The conflict between strength and toughness is a fundamental problem in engineering materials design. However, systematic discovery of microstructured composites with optimal strength-toughness trade-offs has never been demonstrated due to the discrepancies between simulation and reality and the lack of data-efficient exploration of the entire Pareto front. Here, we report a widely applicable pipeline harnessing physical experiments, numerical simulations, and artificial neural networks to efficiently discover microstructured designs that are simultaneously tough and strong. Using a physics-based simulator with moderate complexity, our strategy runs a data-driven proposal-validation workflow in a nested-loop fashion to bridge the gap between simulation and reality in high sample efficiency. Without any prescribed expert knowledge of materials design, our approach automatically identifies existing toughness enhancement mechanisms that were traditionally discovered through trial-and-error or biomimicry. We provide a blueprint for the computational discovery of optimal designs, which inverts traditional scientific approaches, and is applicable to a wide range of research problems beyond composites, including polymer chemistry, fluid dynamics, meteorology, and robotics.