Associate Professor Florida International University, United States
Introduction: : Increased attention is being drawn to the electrospinning technique for its ability to generate fibers from the nanometer to micrometer range; however, controlled fiber formation remains a challenge. The fibers created in the electrospinning process have a diverse range of applications, one of which is a cardiac patch, which aims to regenerate necrotic heart tissue. Cardiac patch design iterations remain in the development stage due to the inability to create such electrospun fibers that emulate cardiac structure; the unpredictability of the electrospinning technique is to blame for this shortcoming. This study addresses this challenge by designing a novel, end-to-end software to accurately predict the fiber formation and subsequent properties of electrospun fibers.
Materials and
Methods: : The platform utilizes various machine learning models to accurately predict the average diameter, average elastic modulus, and distribution of diameters, as well as generate an image prediction of the fibers viewed under a 500x optical magnification, given the volume of polymer solution pumped through the needle tip per unit time (flow rate) and the tip-to-collector distance—a massive improvement over current solutions that can only predict the average diameter. This platform was developed in two parts, the first to address the regression tasks of this system—accurately predict the average diameter, average flow rate, and distance of diameters. For the prediction of the average diameter and flow rate, linear regression, sinusoidal regression, support vector regression, and polynomial regression were created and tested. Three additional models were considered in the prediction of the distribution of fiber diameters: the random forest, the feed-forward neural network, and a custom transformer model.
Results, Conclusions, and Discussions:: In predicting the average diameter, the sinusoidal regression model largely outperformed the linear regression, support vector regression, and polynomial regression models. The best polynomial regression model outperforms all current solutions with its incredibly high R² score. Predicting the average elastic modulus yielded different results with the polynomial regression model barely outperforming the sinusoidal regression model but outclassing the linear regression and support vector regression models by a considerable margin. The Random Forest model was the premier model in predicting the distribution of fiber diameters. This was followed by the linear regression model, the polynomial regression model, the sinusoidal regression model, the custom-built transformer model, and finally, the feed-forward neural network. The culmination of this research arose in the second step of this pipeline, where a novel machine learning architecture was created; this model used the inputs generated by the regression models to generate an image prediction of the fibers viewed under a 500x optical magnification. The model architecture outperformed traditional industry solutions like a Generative Adversarial Network. The R², SSIM, and MSE scores of the best models indicate statistical significance and accuracy in predicting an otherwise unpredictable electrospinner.
Acknowledgements and/or References (Optional): : I would like to acknowledge Florida International University and the Prasad Lab for Materials Research.
