Sprays have a wide variety of applications including internal combustion engines, agricultural sprays, pharmaceutical powder production through spray drying process, aerosol generation for drug delivery, fire suppression, gas turbines, and many more [1]. The spray characterizations are needed in order to properly design systems that use sprays. Spray characterizations include spray droplet size, velocity, and number density distributions. Currently, spray characterizations are obtained outside of the actual industrial use of the spray and at a research stage of the system design. The same information is then used for the design of the actual system for the actual operating conditions. Since the operating conditions influences the spray characterizations, the design parameters have to be adjusted to optimize the system performance. This is a trial and error process with significant cost and time delays.
The objective of the present study is to develop an AI based method to characterize spray in the actual operating conditions. This can be achieved if the near nozzle information can be related to the spray characterization. Since it is much easier to determine the near nozzle information in most actual systems, the spray characterization in actual operating conditions can be determined in-situ and at low cost.
Our approach to this problem is through implementation of deep neural network artificial intelligence (AI) to produce statistically valid and accurate method of spray characterization with great predictive power. There are many examples of such scientific applications of deep neural networks, such as the work done by Ma, et al. [2], where deep neural networks were utilized for predicting the activity of potential drug molecules; the works of Wei, et al. [3] in predicting organic chemistry reactions using deep learning; and the most relevant to the work done in this study, the introduction of convolutional neural networks for image recognition pioneered by Krizhevsky, et al. [4], which classified the ImageNet database images with astonishing accuracy.
The present study examines the effectiveness of deep convolutional neural networks (CNN) as models for spray characterization. The most important characteristic of a spray is the droplet size distribution. The droplet size distribution is obtained by a variety of methods. Most commonly used methods are either based on direct imaging of droplets and further image processing to determine the size of the droplets, or light scattering methods, such as the phase Doppler anemometry system that operates based on the Doppler effect in Mie scattering signals coming from the droplets. In a review done by Lee, et al. [5], it is mentioned that although the phase-Doppler method can be an accurate method of droplet sizing, because of difficulty in optical alignment and expensive equipment cost, it is mainly used by research laboratories and not in actual operating conditions.
Some of the limitations in the aforementioned methods can be mitigated and minimized by introducing deep learning AI. Specifically, by moving away from particle size measurements towards relating other aspects of the atomization process to the particle size distribution. This provides a pathway to move from early stages of atomization to characterizing the spray. Correspondingly, the need for the difficult optical setups and image processing will be eliminated. This stems from the fact that the particle measurement process will only be present once during the development of the database for training the deep neural network. In this study, images from the early stages of spray formation are used to distinguish between different operating conditions without any insight from the spray itself, the size distribution, or the physical parameters of the experiment. This is done using CNNs trained on the aforementioned images to distinguish between the different operating conditions. Here, we study the capability of CNNs in identifying and using physically relevant patterns in the spray formation process to provide useful information with respect to the spray.
We have developed convolutional neural networks (CNN) that are capable of identifying operating pressures of a spray without looking at the droplets formed by the spray, but rather by looking at the spray formation process. For these models a ss4001 nozzle was used at different operating conditions. In our main models that are much larger and not practical for demo we have been able to get upto 94% accuracy in determining operating pressures in a +/- 25psi range. The purpose here is to demonstarte these models, however for practicallity lighter less accurate versions of the models were developed for deployment in this demo. The accuracy of the models presentd here are 80% for both models.
We have made lighter versions of our model to be deployed online (for this version we have simplified the model to reduce its size and the required computing power to run it) to showcase the our work at it's early stages and to update the website as we develop more models and further improve our work, so stay tuned.
In this demo you will find two different models. One of these models uses closeup images of a spray before droplets are formed to determine the operating pressure of the spray. The other model uses images of the spray from a distance, again before droplet formation, to predict the operating pressure. The details of each of these models can be found below.
For this model grayscale images of the spray from an SS4001 nozzle were used at 50psi, 70psi, and 90psi operating conditions. The images were taken from the spray at the very early stages so as to highligh the atomization process and the potential in connecting this process to the spray characteristics. These images were taken very close to the spray to capture the smallest details of the spray formation process. The full model is able to identify between the three pressures with 88% accuracy, however the lighter model used in the demo is only 80% accurate.
For this model grayscale images of the spray from an SS4001 nozzle were used at 75psi, 100psi, 150psi, 200psi, 250psi, 300psi, 350psi, and 400psi operating conditions. The images were taken from the spray at the very early stages so as to highligh the atomization process and the potential in connecting this process to the spray characteristics. The images for this model were take further away from the spray to focus on the larger scale patterns of the spray. The full model is able to identify between the 8 pressures with 94% accuracy, however the lighter model used in the demo is only 80% accurate.
In this demo we have also included the grad-CAM analysis of the models. This method was developed by Selvaraju, et al. [6], and is used to understand the CNN's behaviour. The purpose of including this is to allow the users to take a look at what the CNN is focusing on to make the decisions it is making. Although it is important to mention that the simple models in this demo are slightly overfitted and show some of the patterns that are not physically relavent.
To use the demo simply follow the steps from the start to the finish, it is very intuitive and easy to use.
Some important notes on using the demo:
&tab; All images will be converted to grayscale before being inputed to the CNN.
&tab; Any image you upload to try out the model must be from an SS4001 nozzle either closeup or far images taken before droplet formation to get any meaningful output from the demo.
&tab; We recommend you check out the sample images to try things out, as the model is in preliminary stages and probably will not give you meaningful results on your own images at this point.
Download a more detailed description of the project: Download
[1] N. Ashgriz, Handbook of atomization and sprays, New York: Springer, 2011.
[2] J. Ma, R. P. Sheridan, A. Liaw, G. E. Dahl and V. Svetnik, “Deep Neural Nets as a Method for Quantitative
Structure−Activity,” Journal of Chemical Information and Modeling, vol. 55, pp. 263-274, 2015.
[3] J. N. Wei, D. Duvenaud and A. Aspuru-Guzik, “Neural Networks for the Prediction of Organic Chemistry Reactions,”
ACS Central Science, vol. 2, pp. 725-732, 2016.
[4] A. Krizhevsky, I. Sutskever and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,”
Advances in neural information processing systems, pp. 1097-1105, 2012.
[5] S. Y. Lee and Y. D. Kim, “Sizing of Spray Particles Using Image Processing Technique,” KSME International
Journal, vol. 18, pp. 879-894, 2004.
[6] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh and D. Batra, “Grad-CAM: Visual Explanations from
Deep Networks via Gradient-based Localization,” International Journal of Computer Vision, vol. 128, p. 336–359,
2019.