We propose a semi-supervised domain transfer approach for bridging the sim2real gap in simulation-based inference, leveraging mini-batch optimal transport for scalable, inductive posterior approximation.