Surrogate modeling with insufficient data can lead to high prediction uncertainty and errors. A promising remedy to address this issue is the use of transfer learning techniques that leverage models built using data from other problems that are implicitly related to the problem of interest. We present an algorithm that uses transfer learning and mixtures of experts across different design space regions to improve the predictive capability of surrogate models. The algorithm uses existing data to divide a problem's design space into clusters and build ensembles of surrogate models in each cluster using a multi-criteria weighting method. The proposed algorithm is shown to be both accurate and flexible, allowing for automated transfer learning with tuning parameters that cater for different problem types. The multi-criteria approach enables transfer learning in constrained Bayesian optimization by weighing models based on their shape, accuracy, and variance. The proposed method is demonstrated using aircraft conceptual design examples and showed up to 10\% reduction in prediction errors.