Combining Machine Learning and Simulation for Photovoltaics

/ Fluxim AG/

Here's what we've got to share this month:

  • 👩‍🔬 Project Results - AIPV (AI-Assisted Thermal and Electrical Characterization of Large-Area PV Modules)

  • 💻 New Software Release- Laoss 4.1

  • 📈 The Revolutionary Characterization Platform

  • 📺 Scientific Tutorial Videos

  • 🤝 We're hiring - Join the Fluxim software team.

👩‍🔬 Research Project Results

 

AIPV Project
(AI-Assisted Thermal and Electrical Characterization of Large-Area PV Modules)

An Innosuisse funded project between ZHAW and Fluxim

Contributors from ZHAW: E. Comi, C. Kirsch, E. Knapp

Project duration: 2019 -2021

AIPV Project.png

AIPV Goals
To detect defects in inorganic, organic, and hybrid  (e.g perovskite) solar panels during the manufacturing process through the use of applied experimental characterization, modeling, and simulation of photovoltaic cells with our software Laoss and Setfos. Machine Learning (ML) and traditional methods are used and compared to estimate model parameters. 

The AIPV project is a collaboration between research groups at Fluxim AG and Zurich University of Applied Sciences (ZHAW). The project is coming to a positive end, and we're pleased to be able to share some of  the impactful results here: 


Scientific Results

1.     XGBoost trained on synthetic data to extract material parameters of organic semiconductors

        E Knapp, M Battaglia, T Stadelmann, S Jenatsch, B Ruhstaller
        8th Swiss Conference on Data Science, Lucerne, Switzerland, 9 June 202
        DOI: 160.85.104.64/handle/11475/22414

XGBoost Trained on Synthetic Data to Extract Material Parameters of Organic Semiconductors.png

"We successfully trained an XGBoost model on the synthetic data to a multi-target regression problem to determine underlying material parameters from current voltage and impedance spectroscopy measurement data. The ML network was first trained by synthetic data generated from our (Setfos) before measured data (Paios) was presented to the model. The trained ML system was able to determine the underlying physical device parameters, as demonstrated by the close agreement between simulation and measurement. The concept of merging machine learning and physical modeling for data generation is a powerful alternative to classical fitting algorithms provided the simulation times for the physical modeling are short."

2.   Silicon solar cell parameter estimation by convolutional neural networks trained on simulated imaging data
     
      Mattia Battaglia, Evelyne Knapp, …, Beat Ruhstaller et al.
     (not yet available online)

In previous work (doi.org/10.1016/j.solener.2020.08.058), it was demonstrated that Laoss can be used to quantify defects in solar modules. In this work, the authors investigated whether the step of manually fitting the parameters of the solar cell and the defects can be replaced by machine learning.

In their approach, a convolutional neural network (CNN) was trained on a set of simulated EL images and later applied to real measured data to estimate the underlying physical model parameters of the solar cells and the defects. This method would then allow one to design and train a CNN that can almost instantaneously create a digital twin of the measured sample without relying on a large training set of measured EL images. Three training runs were created with a different randomized split between the training and the validation data. Despite the efforts to create simulated images that closely resemble the measured data, the CNN model is able to predict the parameters on the simulated image significantly better than on the measured image. This demonstrates that the successful use of synthetic data in machine learning applications requires careful reconstruction of the properties of the real data to be expected later. If this simulation reality gap can be successfully controlled and removed, such a physics-informed machine learning approach can potentially be applied to various applied physics and engineering problems.


3. Experimental Validation of an Electro-Thermal Small-Signal Model for Large-Area Perovskite Solar Cells

Author: Ennio Luigi Comi

Master’s Thesis in Energy and Environment, Zurich University Of Applied Sciences - ZHAW Institute Of Computational Physics - ICP

Feb 22, 2021

Small-signal DLIT simulation (left) and measurement (right) showing the temperature amplitude of a perovskite dual cell with two shunts in the upper cell and an interconnection in the center. The simulation allows to better understand the nature of …

Small-signal DLIT simulation (left) and measurement (right) showing the temperature amplitude of a perovskite dual cell with two shunts in the upper cell and an interconnection in the center. The simulation allows to better understand the nature of the defects and to quantify them.

