Machine learning-driven materials synthesis and design

Abstract

With the rapid progression of our society, there is a growing demand for novel materials, and the design requirements for functional materials have become increasingly complex. Yet, the trial-and-error experiments or conventional computational approaches often take months or even years to complete the research and development cycle, greatly impeding the progress in material science. Machine learning (ML) has recently emerged as a promising solution in accelerating material development, offering prediction capacities, scalability and efficiency. It has been widely applied across numerous material classes and at all stages of material development. However, driven by the significant advancements in fields like computer vision or natural language processing, the potential of applying ML in material science remain largely untapped.

The application of ML in material science faces several challenges, including the scarcity of high-quality databases, the difficulty of simultaneously meeting diverse design requirements for multiple target properties, and the need for generalizable representation of material. Additionally, a review of existing applications reveals a lack of research into guided material synthesis.

Material synthesis with the minimal number of trials is crucial for accelerating advanced materials development, and the lack of high-quality data needs to be addressed. Existing multi-variable synthesis methods are highly uncertain, time consuming and expensive. In this thesis, we first expand the application of ML to guide material synthesis, outlining our methodology for construction and optimization of model, and progressive adaptive model (PAM), for selectively increasing data in iterative loops. Two representative multi-variable systems are studies: a classification ML model for optimizing synthesis conditions of chemical vapor grown MoS2 to attain a improved success rate of synthesis, and a regression model for enhancing the process-related properties of the hydrothermal-grown carbon quantum dots (CQDs), such as the photoluminescence quantum yield. These ML models extract critical insights into the influence of synthesis parameters to experimental outcomes. Furthermore, offline analysis demonstrates that utilizing effective feedback loops in PAM can enhance experimental outcomes with a minimized number of trials, showcasing the significant potential of integrating ML into the early stages of material synthesis. This study serves as a demonstration of the concept, illustrating the feasibility and remarkable capabilities of ML in expediting inorganic material synthesis.

CQDs have versatile applications in luminescence, but identifying the optimal synthesis conditions has been challenging due to the numerous synthesis parameters and multiple desired outcomes in the real-life application, giving rising to enormous search space. However, the previously proposed PAM can not optimize multiple target properties concurrently. This thesis then presents a novel multi-objective optimization strategy that utilize ML to intelligently guide the hydrothermal synthesis of CQDs. Our closed-loop approach learns from limited and sparse data, substantially reducing the research cycle and surpassing traditional trial-and-error methods. Moreover, our approach uncovers the hidden relationships between synthesis parameters and target properties, and unifies the objective function to optimize multiple desired properties, including full-color photoluminescence (PL) wavelength and high PL quantum yields (PLQY). We successfully synthesize full-color fluorescent CQDs with PLQY exceeding 60% for all colors with only 63 experiments being performed. Our study represents a significant advancement in the ML-guided CQDs synthesis, paving the way for developing new materials with multiple exceptional properties.

Beyond guided synthesis, this thesis also employs periodic graph transformers for designing two-dimensional (2D) materials with desired optical, thermodynamic and magnetic properties. A generalizable material representation is adopted, namely multi-edge graph with periodic pattern encoding. The effectiveness of models based on graph neural networks (GNNs) and non-GNNs is carefully examined and compared. The results demonstrate that GNNs consistently outperform non-GNN-based methods exceeding a threshold of data size, across all target properties. Material insights regarding the atomic descriptors has been revealed and helped in boosting the prediction performance. In addition, the trained GNN model has been successfully deployed on unseen 2D materials with desired properties in five representative material classes. This work inspires to develop generalizable graph representation that encodes different aspects of material structures.

The ML models developed in this thesis facilitate material development in various material development stages and across multiple material classes.