Proceedings of nanoGe Fall Meeting 2021 (NFM21)
DOI: https://doi.org/10.29363/nanoge.nfm.2021.100
Publication date: 23rd September 2021
X. Rodríguez-Martínez1, E. Pascual-San José1, Zhuping Fei2, A. Sanchez-Díaz1, A. Harillo1, M. Heeney2, R. Gimerà3, M. Campoy Quiles1
1Institut de Ciència de Materials de Barcelona (ICMAB-CSIC), Spain
2Chemistry Department, Imperial College London, UK
3Universitat Rovira I Virgili, Spain
While the strong performance increase in organic photovoltaics is driven by the synthesis of improved materials, the highly entangled thickness-composition-morphology parameter space makes almost impossible the prediction of performance for new systems. One notable exception is the thickness dependence of the active layer, which was formulated in a seminal paper from 1999 [1]. On the other hand, the dependence of the device performance on donor/acceptor ratio has thus far been more elusive; despite composition being one of the parameters that affects efficiency more strongly. In order to address this complex problem, in this talk we will present the results of feeding high-throughput experimental methods into artificial intelligence algorithms in order to predict the composition-performance landscape.
We will first describe a novel methodology for the fast evaluation of donor/acceptor systems for photovoltaics. The new approach is based on the fabrication of samples with gradients in the relevant parameters of interest that represent a large fraction of the corresponding parameter space. In particular, we fabricate gradients in thickness, microstructure, composition and apply hyperspectral imaging to correlate material and device properties. The method is up to 100 faster than conventional optimization protocols, uses less than 50 mg of each active layer material and generates hundreds to millions of data points per system [2,3].
Then we show how this machinery can be used to find design rules for the optimum composition in non-fullerene acceptors based devices. For this, the combinatorial evaluation of more than 15 donor/acceptor systems -which generates thousands of data points in the thickness/composition parameter space-, is coupled to machine learning algorithms. Trained algorithms can predict the composition for maximum efficiency, and even reproduce multi-maximum efficiency vs composition diagrams, from basic material inputs, such as energy levels and optical and electrical traits [4].