diagnostics as future work, and note that extension beyond PID tuning to nonlinear controllers remains open. In the embedded-state instances, ACE itself produced no unstable trajectories, so there UQ-ACE mainly adds a robustness layer.