Robots are taking up the challenges in photovoltaics R&D

High-throughput experimentation platforms and machine learning have shown preliminary success in materials research acceleration. Recently in Joule, Du, Brabec, and colleagues showed the potential of using a robotic platform to optimize organic photovoltaics (OPV) materials and devices. With a short experimental time of 70 h, it was demonstrated to achieve a power conversion efficiency of 14% and a photostability duration of 50 h. High-throughput experimentation platforms and machine learning have shown preliminary success in materials research acceleration. Recently in Joule, Du, Brabec, and colleagues showed the potential of using a robotic platform to optimize organic photovoltaics (OPV) materials and devices. With a short experimental time of 70 h, it was demonstrated to achieve a power conversion efficiency of 14% and a photostability duration of 50 h. Human beings’ demand for energy has made a huge impact on climate change, i.e., 73.2% of CO2 emission is related to energy usage.1Ritchie H. Roser M. Emissions by sector - Our World in Data.https://ourworldindata.org/emissions-by-sector#energy-electricity-heat-and-transport-73-2Google Scholar It is imperative to rapidly develop and deploy the technologies of renewable energy to achieve the carbon-neutral goal. Materials scientists at the 2016 IC6 workshop of Mission Innovation have asked a key question: “Can we tackle the climate challenge in energy materials innovations by utilizing robotic high-throughput platforms and artificial intelligence (AI)?” Within the past few years, we’ve seen the “future laboratory” coming into reality. For example, the flow synthesis robot plans the synthesis route based on all available reaction data in the literature and completes the chemical synthesis experiment autonomously.2Coley C.W. Barzilay R. Jaakkola T.S. Green W.H. Jensen K.F. Prediction of Organic Reaction Outcomes Using Machine Learning.ACS Cent. Sci. 2017; 3: 434-443https://doi.org/10.1021/acscentsci.7b00064Crossref PubMed Scopus (343) Google Scholar The mobile robot chemist was able to complete the entire optimization cycle of 688 experiments in 8 days with 10 process variables without any human intervention.3Burger B. Maffettone P.M. Gusev V.V. Aitchison C.M. Bai Y. Wang X. Li X. Alston B.M. Li B. Clowes R. et al.A mobile robotic chemist.Nature. 2020; 583: 237-241https://doi.org/10.1038/s41586-020-2442-2Crossref PubMed Scopus (238) Google Scholar Most recently published in Joule,4Du X. Lüer L. Heumueller T. Wagner J. Berger C. Osterrieder T. Wortmann J. Langner S. Vongsaysy U. Bertrand M. et al.Elucidating the Full Potential of OPV Materials Utilizing a High-Throughput Robot-Based Platform and Machine Learning.Joule. 2021; 5: 495-506https://doi.org/10.1016/j.joule.2020.12.013Abstract Full Text Full Text PDF Scopus (25) Google Scholar X. Du, C.J. Brabec, and colleagues showed us the impressive potential of a robot-based automated experimental platform “AMANDA Line One” in the rapid development of organic photovoltaics (OPV) devices (Figure 1). In contrast to the conventional experimental methods of trial and error or one-variable-at-time planning, the authors demonstrated that this robotic platform completed quick cycles of fabrication, characterization, and analysis. With 70 h and 100 process conditions, two crucial performance metrics of photovoltaics devices, i.e., power conversion efficiency (PCE) and operational stability illumination, were optimized autonomously. These process conditions were sampled in 10-dimensional parameter space: donor and acceptor weight ratio; solution concentration; spin speed, active-layer annealing temperature; active-layer annealing time; solvent additives; etc. The optimal process conditions were found for the best PCE of 14% under the standard testing condition and the best device photostability of 50 h under degradation testing. For high-dimensional analysis of the process parameter space, a machine-learning regression model (particularly, Gaussian process regression [GPR] was used) can be easily trained with the experimental dataset. Some insights about the process-PCE correlation and process-stability correlation were obtained with the predictive machine-learning model, e.g., (1) annealing temperatures above 140°C for active layer generally reduced the PCE, and longer annealing time leads to instability of the devices; (2) a clear trend that PCE is increasing with an increase in spin speed, especially for low-temperature annealing of the active layer, can be observed; and (3) donor and acceptor weight ratio around 1:1.2 is the optimum for PCE, but a higher ratio shows better stability. All these physical insights could not be easily drawn because human brains are not well trained to analyze a 10-dimensional dataset. For most people, navigating in a 4-dimensional parameter space is already a challenging task. In addition, AMANDA Line One also provides more reliable experimental results (by reducing process variations), uses a shorter experimental time, and consumes a smaller amount of materials. The authors of this work have undoubtedly shined some light on the future of materials research by combining the advancement in the fields of robotics and machine learning. We see the advantages that the robotic platform can free up materials researchers from some repetitive lab work and give them significantly more time to define the experimental hypothesis (rather than conduct process optimization). However, we have also found two crucial questions to be addressed in the community of applied AI development. First, will these automated tools worsen the inequity among different research groups? A fully automated system like this one requires a significant amount of capital investment, and this capital requirement will become a barrier for many smaller groups to access the highly productive research capability. We need to find a way to reduce these barriers. The operation model of “cloud labs” or “paid by hours” could be an interesting approach to democratize these automation tools. Another way could be to democratize the materials data instead of democratizing the tools. However, we must define sustainable incentives for researchers to share their data openly and timely so that everyone can have access to the latest experimental databases. Second, do we want all the lab works to be automated? We’ve heard so many fascinating stories for “discovery by accident” in the history of science. Without getting our hands dirty in the labs, we might easily miss the chance for materials discovery by accident. One possible solution is to develop specialized machine-learning algorithms that can explore the unknown experimental space and alert us about any outlier points. On the other hand, we still need to find a better way to work collaboratively with these robots and machine-learning algorithms. We might even need to fundamentally change the education program so that those future material scientists have the mindsets and skillsets to work collaboratively with the robots. Nevertheless, in the new era of AI, the combination of high-throughput experimental platforms and machine-learning algorithms will improve research productivity and accelerate materials development.5Correa-Baena J.-P. Hippalgaonkar K. van Duren J. Jaffer S. Chandrasekhar V.R. Stevanovic V. Wadia C. Guha S. Buonassisi T. Accelerating Materials Development via Automation, Machine Learning, and High-Performance Computing.Joule. 2018; 2: 1410-1420https://doi.org/10.1016/j.joule.2018.05.009Abstract Full Text Full Text PDF Scopus (123) Google Scholar Especially, with the pressing challenge of climate change, we must break the silos of discipline and use the best tools we have for the innovation tasks of clean energy materials. Robot-Based High-Throughput Screening of Antisolvents for Lead Halide PerovskitesGu et al.JouleJuly 14, 2020In BriefA robotic platform is adopted to conduct a comprehensive solvent engineering for making lead halide perovskites in a high-throughput manner. Deeper insights into the working mechanisms and selection criteria of antisolvents are investigated and summarized. In addition, a reliable antisolvent database is established, and verification tests match well with the theory. Furthermore, our work provides significant guidance for designing functional and environment-friendly mixed solvent systems to fabricate high-quality perovskite materials or devices. Full-Text PDF Open Archive