Predicting Indium Phosphide Quantum Dot Properties Using Machine Learning on Synthetic Procedures

Indium phosphide quantum dots (InP QDs) have emerged as a strong competitor amongst the traditional Cd- and Pb- based materials in the past few years for lighting, displays, and other optoelectronic technologies. Extensive studies over the last decade have been devoted to synthetic optimization for controlling the size uniformity and emissive properties of InP QDs. This creates a timely opportunity to compile the data and pin down the synthetic design rules governing the size, absorption, and emission characteristics of InP QDs.<br/><br/>Recently, machine learning (ML) has emerged as a useful technique in the field of chemical synthesis to optimize chemical reactions, design new procedures, study underlying mechanisms, and simulate chemical processes. In 2020, Santos and coworkers reported the application of ML on quantum dot synthesis, where important synthetic factors were identified in predicting the final size of II-VI and IV-VI QDs.<br/>Using a similar methodology, we have trained and optimized several ML models in two ways, using single and multiple outputs, to predict the diameter, absorption, and emission wavelength of InP QDs based on their synthetic conditions. The dataset was manually extracted from 179 publications. A data augmentation technique was applied to expand the dataset to include 216 syntheses. Both the single output and multiple output models showed high accuracy with mean absolute errors (MAE) as low as 0.33 nm for diameter, and 11.46 nm and 20.29 nm for emission, and absorption wavelength respectively. The ML models also identified nucleation temperature, reaction time, and addition of zinc as the most influential parameters for the synthesis. To allow external users to design syntheses based on the best models, we deployed an accessible and interactive webapp that can be accessed through the URL https://share.streamlit.io/cossairt-lab/indium-phosphide/streamlit/st_all.py. Finally, eight new experiments were designed based on extensions of literature procedures and conducted to confirm the accuracy of the webapp. The experimental results showed MAEs around 24 nm for both absorption and emission wavelengths and 0.35 nm for diameter when compared with the predicted values from the webapp. These results demonstrate the capability of ML in predicting chemical synthesis and contribute to our understanding of InP QD synthesis.