Vikram Ajit, Brudnak Ken, Zahid Arwa, Shim Moonsub, Kenis Paul J A
Department of Chemical and Biomolecular Engineering, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA.
Department of Materials Science and Engineering, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA.
Nanoscale. 2021 Oct 21;13(40):17028-17039. doi: 10.1039/d1nr05497j.
Colloidal semiconductor nanocrystals with tunable optical and electronic properties are opening up exciting opportunities for high-performance optoelectronics, photovoltaics, and bioimaging applications. Identifying the optimal synthesis conditions and screening of synthesis recipes in search of efficient synthesis pathways to obtain nanocrystals with desired optoelectronic properties, however, remains one of the major bottlenecks for accelerated discovery of colloidal nanocrystals. Conventional strategies, often guided by limited understanding of the underlying mechanisms remain expensive in both time and resources, thus significantly impeding the overall discovery process. In response, an autonomous experimentation platform is presented as a viable approach for accelerated synthesis screening and optimization of colloidal nanocrystals. Using a machine-learning-based predictive synthesis approach, integrated with automated flow reactor and inline spectroscopy, indium phosphide nanocrystals are autonomously synthesized. Their polydispersity for different target absorption wavelengths across the visible spectrum is simultaneously optimized during the autonomous experimentation, while utilizing minimal self-driven experiments (less than 50 experiments within 2 days). Starting with no-prior-knowledge of the synthesis, an ensemble neural network is trained through autonomous experiments to accurately predict the reaction outcome across the entire synthesis parameter space. The predicted parameter space map also provides new nucleation-growth kinetic insights to achieve high monodispersity in size of colloidal nanocrystals.
具有可调光学和电子特性的胶体半导体纳米晶体为高性能光电子学、光伏和生物成像应用带来了令人兴奋的机遇。然而，确定最佳合成条件并筛选合成配方以寻找有效的合成途径来获得具有所需光电特性的纳米晶体，仍然是加速胶体纳米晶体发现的主要瓶颈之一。传统策略往往由于对潜在机制的理解有限，在时间和资源上都很昂贵，从而严重阻碍了整个发现过程。作为回应，提出了一个自主实验平台，作为加速胶体纳米晶体合成筛选和优化的可行方法。使用基于机器学习的预测合成方法，结合自动流动反应器和在线光谱，自主合成了磷化铟纳米晶体。在自主实验过程中，同时优化了它们在可见光谱中不同目标吸收波长下的多分散性，同时利用最少的自驱动实验（2天内少于50次实验）。在没有合成先验知识的情况下，通过自主实验训练一个集成神经网络，以准确预测整个合成参数空间的反应结果。预测的参数空间图还提供了新的成核-生长动力学见解，以实现胶体纳米晶体尺寸的高单分散性。