Perovskite thin-film solar cells have attracted a lot of attention in recent years due to rapidly increasing efficiencies. The upscaling of this technology from small laboratory cells to large-area devices without compromising efficiency and stability, however, is still a challenge to be solved for commercialization. Printing perovskite modules entirely by screen printing is an important step towards industrialization, which is why we are carrying out an electro-thermal analysis of screen-printed carbon-based hole transporter-free perovskite dual cells with various interconnection widths. For this purpose, we use the FEM (Finite Element Method) software Laoss that supports the upscaling process from small- to large-area devices by solving for the potential and temperature distribution in 2D top and bottom electrode domains, which are coupled by a vertical 1D coupling law.

We are presenting electrical and thermal DC and AC simulations of dual cells and a reference cell without an interconnection and compare the simulation results with measurements. The software can not only perform electrical and thermal steady-state simulations but also determine the influence of non-ideal electrodes in the frequency domain. Therefore, we also introduce the small-signal dark lock-in thermography (SS-DLIT) method to measure and simulate electro-thermal effects in perovskite solar cells in the dark with high accuracy thanks to the use of a small, periodic voltage modulation at a chosen offset voltage. This adapted DLIT method can be simulated with the thermal AC module in Laoss and allows the investigation and quantification of various defects, such as shunts or the quality of the interconnection of perovskite solar cell modules.


Achievements beside science

Laoss 4.1

The AIPV project has already proven beneficial in the advancement of our simulation software Laoss. As a direct result of the research carried out, we have been able to implement and validate several new features in our simulation software. Thanks to the AIPV project, a major step forward was made in linking Setfos and Laoss simulations as well as with the electro-thermal simulation of small-signal features (AC) for large-area solar cells and modules.


AIPV next stage?

Due to the encouraging results obtained during this project we have submitted a proposal to continue the project and look forward to sharing further results in time.

💻 New Software Release

Laoss 4.1

Laoss (large-area organic semiconductor simulation) is a powerful software package for the design, simulation, and optimization of large-area organic and perovskite solar cells and LEDs (displays, lighting panels, photovoltaic arrays).

  • FEM based electro-thermal modeling

  • Powerful 3D ray tracing for displays and concentrator PV

  • Wide range of different output data and comprehensive result visualization.

  • High-speed computation on standard PCs

  • Intuitive, easy to use Graphical User Interface (GUI) and workflow

The latest release 4.1 see the additional features and improvements:

  • AC Simulation

  • Laoss-Setfos integration

  • Metal finger predefined geometry has additional parameters: number of fingers, angle, base offset

  • New parameters to predefined pixel topography: Number of pixels in X and Y direction

  • Geometry import and predefined geometries can now be used in Laoss Optics as well, including a pixel geometry

  • Cutting triangles step can now be skipped in Laoss Optics

  • Ability to fix azimuth angle phi for (non-azimuthal) BSDFs

  • Check for changed parameters before closing laoss, running a simulation, loading a different simulation etc.

  • Laoss Optics: Possibility of setting individual border type per cardinal direction

  • Laoss Optics: Spectrum plot

  • Undo/Redo in topography import

  • Ability to show interface geometry in result XY plot

  • Project Save As with simulation results

  • Path continuation solver

If you haven't already updated your software please do so now.

 

Would you like to learn how to use simulation to develop large-area PV and LEDs?

📈 The Revolutionary Characterization Platform

Paios is the revolutionary platform for the electro-optical characterization of solar cells and LEDs. How revolutionary? Check out our new video to find out:

 

📺 Scientific Tutorial Videos


Over the years we've recorded hours of video content from various conferences and workshops, some of it has been openly available via our Youtube channel and some of it has been locked away in the vaults of our servers. There are some great insights and knowledge to be gained in this material so we've gathered the most recent and useful recordings and conveniently hosted them on one page. So if you want to learn how to quantify electro-optical cross-talk in white OLEDs or seeing how quickly one can design and simulate an organic solar cell with Setfos then you may want to check out our scientific videos page:

Watch here: https://www.fluxim.com/videos

🤝 We're Hiring - Join the Fluxim Software Team

Are you a C++ or Labview Software Engineer?

Then we have two exciting full-time roles available within our software team. Both roles are based in the city of Winterthur, Switzerland.

Visit our careers page to get the full details https://www.fluxim.com/jobs