Robot- and Machine-Learning-Accelerated Discovery of Complex Materials

Hybrid inorganic/organic materials and doped nanocrystal heterostructures offer modular platforms for generating libraries of materials with complex structure, composition, and properties that can be tailored to applications such as lasers, lighting, and imaging. The large number of components in these complex materials, and the numerous parameters that specify their synthesis, result in an experimental space that is challenging to explore with conventional laboratory techniques. I will discuss the development of data-enabled strategies for discovering new materials through the combination of robotic synthesis, physical models, and machine learning (ML). We used large (>17,000 sample) datasets generated by robotic synthesis to train machine learning classification models for the crystallization of metal halide perovskites and to elucidate the reaction networks that govern chemical transformations of perovskites. Our robot- and ML-accelerated techniques are also used to investigate the complex photophysical networks that govern the optical properties of lanthanide-doped upconverting nanoparticles (UCNPs). Automated synthesis of these materials, coupled with photophysical models validated by high-throughput datasets, have facilitated the discovery of new upconverting materials including photon avalanching nanoparticles that enable sub-diffraction-limited NIR imaging and patterning. To guide automated synthesis, we leverage a high-throughput computational workflow that can autonomously tune UCNP nanostructures in silico for targeted applications. We use kinetic Monte Carlo to simulate the emission spectra of candidate UCNP structures recommended by Bayesian Optimization workflows over iterative cycles. Differentiable, physics-informed neural networks trained on computational datasets enable rapid inverse design of multi-shell UCNPs while domain-specific knowledge is still being developed.