Prediction of air filtration efficiency and airflow resistance of air filter media using convolutional neural networks and synthetic data derived from simulated media.

In the absence of experimental air filtration test data, it is difficult to predict the air filtration efficiency and airflow resistance of filter media. In this novel work, an effort is undertaken to predict the air filtration efficiency and airflow resistance of filter media based solely on scanning electron microscopy (SEM) images of the filter fabric. The analytical single fiber efficiency (SFE) model is incorporated into a machine-learning model designed to analyze digital imagery. A convolutional neural network (CNN) is developed and optimized using synthetic training data produced from digital replication of air filter media. Simulated air filter media were created to replicate the physical characteristics of actual nonwoven nanofibrous air filter media. Digital grayscale images of the simulated media provided the input data for the CNN. Analytical calculations of efficiency and resistance based on the SFE model provided target data for the CNN. The regression model included seven convolution layers and two hidden fully connected layers along with width reduction and depth expansion methods. Twelve model parameters were optimized using training and validation data of 2100 iterations of simulated media. The model was then employed to predict the air filtration behavior of actual air filter media based on an SEM image of the media. The resulting predictions of air filtration efficiency and airflow resistance based solely on an SEM image suggests the potential viability of using synthetic data derived from simulated media to train machine learning models for real-world applications. To the knowledge of the authors, this is the first such application of a CNN to predict air filtration efficiency and airflow resistance using an SEM image, synthetic data created from simulated media, and the SFE analytical model. • Convolutional neural networks enable prediction of air filtration efficiency. • Analytical single fiber efficiency models are applied to simulated media. • Computer vision with CNNs enables prediction of airflow pressure resistance. • Machine learning modeling with CNNs will advance the science of air filtration. [ABSTRACT FROM AUTHOR]