Simulation-based optimization is widely used to improve the performance of an inventory system under
uncertainty. However, the black-box function between the input and output, along with the expensive
simulation to reproduce a real inventory system, introduces a huge challenge in optimizing these performances.
We propose an efficient framework for reducing the total operation cost while satisfying the
service level constraints. The performances of each inventory in the system are estimated by kriging
models in a region-wise manner which greatly reduces the computational time during both sampling
and optimization. The aggregated surrogate models are optimized by a trust-region framework where a
model recalibration process is used to ensure the solution’s validity. The proposed framework is able to
solve general supply chain problems with the multi-sourcing capability, asynchronous ordering, uncertain
demand and stochastic lead time. This framework is demonstrated by two case studies with up to
18 nodes with inventory holding capability in the network.