(134a) Optimization of Process Parameters of Quantum Dot Synthesis Aided By High-Throughput Experimentation and Big Data Techniques

Keywords: Particle formation, focusing/defocusing, mixing, artificial neural networks, support vector machines

Quantum Dots (QD), i.e. small semiconductor nanocrystals that exhibit quantum confinement, have received special interest in the past years due to their unique size-dependent opto-electronic properties leading to applications in emerging fields including photovoltaics, display technologies, biology and telecommunication. However, reproducible production of well-defined materials with high quality is still a clear challenge. To address this issue, a robot for automated particle synthesis under inert conditions was installed that enables high reproducibility, fast sampling down to 6 s and gives unique insight to process-structure and therefrom based structure-property relationships. The automated platform contains a reaction module with six reactors that can be heated up to 340 °C internally which is stirred with magnetically coupled triple blade or anchor stirrers. Thus, the flow field inside the reactors is well controlled and closer to a larger scale stirred vessel compared to magnetic stirrer bars that are usually applied in lab scale syntheses.

As benchmark process of high complexity, synthesis of CdSe QDs by hot injection routine was selected. It is based on the fast injection of a cold reactant (~25 °C) into a hot precursor solution (260 °C) under vigorous stirring [1]. After a rapid burst of nuclei, growth and finally slow Ostwald ripening define the final dispersity. Inline temperature monitoring and rapid sampling to 96 well micro titer plates (MTPs) showed outstanding reproducibility of the robotic system. This was demonstrated by particle size distributions (PSDs) [2] derived from optical absorbance spectra as well as independent results gained by transmission electron microscopy (TEM) and analytical ultracentrifugation (AUC) [3]. Especially the latter clearly revealed highly reproducible, however multimodal samples, that were resolved with Å-resolution. Finally, reproducibility in combination with early stage sampling and controlled mixing allowed us to systematically analyze the influence of stirring on focusing and defocusing of PSDs, that was expressed in terms of the evolution of the relative standard deviation (RSD) [4].

From these findings, we deduce that mixing is a decisive quantity for generating suspensions with optimum properties, i.e. minimum dispersity. However, because of the manifold parameters affecting the mixing time -namely the mixing speed, reaction volume, injection speed, injection location and impeller typ- the parameter space becomes quite large. Therefore the use of high throughput experimentation (HTE) is needed. For that, design of experiments (DoE) is one option as it utilizes building a statistical model from fewer experiments than what would be normally required, saving a substantial amount of material and time. In addition, using machine learning –for our case artificial neural networks (ANNs) and support vector machines (SVMs)- can lead to a similar outcome.

Thus, in the second part of our contribution, we will present how DoE as well as machine learning algorithms were applied to obtain a predictive model for the mixing time as decisive quantity for mixing controlled particle synthesis during hot injection. First, we experimentally explored a large parameter space using a cold model that includes mixing speed, reaction volume, injection speed, injection location and impeller type and identified their effect on the equivalent mixing time, i.e. the mixing time in water at room temperature. The latter was derived by the Villermaux-Dushmann protocol that was designed to resemble the hot injection procedure applied in this study [5-7].

Then, a predictive statistical model and machine learning algorithms for the equivalent mixing time at various conditions were developed and tested with combinations of different parameters (mixing speed, volume, injection speed, injection location and stirrer type) within this multidimensional space. Finally, it was analyzed to what extent the cold model is able to predict for a given impeller Reynolds number the occurrence of focusing and defocusing for different stirrer geometries and, most important with respect to larger scale industrial production, different filling volumes of the reactor. Noteworthy, experimental findings agreed very well, both qualitatively and quantitatively with the expectations.

In conclusion, such in-depth understanding of process parameters on quantum dot formation is a very important step for transferring lab scale syntheses efficiently towards larger reaction volumes. This paves the way towards the development of optimized routes for high quality particles with minimized dispersity.

Acknowledgements

The authors want to thank the funding of Deutsche Forschungsgemeinschaft (DFG) through the Cluster of Excellence “Engineering of Advanced Materials” and project SE 2526/1-1.

Literature

[1] E.M. Chan, C. Xu, A.W. Mao, G. Han, J. Owen, B.E. Cohen, and D.J. Milliron, Nano Lett., 10 (2010) 1874-1885

[2] D. Segets, J. Gradl, R. Klupp Taylor, V. Vassilev, and W. Peukert, ACS Nano, 3 (2009) 1703-1710

[3] J. Walter, K. Löhr, E. Karabudak, W. Reis, J. Mikhael, W. Peukert, W. Wohlleben, and H. Cölfen, ACS Nano, 8 (2014) 8871-8886

[4] A.M. Salaheldin, J. Walter, P. Herre, J.M. Kolle, I. Levchuk, C.J. Brabec, W. Peukert, and D. Segets, Chem. Eng. J., 320 (2017) 232-243

[5] P. Guichardon and L. Falk, Chem. Eng. Sci., 55 (2000) 4233-4243

[6] P. Guichardon, L. Falk, and J. Villermauy, Chem. Eng. Sci., 55 (2000) 4233-4253

[7] J.-M. Commenge and L. Falk, Chem. Eng. Proc., 50 (2011) 979-990