Understanding Hot Injection Quantum Dot Synthesis Outcomes Using Automated High-Throughput Experiment Platforms and Machine Learning

Machine learning (ML) has demonstrated potential toward accelerating synthesis planning for various material systems. However, ML has remained out of reach for many materials scientists due to the lack of systematic approaches or heuristics for developing ML workflows for material synthesis. In this work, we report an approach for selecting ML algorithms to train models for predicting nanomaterial synthesis outcomes. Specifically, we developed and used an automated batch microreactor platform to collect a large experimental data set for hot-injection synthesis outcomes of CdSe quantum dots. Thereafter, this data set was used to train models for predicting synthesis outcomes using various ML algorithms. The relative performances of these algorithms were compared for experimental data sets of different sizes and with different amounts of noise added. Neural-network-based models show the most accurate predictions for absorption and emission peak, while a cascade approach for predicting full width at half-maximum was shown to be superior to the direct approach. The SHapley Additive exPlanations (SHAP) approach was used to determine the relative importance of different synthesis parameters. Our analyses indicate that SHAP importance scores are highly dependent on feature selection and highlight the importance of developing inherently interpretable models for gaining insights from ML workflows for material synthesis.