Our research team at SigOpt has been very fortunate to be able to collaborate with outstanding researchers around the world, including through our internship program. In our Highlight blogs, we take the opportunity to feature our work with these collaborators. This post introduces a collaboration on Bayesian Optimization of High Transparency, Low Haze, and High Oil Contact Angle Rigid and Flexible Optoelectronic Substrates by Sajad Haghanifar, Sooraj Sharma, Luke M. Tomasovic, and Paul. W. Leu, Bolong (Harvey) Cheng, appearing at the upcoming NeurIPS 2018 Machine Learning for Molecules and Materials.
This project is a collaboration between SigOpt and the Laboratory for Advanced Materials at Pittsburgh (LAMP), directed by Professor Paul Leu. LAMP uses computational modeling and experimental research to design and understand advanced materials in fields such as nanostructures, solar power and medicine. Here at SigOpt, we are thrilled to support their groundbreaking accomplishments through our academic program.
Many modern consumer electronic devices such as smartphones and tablets require the use of glass or plastic materials to protect the device’s delicate display. In order to do so, the glass/plastic material must exhibit important characteristics specific to these applications.
First, the material should have high transparency to allow light to pass through. Even small increases in transparency can reduce the brightness demands of the display; this can prolong the battery life of these mobile devices.
Second, the material should minimize haze. Haze measures how much light is scattered; too much haze will cause fine details like text to appear blurry. It is interesting to note that high haze is desirable for applications like solar cell and LEDs, as seen in the previous work by the LAMP group maximized haze (Fig. 1). This new work focuses on minimizing haze for display applications.
Fig 1. Optical images of smooth glass and glass nanograss when placed (left) directly on paper with text and (right) about 1 cm above (source: Haghanifar et. al., 2017).
Lastly, the material should be resistant to dust, grease, water, etc. This last characteristic has been linked to the property of superomniphobicity, which affects a material’s ability to repel liquids. This is especially interesting because of new developments in nanostructured surfaces inspired by similarly occurring nanostructures found in nature (such springtail insects shown in Fig 2). By adding little hair-like spikes to the surface of the material, we can adjust the liquid contact angles on the surface (see Fig 3).
Fig 2. Springtail insect with omniphobic skin, mushroom like structure (source: Choi et. al., 2017)
Fig 3. Nanostructured grass-like glass with different heights (source: Haghanifar et. al., 2017).
Currently, nanostructured surface research is slow and fragmented due to the use of trial-and-error design methods; these often rely on the knowledge and intuition of researchers. Nanostructured surface experiments require the precise selection of various fabrication parameters, such as the flow rate of various gases, ion etching time, chamber pressure, etc. Moreover, the fabrication process is time consuming: one step in such an experiment requires 16 hours of chemical vapor deposition.
So how do we efficiently search for the desired fabrication parameters? The answer is Bayesian optimization. In our paper, we use Bayesian optimization in a two-stage approach to efficiently design the optimal material and fabrication process for displays. During the first stage, we maximize transparency and minimize haze in a multiobjective setting. This produces a material that is crystal clear for us to optimize in the next stage. In the second stage, we optimize the liquid contact angle of the materials, as a proxy objective for the superomniphobicity characteristic. This sequence of two Bayesian optimization experiments yields a material which effectively maximizing both transparency and superomniphobicity while requiring as few experiments as possible.
Fig 4. Oil drop shape and contact angles on bare glass (left) and superomniphobic glass (right).
To learn more about our this fabrication process, and the valuable role Bayesian optimization played in efficiently designing it, please visit our poster at the NeurIPS 2018 Workshop on Machine Learning for Molecules and Materials on December 8, 2018.
Special thanks to Taylor Jackle Spriggs for his contributions to and editing of this blog post.