Large Language Model-Guided Bayesian Optimization of Ag+ Doping in CdSe Nanoplatelets Enables Deterministic Optical Prediction across the Gamut

Heterovalent doping is a versatile strategy for modulating the otherwise discrete optical properties of cadmium selenide nanoplatelets. However, identifying the correlation between ideal synthetic parameters and desired optical outcome is challenging due to the large experimental space. We present a machine learning–driven application for rapidly optimizing post-synthetic Ag+ doping in CdSe nanoplatelets. Using Bayesian Optimization with In-Context Learning (BO-ICL) on large language models, our approach directly integrates continuous and categorical synthesis and cationexchange experimental variables to predict and improve photophysical properties. Specifically, we jointly optimize the total quantum yield and dopant peak area to capture the trade-off between emissive efficiency and spectral tunability. Compared to a Gaussian Process baseline, the LLM-based optimizer recommends high-performing experimental protocols with fewer trials. Moreover, we illustrate interpretability via SHAP values, revealing that doping percentage, dopant mass, reaction volumes, and NPL dimensions are the most influential factors in controlling the photoluminescence response. In addition, we demonstrate the use of accurate multidimensional color predictions from BO-ICL, without excessive fine-tuning, to obtain experimentally accurate optical responses on a CIE chromaticity plot. This one-step method underscores the promise of large-language models to accelerate materials innovation and holds the potential to guide the rational design of doped nanocrystal systems.