High-Throughput and Machine Learning Strategies for the Optimization of Optoelectronic Thin-Film Devices

Description

The optimization of optoelectronic thin-film devices poses a difficult challenge, especially when the preparation method consists of multiple parameters which influence the resulting properties. Even if the individual materials and processes are well established, their combination may not be straight-forward and requires fine-tuning of the process parameters. In this case, combinatorial and high-throughput approaches are advantageous since they allow for fast screening of the full-factorial parameter space. Further acceleration can be achieved through the utilization of machine learning methods which are able to uncover complex relations.
In a previous work, we demonstrated the performance of p-type Cu2O layers in photovoltaic devices which show high rectification and an open circuit voltage of 940 mV, thus implicating the applicability as hole-transport layers. The heterojunctions are fabricated on indium-tin-oxide (ITO)-coated glass by stacking ultrasonic spray pyrolyzed (USP) Ga2O3 with reactively sputtered Cu2O [1]. The latter material is also achieved in a water-based USP process [2], paving the way to all-sprayed photovoltaic devices. Additionally, the USP is well-suited to create combinatorial composition gradients [3], enabling high-throughput screening of dopant inclusions to further optimize the material’s performance. To this end, a platform for combinatorial device characterization is developed, designed to measure a matrix of 8x8 individual solar cells on a single sample of 25x25 mm² size.
However, the transition from established processes introduces several challenges. Foremost, the Ga2O3 layer changes the reaction kinetics of the Cu2O USP deposition, yielding a mixture of copper oxide phases. This results in impeded device performance and reduced yield of functioning cells on the 8x8 matrix. In order to obtain the USP process parameters which lead to pure Cu2O and defect-free interfaces, a data-driven approach is followed. To this end, Latin hypercube sampling [4] is employed to efficiently represent the multidimensional parameter space by only 12 deposition experiments. This preliminary data is used to obtain surrogate models, based on Gaussian processes [5], which are subjected to Bayesian optimization [6]. Following this approach, the USP process is optimized in terms of precursor concentration, flow rate, temperature, nozzle to substrate distance, nozzle speed, and number of deposition cycles, to obtain all-sprayed Cu2O-Ga2O3 heterojunctions.

[1] Dimopoulos T., et al. “Heterojunction devices fabricated from sprayed n-type Ga2O3, combined with sputtered p-type NiO and Cu2O.” Nanomaterials (2024, Accepted Manuscript).
[2] https://doi.org/10.1002/cnma.202000006
[3] https://doi.org/10.1039/D3MA00136A
[4] https://doi.org/10.1080/00401706.2000.10485979
[5] Gramacy, Robert B. Surrogates. “Gaussian process modeling, design, and optimization for the applied sciences.” CRC press (2020).
[6] https://doi.org/10.48550/arXiv.1807.02811

Period	28 May 2024
Event title	Spring Meeting of the European Materials Research Society
Event type	Conference
Location	Strasbourg, FranceShow on map
Degree of Recognition	International