Data-Driven Model for Prediction of Radiative Characteristics
Miguel Moreno
Author
04/05/2021
Added
17
Plays
Description
Advancement of surface laser processing for thermal management applications via a combined deep learning and machine learning model that predicts surface patterns and thermal emissivity.
Searchable Transcript
Toggle between list and paragraph view.
- [00:00:00.700]Hello everyone. My name is Miguel Moreno
- [00:00:02.290]and this is Data-Driven Model for Prediction
- [00:00:04.610]of Radiative Characteristics in collaboration
- [00:00:06.241]with Andrew Reicks and Dr. Mohammad Ghashami, as my advisor.
- [00:00:11.470]I need to start
- [00:00:12.303]by first defining important background information.
- [00:00:14.770]Like what emissivity is.
- [00:00:16.560]Which in this case, is the potential
- [00:00:18.150]of a surface to emit thermal energy in form
- [00:00:20.640]of radiation, is given as a number from zero to one
- [00:00:23.690]because it's a ratio to the black body radiation
- [00:00:27.490]which is the theoretical limit.
- [00:00:30.160]It's wavelength, angle and temperature dependent.
- [00:00:33.610]There is a desire
- [00:00:34.443]for materials with high terminal emissivity in industries
- [00:00:37.370]like aerospace, photovoltaics or basically solar panels
- [00:00:41.380]and also radiative cooling which is a promising new field.
- [00:00:45.320]For this research,
- [00:00:46.220]We want to focus on laser processed surfaces.
- [00:00:49.430]So what is surface processing?
- [00:00:51.786]It's goal is to optimize emissivity with surface patterns,
- [00:00:56.070]but why does this improve emissivity?
- [00:00:59.228]Well, microstructures increase surface area
- [00:01:02.250]and act as "light traps"
- [00:01:04.630]meaning they are locations where light concentrates
- [00:01:09.640]and facilitates outgoing radiation.
- [00:01:12.650]We can see this in nature, like in the case of silver ants
- [00:01:16.320]whose hair act as tiny radiators.
- [00:01:20.610]Like we can see here on the image on the right.
- [00:01:24.730]We can also see in leaves and even in some butterflies.
- [00:01:30.110]Some butterflies have micro structures in the wing
- [00:01:33.320]that not only reflect the color,
- [00:01:36.870]for example, vivid blue
- [00:01:39.460]but also help them to regulate their temperature.
- [00:01:42.670]The advantages of doing surface processing
- [00:01:45.550]with laser specifically, is that it is low cost, scalable,
- [00:01:50.270]easy to modify parameters, and it has no effect
- [00:01:53.820]on material's bulk properties.
- [00:01:56.280]Which means, for example, if we were to do this on a metal
- [00:02:00.140]it would not affect its strength.
- [00:02:02.493]Now focusing on our search research.
- [00:02:05.920]Our objective was to predict emissivity
- [00:02:08.230]of surface profiles to improve fabrication and
- [00:02:11.450]development performance of laser process surface techniques.
- [00:02:15.900]We can divide our research into three main parts.
- [00:02:19.640]The first one being fabrication
- [00:02:21.220]of samples with different surface patterns.
- [00:02:23.520]Then we move on to the surface characterization
- [00:02:26.790]or analysis of the surface.
- [00:02:29.160]Now finally, the combined deep learning
- [00:02:31.030]plus machine learning model.
- [00:02:33.800]The fabrication part carried out
- [00:02:35.440]by Andrew Reicks consists of 116 laser processed samples
- [00:02:41.200]of polished aluminum alloy 6061
- [00:02:44.180]using a femtosecond laser here
- [00:02:46.010]at NERcF at UNL with changing parameters,
- [00:02:49.910]such as peak fluence and pulse count
- [00:02:52.080]to create different patterns in every sample.
- [00:02:54.420]We can see here on the image in the right.
- [00:02:57.890]An aluminum sample, with six different surface patterns
- [00:03:03.630]which are the black squares.
- [00:03:06.370]They don't have to be exactly one centimeter
- [00:03:10.550]by one centimeter
- [00:03:12.050]because it is not important for this case.
- [00:03:16.870]The size at least.
- [00:03:18.830]It was an open air fabrication
- [00:03:20.240]which means there was no chamber involved.
- [00:03:23.400]So the sample was in contact with the ambient air.
- [00:03:27.260]We now move on to the surface characterization.
- [00:03:30.420]We started by using a thermal imaging camera
- [00:03:32.980]to measure the spectral directional emissivity.
- [00:03:36.550]Which was measured at multiple angles
- [00:03:38.560]from zero to 85, in five degrees increments.
- [00:03:42.690]We then integrated these emissivity over all emission angles
- [00:03:46.330]at a fixed wavelength and temperature,
- [00:03:49.070]which in this case was 50 degrees.
