|
|
Autonomous laboratories based on robotics and artificial intelligence (AI) has been recently conducted widely for accelerated search of materials. On the other hand, an inverse design that suggest specific synthesis conditions to achieve various target properties is required for bespoke synthesis of materials. However, because the inverse design needs to solve questions in a high-dimensional parameter space, it is greatly complex, so that researchers need to perform enormous numbers of experiments for the inverse design to find the optimal synthesis condition. It is very difficult to predict the next synthesis condition by understanding the correlations between experimental data via human intelligence. Usually, material properties have been individually explored with a Bayesian optimization (B.O.) model in recent autonomous laboratories; likely leading to successive problems caused by a number of experiments and a lot of physical time although it is more efficient than the high-throughput screening (HTS) process. The conventional B.O. does not share experimental data simultaneously during the entire optimization process. In other words, the inverse design for bespoke synthesis of materials must be performed in parallel for the B.O. models while sharing experimental data in the single variable space. This parallelization allows to accumulate experimental results in the same parameter space without redundant experiment conditions. We tried to optimize the synthesis of silver nanoparticle (Ag NPs) to demonstrate the efficiency of parallelized B.O., in which a home-made automatic apparatus was used for synthesis of Ag NPs. Our automatic synthesis system and B.O. model were used to identify the optimal synthesis condition for various combinations of optical target properties (e.g., λmax, full width half maximum, intensity, etc.) at the same time. Then, we compared the efficiency of HTS, conventional B.O., and parallelized B.O. with increasing the numbers of input synthesis variables and target properties. Our work provides a strong potential to solve the high-dimensional space issue for bespoke design of materials via an autonomous laboratory. In addition, our demonstration of autonomous nanoparticle synthesis will be useful in energy applications such as catalysis, photovoltaics which are made up of nanoparticles.
|
|
Autonomous laboratories based on robotics and artificial intelligence (AI) has been recently conducted widely for accelerated search of materials. On the other hand, an inverse design that suggest specific synthesis conditions to achieve various target properties is required for bespoke synthesis of materials. However, because the inverse design needs to solve questions in a high-dimensional parameter space, it is greatly complex, so that researchers need to perform enormous numbers of experiments for the inverse design to find the optimal synthesis condition. It is very difficult to predict the next synthesis condition by understanding the correlations between experimental data via human intelligence. Usually, material properties have been individually explored with a Bayesian optimization (B.O.) model in recent autonomous laboratories; likely leading to successive problems caused by a number of experiments and a lot of physical time although it is more efficient than the high-throughput screening (HTS) process. The conventional B.O. does not share experimental data simultaneously during the entire optimization process. In other words, the inverse design for bespoke synthesis of materials must be performed in parallel for the B.O. models while sharing experimental data in the single variable space. This parallelization allows to accumulate experimental results in the same parameter space without redundant experiment conditions. We tried to optimize the synthesis of silver nanoparticle (Ag NPs) to demonstrate the efficiency of parallelized B.O., in which a home-made automatic apparatus was used for synthesis of Ag NPs. Our automatic synthesis system and B.O. model were used to identify the optimal synthesis condition for various combinations of optical target properties (e.g., λmax, full width half maximum, intensity, etc.) at the same time. Then, we compared the efficiency of HTS, conventional B.O., and parallelized B.O. with increasing the numbers of input synthesis variables and target properties. Our work provides a strong potential to solve the high-dimensional space issue for bespoke design of materials via an autonomous laboratory. In addition, our demonstration of autonomous nanoparticle synthesis will be useful in energy applications such as catalysis, photovoltaics which are made up of nanoparticles.
|
|
|
|
Sang Soo Han is a principal research scientist at Computational Science Research Center in Korea Institute of Science and Technology (KIST). Also, he has been serving as the Head of Computational Science Research Center in KIST since 2020. Before joining KIST in 2013, he worked in Korea Research Institute of Standards and Science as a senior research scientist during 2009-2013. He also worked at Department of Chemistry in California Institute of Technology as a post-doctoral scholar during 2005-2009. He received the B.S. and M.S. degrees from Hanyang University, Korea, in 1999 and 2001, respectively, and the Ph. D. degree from Korea Advanced Institute of Science and Technology in 2005. He has a specialty of multiscale simulations such as reactive force field for catalysis, porous materials, and batteries. Recently, his research interests have been extended to data-driven materials design, in particular, bypassing quantum mechanical calculations and inverse design. Moreover, he is developing autonomous labs for development of nanomaterials based on AI and robots. |
