(403b) Machine Learning-Guided Flow Synthesis of Inorganic Metal Halide Perovskite Quantum Dots | AIChE

(403b) Machine Learning-Guided Flow Synthesis of Inorganic Metal Halide Perovskite Quantum Dots


Epps, R. - Presenter, North Carolina State University
Volk, A., North Carolina State University
Abdel-Latif, K., North Carolina State University
Reyes, K. G., Princeton University
Abolhasani, M., NC State University
The controlled synthesis of advanced colloidal nanomaterials is limited by the currently available reactors utilized for their synthesis and the disjointed workflow between experimentation and data processing. Manual flask-based experimentation strategies, although applied across most research laboratories, are unlikely to sufficiently capture and control the complex and highly sensitive reaction systems of many colloidal nanomaterial syntheses. These batch reactors are slow, labor-intensive, and subject to batch-to-batch variations in experimental data sets. In response, over the last decade different microfluidic synthesis strategies have been developed and effectively applied towards the controlled synthesis of different classes of colloidal nanomaterials. Such flow synthesis strategies notably enhance the efficiency of colloidal synthesis studies, while simultaneously unlocking previously unattainable precision in synthesis parameter control. Despite the effectiveness of flow synthesis microreactors in accelerating reaction condition screening of colloidal nanomaterials, they still rely on user-directed experiment selection. Broad reaction space screening using design of experiments is materially inefficient, and expertly guided experiments are often too slow to effectively cover a synthesis space within the lifetime of unstable chemical precursors. These bottlenecks can be addressed through development of self-optimizing fluidic reactors and integrating fully automated experimentation, materials diagnostics, and data analysis with artificial intelligence-guided experiment-selection algorithms. Such autonomous materials synthesis platforms will be able to isolate optimal reaction parameters entirely without user intervention.

Colloidal semiconductor nanocrystals (known as quantum dots) are a particularly suitable testbed for autonomous materials development strategies. Quantum dot syntheses possess a large and highly sensitive reaction parameter space.[1] Proper control of their reaction conditions is difficult to attain both within a single set of precursors and across research labs (batch to batch variation). As a result, despite extensive efforts and constantly growing applications, many synthesis paths are either unexplored or not fully optimized. One of the recently emerging classes of quantum dots is colloidal metal halide perovskites, which have highly favorable properties for applications in optoelectronic devices. Due to the combinatorially increasing synthesis space of metal halide perovskite quantum dots, many of their possible synthesis routes are novel and unexplored. Therefore, these reaction systems present an exciting opportunity for exploration within an autonomous flow synthesis platform.

In this work, we present the first machine learning guided-fluidic reactor for the controlled synthesis of high-quality colloidal quantum dots. While the strategies and techniques detailed here may be easily applied to a large number of solution-processed materials, as a case study we analyzed the halide exchange of cesium-lead-bromide perovskite quantum dots with zinc halide salts and two different surface ligands.[2] We developed a novel in situ material diagnostic module to attain accurate real-time access to peak emission energy, emission linewidth, and photoluminescence quantum yield of the in-flow synthesized perovskite quantum dots. These output parameters, corresponding to the physicochemical and optoelectronic properties of quantum dots, were then simultaneously optimized through an objective function integrated within a custom neural network ensemble model and various decision-making policies (used to select conditions from the model). Utilizing the developed autonomous materials development platform, we evaluated the performance of eight different experiment selection algorithms (spanning machine learning to evolutionary algorithms) over the course of more than 1400 experiments. The controlled syntheses of perovskite quantum dots were performed without user intervention and required less than 400 µL of starting quantum dot solution per experimental condition. The methods detailed in this work offer autonomous reaction optimization across multiple vital performance parameters. Further use of similar self-optimizing systems in the field of colloidal nanoscience would reduce the cost and expedite the process of materials development and discovery.


[1] R. W. Epps, K. C. Felton, C. W. Coley, M. Abolhasani, Lab Chip 2017, 17, 4040.

[2] K. Abdel‐Latif, R. W. Epps, C. B. Kerr, C. M. Papa, F. N. Castellano, M. Abolhasani, Adv. Funct. Mater. 2019, 29, 1900712.