- [00:03:51.590]And then we obtained a spectral hemispherical
- [00:03:54.010]emissivity value, which in the case
- [00:03:56.040]of this surface is shown here as Eh
- [00:03:59.610]which is equal to 0.956.
- [00:04:02.540]Which is already way above the emissivity
- [00:04:05.550]of normal aluminum.
- [00:04:08.160]For the surface imaging,
- [00:04:09.340]we use the Keyence Laser Scanning Microscope,
- [00:04:12.540]also found here at NERcF.
- [00:04:16.035]I then took three images per sample
- [00:04:18.180]each from a different section to account
- [00:04:19.830]for difference in pattern.
- [00:04:22.020]And then I applied a height filter to
- [00:04:25.012]facilitate further analysis,
- [00:04:26.851]which is where we can see here on these images.
- [00:04:29.670]The most orange means higher microstructure
- [00:04:34.210]and the blue means shallow places.
- [00:04:38.742]And we can already see how different the patterns
- [00:04:41.401]can be ranging from a maximum height of 187.5 micrometers
- [00:04:45.270]to all the way down to 17.9 micrometers.
- [00:04:49.250]It is worth noting that they are all at 50X resolution
- [00:04:53.230]which means we can appreciate the real difference
- [00:04:57.320]in the patterns.
- [00:05:01.170]I then used the VK-analyzer software
- [00:05:03.585]which is from Keyence to obtain surface roughness parameters
- [00:05:08.114]like the maximum height that I reported in the last slide.
- [00:05:11.550]And after that, we take these processed images.
- [00:05:16.400]So the images with the height filters, and we take them
- [00:05:22.196]as input to a MATLAB model, which calculates the density
- [00:05:27.021]of micro structures using a Fast Fourier Transform.
- [00:05:29.260]This was carried out by Andrew Reicks.
- [00:05:33.330]But basically how it works is that,
- [00:05:36.702]it splits the image into three different layers
- [00:05:41.477]and from each layer, it calculates the FFT.
- [00:05:44.020]And after that we get three different density FFT values,
- [00:05:48.020]which they are usually different,
- [00:05:50.597]but at the end we take the one that makes more sense
- [00:05:55.165]and we average them and we obtain the general density.
- [00:05:59.870]The final part of this was the Data-Driven Model itself.
- [00:06:06.460]So all these surface roughness parameters, all the images
- [00:06:10.474]with the height filters, emissivity values
- [00:06:13.604]are inputs for this model.
- [00:06:16.420]And then we can obtain the prediction of patterns
- [00:06:19.826]for a given emissivity.
- [00:06:22.064]This graph shows in blue dots the experimental emissivity
- [00:06:26.723]obtained from a given, from given patterns
- [00:06:30.270]and the orange dots are predicted emissivities
- [00:06:33.960]using this model.
- [00:06:38.540]We can see how good this model is by looking at this graph,
- [00:06:43.484]the percent error .
- [00:06:44.610]We can see that almost all of it is below 8%
- [00:06:50.782]with some outliers above 8%,
- [00:06:55.857]but we believe
- [00:06:56.961]or we know for sure that this model can be improved
- [00:07:01.643]by increasing the number of samples.
- [00:07:04.876]I then show two images of the same pattern
- [00:07:08.410]but one is taken directly from the Keyence
- [00:07:11.500]and the other one is a surface built by the model
- [00:07:13.840]to obtain a given emissivity.
- [00:07:17.279]In conclusion, this model demonstrates the power
- [00:07:20.776]of combined deep learning plus machine learning
- [00:07:23.457]for novel applications like radiative cooling.
- [00:07:27.259]It helps in the advancement
- [00:07:28.968]of laser processing knowledge within the field.
- [00:07:31.632]And it opens a door
- [00:07:32.465]for novel future applications, like radiative cooling
- [00:07:35.878]which is going to be a our future focus.
- [00:07:39.290]Thank you so much for your attention and have a great day.
The screen size you are trying to search captions on is too small!
You can always jump over to MediaHub and check it out there.
Log in to post comments
Embed
Copy the following code into your page
HTML
<!-- To force a 16x9 aspect ratio use 'padding-top: 56.25%;' instead of 'padding-top: 75%;' --> <div style="padding-top: 75%; overflow: hidden; position:relative; -webkit-box-flex: 1; flex-grow: 1;"> <iframe style="bottom: 0; left: 0; position: absolute; right: 0; top: 0; border: 0; height: 100%; width: 100%;" src="https://mediahub.unl.edu/media/16424?format=iframe&autoplay=0" title="Video Player: Data-Driven Model for Prediction of Radiative Characteristics" allowfullscreen ></iframe> </div>
Comments
0 Comments