Predicting Indium Phosphide Quantum Dot Properties from Synthetic Procedures Using Machine Learning
Predicting Indium Phosphide Quantum Dot Properties from Synthetic Procedures Using Machine Learning
复制标题
使用机器学习从合成过程预测磷化铟量子点特性
DOI：
10.1021/acs.chemmater.2c00640
复制
发表时间：
2022
影响因子：
8.6
通讯作者：
Cash, Melanie
中科院分区：
文献类型：
--
作者：
Nguyen, Hao A.;Dou, Florence Y.;Park, Nayon;Wu, Shenwei;Sarsito, Harrison;Diakubama, Benedicte;Larson, Helen;Nishiwaki, Emily;Homer, Micaela;Cash, Melanie
利用机器学习（ML）预测化学反应结果已成为推进材料合成的有力工具。然而，这种方法需要庞大而多样的数据集，由于文献中不一致和非标准化的报道以及缺乏对合成机制的理解，这些数据集在纳米材料合成领域极为有限。在这项研究中，我们从72篇出版物中提取了InP量子点（QD）合成的参数作为我们的输入，并将所得性质（吸收，发射，直径）作为我们的输出。我们使用数据输入方法“填充”缺失的输出，以准备包含216个条目的完整数据集，用于训练和测试预测ML模型。我们基于化学特性或试剂的作用以两种方式（浓缩和扩展）定义描述符空间，以探索对输入特征进行分类的最佳方法。使用我们最好的ML模型，我们的吸收、发射和直径的平均绝对误差（MAEs）分别低至20.29、11.46和0.33 nm。我们使用这些模型来部署一个可访问的交互式web应用程序，用于设计InP的合成（https://share.streamlit.io/cossairt-lab/indium-phosphide/Hot_injection/hot_injection_prediction.py）。通过这个web应用程序，我们研究了InP合成的化学趋势，例如锌盐和三辛基膦等常见添加剂的影响。我们还设计并进行了基于文献程序扩展的新实验，并将实验测量的属性与预测进行了比较，从而评估了我们模型的“现实生活”准确性。相反，我们使用逆设计来获得具有特定性质的InP量子点。最后，我们通过扩展先前发布的数据集，应用相同的方法来训练、测试和启动CdSe量子点的预测模型。总之，我们的数据预处理方法和ML实现展示了即使面对有限的数据资源，也能设计具有目标属性的材料并探索潜在的反应机制。
Prediction of chemical reaction outcomes using machine learning (ML) has emerged as a powerful tool for advancing materials synthesis. However, this approach requires large and diverse datasets, which are extremely limited in the field of nanomaterials synthesis due to inconsistent and nonstandardized reporting in the literature and a lack of understanding of synthetic mechanisms. In this study, we extracted parameters of InP quantum dot (QD) syntheses as our inputs and resultant properties (absorption, emission, diameter) as our outputs from 72 publications. We “filled in” missing outputs using a data imputation method to prepare a complete dataset containing 216 entries for training and testing predictive ML models. We defined the descriptor space in two ways (condensed and extended) based on either chemical identity or the role of reagents to explore the best approach for categorizing input features. We achieved mean absolute errors (MAEs) as low as 20.29, 11.46, and 0.33 nm for absorption, emission, and diameter, respectively, with our best ML model. We used these models to deploy an accessible and interactive web app for designing syntheses of InP (https://share.streamlit.io/cossairt-lab/indium-phosphide/Hot_injection/hot_injection_prediction.py). Using this web app, we investigated chemical trends in InP syntheses, such as the effects of common additives, like zinc salts and trioctylphosphine. We also designed and conducted new experiments based on extensions of literature procedures and compared our experimentally measured properties to predictions, thus evaluating the “real-life” accuracy of our models. Conversely, we used inverse design to obtain InP QDs with specific properties. Finally, we applied the same approach to train, test, and launch predictive models for CdSe QDs by expanding a previously published dataset. Altogether, our data preprocessing method and ML implementations demonstrate the ability to design materials with targeted properties and explore underlying reaction mechanisms even when faced with limited data resources.