Deep Learning Models for Colloidal Nanocrystal Synthesis

Abstract

Colloidal synthesis of nanocrystals usually includes complex chemical reactions and multi-step crystallization processes. Despite the great success in the past 30 years, it remains challenging to clarify the correlations between synthetic parameters of chemical reaction and physical properties of nanocrystals. Here, we developed a deep learning-based nanocrystal synthesis model that correlates synthetic parameters with the final size and shape of target nanocrystals, using a dataset of 3500 recipes covering 348 distinct nanocrystal compositions. The size and shape labels were obtained from transmission electron microscope images using a segmentation model trained with a semi-supervised algorithm on a dataset comprising 1.2 million nanocrystals. By applying the reaction intermediate-based data augmentation method and elaborated descriptors, the synthesis model was able to predict nanocrystal's size with a mean absolute error of 1.39 nm, while reaching an 89% average accuracy for shape classification. The synthesis model shows knowledge transfer capabilities across different nanocrystals with inputs of new recipes. With that, the influence of chemicals on the final size of nanocrystals was further evaluated, revealing the importance order of nanocrystal composition, precursor or ligand, and solvent. Overall, the deep learning-based nanocrystal synthesis model offers a powerful tool to expedite the development of high-quality nanocrystals.

Figures (26)

• Figure : Fig. 1 | Schematic diagram of the segmentation and synthesis models. Upper panel: The segmentation model for nanocrystals was trained using a segmentation network combined with a semi-supervised learning strategy on the TEM image dataset, enabling instance segmentation and shape clustering of nanocrystals. The sizes obtained from the segmentation model were utilized as labels for training the synthesis model. Bottom panel: The synthesis model of nanocrystals was trained using deep learning algorithms. Condition descriptors and chemical descriptors were extracted from the synthesis recipe dataset and used as input features to predict nanocrystal size and shape with the reaction intermediate-based data augmentation method.
• Figure : Fig. 2 | Segmentation and shape analysis of nanocrystals from TEM images. a. Schematic of the semi-supervised learning process for nanocrystal segmentation. b. TEM images before and after nanocrystal segmentation. c. Clustering and visualization of 130,000 PbSe nanocrystals based on shape descriptors, where different colors represent distinct shapes: black (C1, irregular), red (C2), orange (C3, star), green (C4), blue (C5, square), purple (C6), and pink (C7). d. Proportions of the seven nanocrystal shapes synthesized at different reaction temperatures, with colors corresponding to the shape clusters.
• Figure : Fig. 3 | Development and evaluation of the nanocrystal synthesis model. a, Data processing procedure for input features to the synthesis model. b, Parity plot of size prediction using the transformer-based model. c, Visualization of input descriptors to the synthesis model using t-distributed stochastic neighbor embedding method. d, Size 2 and (e) shape evolution of PbSe nanocrystals with different molar amount of OLA and OA.
• Figure : Fig. 4 | Analysis of attention weights in the synthesis model. a, Heatmaps of the attention weights between chemicals in different layers. b, Maximum average attention weight maps of CLS for various chemicals across different transformer layers. The vertical lines indicate the maximum average attention weight of CLS for a given chemical, with the corresponding horizontal coordinate representing the chemical names and the vertical axis indicating the transformer layer. c, Categorization and distribution of highlighted chemicals across transformer layers.
• Figure : Extended Data Fig. 1 | Workflow for generating weak labels from unlabeled images by a pre-trained object detection network and morphological operations.