有机合成与机器学习结合发现钙钛矿太阳能电池的最佳空穴传输层
一个由德国研究机构 Forschungszentrum Jülich 的分支机构 ——Helmholtz Institute Erlangen-Nürnberg for Renewable Energy（HI ERN）牵头的国际团队，创新性地将有机合成与预测模型相结合，在闭环流程中展开探索，成功发现了新的功能性材料。这种被称为混合方法的研究模式，在寻找钙钛矿太阳能电池中空穴传输层材料方面发挥了关键作用。该方法的独特之处在于，借助基于实验数据、机器学习（ML）技术以及通过进一步高通量实验验证的大型分子描述符数据库。
研究团队发现，在运用机器学习和自主优化算法时，充足的数据量固然重要，数据的多样性同样不可或缺。为了攻克这一难题，团队率先构建了一个结构公式数据库，其中囊括了约 100 万个可由市售物质制备的虚拟分子。这一数据库的建立，为后续研究提供了坚实的数据基础。
展望未来，该研究小组指出，他们计划把材料发现与器件优化融入一个无缝衔接的闭环流程。研究小组着重强调：“达成这一目标，需要在跨学科研究中齐心协力，融合材料科学、工程学以及先进计算技术等多领域的见解，从而构建起一个协同共进的工作流程。”
最近发表在《Science》杂志上的论文“Inverse design workflow discovers hole-transport materials tailored for perovskite solar cells,”中描述了这项研究工作。
The best hole transport layers for perovskite solar cells
An international team has combined organic synthesis with predictive models to discover new functional materials that enhance performance of hole transport layers used in perovskite solar cells. The team asserts that optimizing for other solar cell properties is possible with the platform, as well as using it for development of materials for other kinds of devices.
An international team led by Helmholtz Institute Erlangen-Nürnberg for Renewable Energy (HI ERN), a branch of German research institute Forschungszentrum Jülich, has combined organic synthesis with predictive models to discover new functional materials in a closed-loop process. The so-called hybrid approach was applied to discovering material for hole transport layers in perovskite solar cells. It featured the use of a large molecular descriptor database based on experiments, machine learning (ML) and validation through further high-throughput experiments.
The team found that to be able to use machine learning and autonomous optimization algorithms requires not only a sufficiently large data volume but also data diversity. To address that challenge it began with a database of structural formulas of around one million virtual molecules that could be produced from commercially available substances.
Density functional theory (DFT) calculations and other computational tools were used and tailored to calculate a large set of possible hole transport materials (HTMs) for enhancing the performance of perovskite solar cells based on relevant structural features. “Then we synthesized about 100 initial materials and tested them in solar cells, which allowed us to train machine learning models to select two more batches of about 20 materials in order to maximize the solar cell device efficiency,” the research's corresponding author, Pascal Friederich, told pv magazine.
Molecules with differing characteristics were automatically produced at HI ERN with the help of a robotic system and used to manufacture otherwise identical solar cells, dopant-free HTMs in p-i-n structured perovskite solar cell devices. The results for the series of HTMs showed an initial power conversion efficiency exceeding 20% after Bayesian optimization, according to the paper. These materials reached an efficiency as high as 26.2% and a certified efficiency of 25.9%.
Furthermore, the devices maintained 80% of the initial performance for more than 1000 hours, which was attributed to the passivating properties of the new HTM, suppressing nonradiative recombination at the interface more effectively. “For model selection, we trained different ML models on a random 10-fold cross-validation of the 101 experimental molecular data points. Tested ML models included random forest regression, linear regression, neural networks, GP regression, and kernel-ridge regression. All simple models performed equally well,” noted the team.
“We were able to use the high throughput (HT) synthesis we developed for the wide range of molecules, and couple experimental and theoretical data to discover new hole-transport materials using Bayesian optimization. This is a promising research direction that uses machine learning to learn and understand, rather than just accelerating discovery,” Anastasia Barabash, research team member, told pv magazine, stressing that the trained ML models can be used to learn more about the underlying relationship between microscopic molecular properties and macroscopic device performance.
Research first author, Jianchang Wu, told pv magazine that one of the challenges of the study was to purifiy hundreds of molecules consistently, maintaining reproducibility from batch to batch, as well as keeping uniform conditions of the “many environmental, crystallization, and process factors that can “substantially” affect the outcome. Such models can be further explored as autonomous workflows “to identify and predict further novel molecules” tailored for a variety of applications, noted the group, adding that it could also be used to predict perovskite device performance of new molecular designs using design rules extracted and “elucidated from a fully trained model.”
Looking ahead, the group said it will integrate material discovery and device optimization into a seamless, closed-loop process. “Achieving this will require a concerted effort in interdisciplinary research, combining insights from materials science, engineering, and advanced computational techniques to create a synergistic workflow,” it said.
The research work is described in the paper “Inverse design workflow discovers hole-transport materials tailored for perovskite solar cells,” recently published in Science.
The scientists participating in the study were from Helmholtz Institute Erlangen-Nürnberg for Renewable Energy (HI ERN), a branch of Forschungszentrum Jülich, Karlsruher Instituts für Technologie (KIT), Friedrich Alexander University Erlangen Nürnberg (FAU), Ulsan National Institute of Science in South Korea, in China, University of Groningen, in Netherlands, as well Xiamen University and the University of Electronic Science and Technology, both based in China.
