This paper proposes a machine learning–based multifidelity modeling (MFM) and information-theoretic Bayesian optimization approach where the associated models can have complex discrepancies among each other. Advantages of MFM-based optimization over a single-fidelity surrogate, specifically under complex constraints, are discussed with benchmark optimization problems involving noisy data. The MFM framework, based on modeling of the varied fidelity information sources via Gaussian processes, is augmented with information-theoretic active learning strategies that involve sequential selection of optimal points in a multiscale architecture. This framework is demonstrated to exhibit improved efficiency on practical engineering problems like high-dimensional design optimization of compressor rotor via implementing its multiscale architecture and calibration of expensive microstructure prediction model. From the perspective of the machine learning–assisted design of multiphysics systems, advantages of the proposed framework have been presented with respect to accelerating the search of optimal design conditions under budget constraints.