Leveling Up Upconverting Nanoparticles with Machine Learning

ConspectusUpconverting nanoparticles (UCNPs) transform low-energy light into higher-energy photons, enabling applications in subwavelength and subsurface imaging, nanoscale sensing, therapeutics, optogenetics, printing, and optical computing. However, the widespread adoption of UCNPs is hindered by their low brightness and limited spectral tunability. Predicting the ideal nanoparticle architectures to overcome these limitations is challenging because UCNP photophysics are governed by highly nonlinear, complex energy transfer networks that span the excited states of lanthanide dopants. Due to the large number of possible combinations of dopants, concentrations, host matrices, heterostructures, and reaction conditions, optimizing the compositional and synthetic parameters of UCNPs using conventional trial-and-error approaches is intractable.This Account explores how researchers can overcome these challenges and enhance the properties of UCNPs using artificial intelligence (AI) and machine learning (ML). We first review how the early foundations of AI-guided discovery were established with automated experimental workflows and physical modeling. Using robotic synthesis platforms and differential rate equation models, researchers have successfully navigated high-dimensional compositional spaces to reveal optical phenomena, such as energy looping and photon avalanching, in nanoparticles.Building on these data-driven approaches, ML has been integrated into UCNP research initially for processing raw characterization data, such as automating the analysis of TEM images and time-resolved luminescence curves. AI approaches have been extended to interpret signals in applications that utilize UCNPs, such as classifying the cytotoxicity of drugs based on upconversion luminescence microscopy data. Most significantly, ML is driving the design of new UCNP compositions and structures, including our recent development of closed-loop active learning of UCNP core-shell heterostructures. By coupling Bayesian optimization with kinetic Monte Carlo (kMC) simulations, we achieved 110-fold enhancement in UCNP emission over just 40 iterations. To bypass the steep computational cost of simulating UCNP heterostructures with up to 9 shells, we leveraged differentiable deep learning surrogate models based on heterogeneous graph neural networks to perform inverse design. Notably, these hetero-GNNs were able to extrapolate far outside of the model's training data and predict UCNP heterostructure compositions with 6.5-fold more intense emission than the brightest UCNP in the training set.In the future, we predict that AI/ML approaches will become integral to the UCNP research. UCNP experiments may soon be accelerated by autonomous self-driving laboratories in which robotic synthesis, in-line characterization, and ML agents operate in a closed feedback loop to intelligently investigate underexplored chemical spaces. Large language models (LLMs) could parse literature to develop overarching hypotheses and detailed recipes for these autonomous workflows, with generative models suggesting novel structures to test. Together with human creativity and critical analysis, these AI tools will accelerate the discovery of advanced upconverting nanomaterials, aiding fundamental understanding of their mechanisms and inspiring a broader array of photonic applications.