Lanthanide-activated inorganic scintillators underpin radiation detection in medical imaging, high-energy physics, space science and nuclear security, yet are almost always used ?off the shelf? rather than designed for the application. A strict prerequisite for scintillation is that the activator?s 4f ground state and lowest 5d? excited state both lie within the host band gap. This talk presents a physics-informed machine learning strategy that, by combining established chemical-shift trends for lanthanide 4f and 5d binding energies with first-principles band edges, places the activator levels in any candidate host and turns scintillator-versus-non-scintillator classification into a rapid, high-throughput screen. The approach is demonstrated on ~5 million double-perovskite oxides via hierarchical down-selection, on Ce-doped garnets using a fast ceramic-pellet feedback loop, and on chloride elpasolites, where ML-predicted compositions were grown and shown to luminesce. Challenges specific to data-driven materials design?small, biased datasets, domain-knowledge integration and slow experimental feedback?are discussed throughout.