A review of artificial intelligence approaches for the development of upconversion and downconversion phosphors

Upconversion (UC) luminescent materials enable applications in biomedical imaging, renewable energy, and anti-counterfeiting by converting near-infrared photons into visible or ultraviolet emissions. Their rational design, however, is limited by multiphoton kinetics, sensitivity to local structure, and the slow pace of empirical synthesis. In contrast, downconversion (DC) phosphors rely on simpler single-photon processes, exhibit higher chemical robustness, and dominate mature lighting and display technologies, making them a natural testbed for data-driven and Artificial Intelligence (AI)-assisted materials design strategies. This review examines how machine learning (ML) and deep learning (DL) address these barriers by predicting key UC/DC properties—including emission wavelength, quantum yield, and thermal stability—and by revealing structure–property relationships that conventional analysis often overlooks. We systematically compare ML performance in UC and DC systems, showing that while conventional models often remain effective for DC materials, UC systems generally require descriptors that explicitly encode the underlying energy-transfer mechanisms. We also highlight recent advances in inverse design using graph neural networks and reinforcement learning, as well as emerging ML strategies for understanding thermal quenching and optimizing host lattices, dopant distributions, and core–shell architectures in UC and DC systems. A detailed discussion of data quality, reproducibility, and the need for physics-informed ML underscores current limitations in UC phosphor development. Overall, this review provides an integrated perspective on how data-driven and hybrid approaches can accelerate discovery and enable more interpretable, generalizable design of next-generation UC and DC ceramic phosphors.