【Abstract】Driven by critical challenges in biomanufacturing, including high complexity and high uncertainty, we propose global optimization methods on the policy-augmented Bayesian network (PABN), characterizing risk- and science-based understanding of underlying bioprocess mechanisms, to guide the optimal control. We first develop a sequential optimization algorithm based on deep kernel learning (DKL) for PABN with general state transition dynamics, which can learn the spatial dependence of mean response through a deep neural network. In addition, to improve the interpretability and computational efficiency of policy optimization, a global metamodel is introduced to guide linear Gaussian PABN optimization, which explicitly accounts for the correlation of input-to-output pathways obtained under different candidate policies. Our empirical study provides the ablation analysis and the interpretation analysis of the DKL, and also shows that both proposed approaches demonstrate promising performance compared to the standard Bayesian optimization with Gaussian process.