Active Learning Guided Materials Synthesis and Full-map Understanding

Xiaonan Wang

 Department of Chemical and Biomolecular Engineering, National University of Singapore,

4 Engineering Drive 4, Singapore 117585, Singapore.

ABSTRACT: The rapid development of deep learning enabled artificial intelligence (AI) has brought new opportunities to accelerate conventional experimental design in materials science. The discovery of high-dimensional synthesis recipes that yield desired material properties used to be costly and time-consuming, while advances in machine learning (ML) driven high-throughput experiments are able to rapidly achieve optimal conditions to produce the target materials. In this talk, I will first provide an overview on ML tools for material discovery and sophisticated applications of different ML strategies, which shows how AI strategies are applied through material discovery stages (including characterization, property prediction, synthesis, and theory paradigm discovery. See Figure 1). Our recently developed online active learning approaches will then be discussed, which can effectively guide experiments and achieve full-map understanding of the design space. Several case studies will be demonstrated including: 1) a two-step active learning framework that combines Bayesian Optimization (BO) and Deep Neural Network (DNN) in a loop with a high-throughput microfluidic platform, to optimize the synthesis of silver nanoparticles with the desired absorbance spectrum, and 2) a hierarchical AI framework, tested for strain sensor material synthesis.

Moreover, we have developed a series of interpretation and visualization approaches to understand "black-box" ML models and make them more favoured by domain experts to extract knowledge of complex systems that embed a huge amount of hidden information. Although promising opportunities are identified, many challenges exist this highly interdisciplinary field, such as construction of valuable and open datasets, in-depth understandings of descriptors, lack of standard algorithms workflow and fully autonomous experimental platforms. We are aiming to enable a better integration of AI methods with the material discovery process, with the keys to successful applications of AI in material discovery and challenges fully addressed.

Figure 1. AI Applications through the Whole Life Cycle of Material Discovery.

Keywords：active learning; materials synthesis; deep neural network; high-throughput experiment