In our previous post, we had discussed the challenge of climate change and the mitigation pathways outlined by the IPCC to curb global warming to 1.5°C. In this post, we detail our collaboration with the Laboratory of Advanced Material at Pittsburgh, where we use Bayesian optimization to discover a fabrication process for an ideal type of nanostructured glass for solar panels.
We discuss the article Using Bayesian Optimization to Improve Solar Panel Performance by Developing Antireflective, Superomniphobic Glass, by Sajad Haghanifar, Bolong Cheng, Michael McCourt, and Paul Leu, appearing at the ICML 2019 workshop, Climate Change: How can AI help?. For a discussion on the algorithmic components of this research, please see the complementary article Practical Complications in Applying Bayesian Optimization to Understand Tradeoffs between Substrate Fabrication Processes, appearing at the Real-World Sequential Decision Making workshop.
This project is a continuation of our earlier collaboration effort, in which we adapt Bayesian optimization (BO) to accelerate material science research. In the current project, for the resulting material to be applicable for solar panel covers, we want it to exhibit the following properties:
- High transmission: maximizing the amount of light passing through the glass surface.
- Low haze: minimizing the scattering of light inside of the glass surface.
- High liquid contact angle: water can bounce off the surface and carry away dust and dirt.
In addition, we are also studying properties such as resistance to fogging and reflection and recovery from abrasion. Abrasion recovery directly allows high durability of the glass surface, thus further reduce the maintenance cost of solar panels.
The nanofabrication process is performed in two steps:
- Reactive ion etching (RIE): Creating the sub-wavelength nanostructures on the top surface of the glass.
- Plasma enhanced chemical vapor deposition (PECVD): Creating re-entrant structures and a low energy surface on top of the nanostructures.
The fabrication process produces randomized nanostructures on the top surface of the glass, as shown in Figure 1. While the fabrication procedure is scalable, there are many parameters involved in each of the two steps. Amongst all the parameters of the fabrication process, we have identified 9 parameters that are most impactful to the resulting material.
Figure 1. Cross sectional scanning electron microscopy (SEM) images of fabricated glass with different magnifications.
The entire search process for the fabrication parameters is illustrated in Figure 2. We modified the Bayesian optimization to adapt to the workflow of the experimentalists. In particular, we give the researchers the ability to tag unpromising fabrication parameters as failures without conducting the experiments. We also rephrased the optimization problem as a constrained optimization problem to only search for fabrications that conform to a baseline performance thresholds.
Figure 2. Illustration of experimental fabrication and Bayesian optimization process for nanostructured glass.
After 64 fabrications driven by Bayesian optimization, we produced five fabrication setups that are Pareto-optimal. In Figure 3, we demonstrate one of the key properties of the nanostructured glass: high liquid contact angle.
Figure 3. Droplets of different liquids on (i) normal glass and (ii) our nanostructured glass with high liquid contact angle.
We also measured the total transmission and haze of the nanostructured glass across a wide domain of wavelength that is the key range for most photovoltaic cells. In Figure 4, we see that the nanostructured glass (etched on one side or on both sides) uniformly outperforms normal glass over the entire spectrum: it has higher transmission, lower haze and lower reflectivity.
Figure 4. Transmission (i), haze (ii) and specular reflection (iii) as a function of wavelength for bare, single side, and double side etched glass.
SigOpt’s vision is to empower the world’s experts. This work provided an exciting opportunity for the SigOpt research team to contribute to a project with broader impacts for our entire planet. We are thankful for the opportunity to collaborate with our friends at the University of Pittsburgh and look forward to future opportunities to bring Bayesian optimization and efficient search to more of the scientific community.
Use SigOpt free. Sign up today.