COBALT: COnstrained Bayesian optimizAtion of computationaLly expensive grey-box models exploiting derivaTive information

• Novel constrained grey-box optimization framework using Gaussian process models. New almost everywhere differentiable acquisition function for composite functions. Efficient moment-based approximation of chance constraints. Tailored algorithm enrichment sub-problem that exploits model structure. Performance comparison with Bayesian on diverse set test problems. Many engineering problems involve the computationally expensive models which derivative information is not readily available. The (BO) a particularly promising approach solving these problems, uses (GP) and an expected utility to systematically tradeoff between exploitation exploration design space. BO, however, fundamentally limited by black-box assumption does take into account any underlying problem In this paper, we propose new algorithm, COBALT, combines multivariate GP novel whose structure can be exploited state-of-the-art nonlinear programming solvers. COBALT compared traditional BO seven including calibration genome-scale bioreactor experimental data. Overall, shows very performance both unconstrained