Screening Efficient Tandem Organic Solar Cells with Machine Learning and Genetic Algorithms

Abstract

Tandem organic solar cells can potentially drastically improve the power conversion efficiency over single-junction devices. However, there is limited research on device development and often ca. 1% improvement over single-junction devices. Because of the complex nature of organic material compatibility and properties, such as energy-level alignment and maximizing absorption spectra, and the vastness of chemical space, computational guidance is vital. The first part of this work uses a new data set of 1225 donor/non-fullerene acceptor (NFA) pairs containing 1001 unique pairs, one of the largest to date, to train an ensemble machine learning model to predict device efficiency (RMSE = 1.60 ± 0.14%). Next, a series of genetic algorithms (GAs) are used to discover high-performing NFAs and polymer donors and then combinations of them for potential high-efficiency tandem cells. Interesting design motifs show up in high-performing NFAs, such as diphenylamine substituents on the core and 3D terminal groups. The donor polymers from the GAs reveal that arranging the monomers as a smallblock copolymer may be beneficial instead of the typical alternating copolymer. The GAs for selecting tandem cell materials successfully find material combinations that, when in a device together, have strong absorption across the entire visible−near-IR spectrum. Computational guidance is critical for the selection of tandem OSC materials, with genetic algorithms proving a highly successful technique.

Beyond Molecular Structure: Critically Assessing Machine Learning for Designing Organic Photovoltaic Materials and Devices

“…65,66 Despite the modest R 2 scores, we find that our top-performing models (RF and HGB) either match or surpass the state-of-the-art for other similar OPV datasets (Figure S2). 22,29 In particular, we achieve better predictive performance (R 2 = 0.5±0.03) using only ECFP and the material properties available in the dataset in comparison to the performance (R 2 = 0.4) achieved when using computed descriptors. 22 We also explored the ability of multi-output models for predicting PCE.…”

“…22,29 In particular, we achieve better predictive performance (R 2 = 0.5±0.03) using only ECFP and the material properties available in the dataset in comparison to the performance (R 2 = 0.4) achieved when using computed descriptors. 22 We also explored the ability of multi-output models for predicting PCE. While the most common approach is to predict PCE as the single output, PCE can also be predicted as the product of its components -JSC, VOC, and FF (Equation 1).…”

“…A similar approach has recently incorporated genetic algorithms in an effort to design high-performing non-fullerene acceptorbased (NFA) devices. [20][21][22] Others have generated one-hot encodings based on human intuition regarding molecular substructures. 23 However, the most accurate models to date are those that have used high-throughput domain-specific descriptors.…”

A universal platform of molecular orbital energy level prediction and molecular design for organic materials

“…Confronted with the challenges of processing high-dimensional, nonlinear, and multivariate data, machine learning excels in extracting key insights from vast experimental and computational datasets [29]. Machine learning extract essential data to make it an invaluable tool for developing structure-property-process models, predicting new material properties, and steering design and optimization [30][31][32]. In organic optoelectronic device research, machine learning signi cantly enhances device performance, particularly in OPVs and OLEDs, through the optimization of HOMO and LUMO levels [33,34].…”

Determining best practices for using genetic algorithms in molecular discovery

“…Although GAs have already been used for a variety of chemical applications 3,8,10,[13][14][15][16][17][18][19][20][21][22][23][24][25] , to our knowledge their convergence strategies and multiple hyperparameters have not been thoroughly investigated for molecular discovery. For example, is it better to have a small population size and allow the GA to run for many generations or a large population size with fewer generations?…”