Researchers at Nanyang Technological University want to take it a step further. Writing in Materials Today [DOI: 10.1016/j.mattod.2020.06.010], they report that machine learning could be used to optimise the synthesis of new materials. In this study, they looked at two specific systems – 2D MoS2, grown by chemical vapour deposition (CVD), and hydrothermally-grown carbon quantum dots. In both cases, they used data taken from archived lab notebooks to ‘train’ a series of ML algorithms, and used them to predict experimental outcomes, given a set of input parameters. In addition, they applied a progressive adaptive model to optimise and speed-up the synthesis process.
For their CVD synthesis, the team’s initial focus was to better understand the role of each system variable, e.g. chamber pressure, reaction temperature, etc. They took a dataset of 300 MoS2 synthesis experiments, 183 of which had resulted in successful crystal growth, and inputted them into four different ML models. One (XGBoost-Classification) was found to accurately reproduce the results of their lab-based synthesis efforts, so was chosen for further optimisation. That analysis showed that “…the gas flow rate plays the most important role in determining whether MoS2 can be synthesized, followed by the reaction temperature and reaction time.” They used these findings to identify the combination of synthesis conditions that would be most likely to result in a successful synthesis. Ten were identified, and tested in the laboratory. MoS2 was synthesized in all cases, surpassing the 61 % success rate that had been achieved by experiment alone. By applying their progressive adaptive model, they optimised the system further, measuring a success rate of more than 86 %.
Their second target was the hydrothermal synthesis of carbon quantum dots, where they wanted to use ML to optimise a key property – the photoluminescence quantum yield (PLQY). For this, they identified six experimental parameters and assembled a database of 467 experimental records to train their ML model (in this case, XGBoost-Regression). pH was found to play the most important role, followed by the reaction temperature and reaction time. Their model identified a series of synthesis conditions that would provide the highest values of PLQY. The average value they achieved was 53.56 %, more than twice the average value of the training set.
These results highlight just how powerful machine learning can be for materials synthesis, but the authors identify some gaps. They highlight the need for a more comprehensive model that includes chemical variables like solubility and reactivity, as well as an integrated database across inorganic material synthesis systems.
This story was written for Materials Today. It was published there on 21 July 2020: https://www.materialstoday.com/computation-theory/news/making-materials-with-machine-learning/
Research paper ($): Bijun Tang, Yuhao Lu, Jiadong Zhou, Tushar Chouhan, Han Wang, Prafful Golani, Manzhang Xu, Quan Xu, Cuntai Guan, Zheng Liu. “Machine learning-guided synthesis of advanced inorganic materials”, Materials Today, article in press. DOI: 10.1016/ j.mattod.2020.06.010