Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a novel Bayesian optimization framework for engineering design problems where objective and constraints are state-dependent, specifically addressing the challenge of maintaining selected state observables near trusted values (state preservation constraints). The core innovation is to transform state-space constraints into design-space constraints by pre-computing controls that satisfy linearized constraint responses within tolerance, forming an ellipsoid for efficient candidate sampling. The method is validated end-to-end on Tokamak divertor optimization, demonstrating its efficacy in a complex, anisotropic feasible set.
State-preserving constraints in Bayesian optimization are transformed into an efficient, anisotropic ellipsoid in design space, enabling robust optimization of complex engineering systems.
In many engineering design problems, the objective and constraints depend on the state: the solution of a PDE determined by the design parameters. We consider improving a design while holding selected state observables near trusted values, which we call state preservation constraints. Constrained Bayesian optimisation handles these with a learnt feasibility model, but struggles with this problem's highly anisotropic feasible set. Our central idea is to pre-compute the set of controls whose linearised constraint response stays within tolerance, thereby pulling back the state-space constraint into design space. This linearisation defines an ellipsoid from which we can efficiently draw a large number of well-spread candidates. The underlying linear response map is refined online, and the ellipsoid is rebuilt accordingly. We demonstrate the method end-to-end on our key application - Tokamak divertor optimisation under plasma-boundary preservation.