Shocker Open Access Repository
Featured Items
Communities in SOAR
Select a community to browse its collections.
Recent Submissions
Item Metadata only 2026-04-21 University Staff Senate meeting(Wichita State University, 2026-04-21)Item Open Access 2026-03-24 University Staff Senate Meeting(Wichita State University, 2026-03-24)Item Open Access 2026-02-17 University Staff Senate meeting(Wichita State University, 2026-02-17)Item Open Access Computational precision control of electrospun polymer nanofiber morphology using machine learning and genetic algorithm optimization(Springer, 2026-04-29)Abstract: Electrospinning offers exceptional flexibility in fabricating nanofibers, yet achieving precise control over fiber diameter, which is crucial for tailoring material properties, remains a challenge due to the complex and nonlinear interactions among various process parameters. This study presents a literature-derived, multi-material predictive-optimization strategy for the computational control of electrospun fiber diameter. A dataset containing 3000 data points was collected from published scientific research articles on electrospinning, covering a wide range of polymers, solvents, and process parameters, and was used to develop predictive machine learning (ML) models for fiber diameter prediction. Among the developed models, the extreme gradient boosting (XGB) model performed best, achieving a coefficient of determination (R2) value of 0.94 and delivering low prediction errors (root mean square error [RMSE]: 275.02 nm, mean absolute error [MAE]: 75.40 nm) on the unseen data. Experimental validation using polystyrene (PS) and polyvinyl chloride (PVC) nanofibers showed close alignment between predicted and measured fiber diameters, supporting the practical reliability of the predictive model under real electrospinning conditions. To move beyond forward prediction, the trained XGB model was integrated with a genetic algorithm (GA) to enable surrogate-based inverse design of electrospinning parameters for user-defined target fiber diameters. The GA identified parameter combinations corresponding to target diameters between 100 nm and 4000 nm with low surrogate-model fitness errors. For target diameters of 400 nm and 600 nm, the GA achieved fitness errors of 0.03 nm and 0.35 nm, respectively. Furthermore, the integrated XGB-GA approach yielded a perfect linear correlation (R2 = 1.00) between the target fiber diameters and the predicted fiber diameters within the GA-optimized subset, demonstrating the effectiveness of the optimization strategy for precise diameter control across various electrospinning conditions. Overall, this study provides a data-driven approach for predicting fiber diameter and guiding electrospinning parameter selection while reducing dependence on trial-and-error experimentation within the material systems represented in the dataset. © The Author(s) 2026.Item Metadata only Impact of the TensorFlow library on neural network performance for harnessing hardware parallelism(IEEE Computer Society, 2026-04-07)Neural networks constitute a well-established and extensively studied area of modern computational technology. Among available frameworks, TensorFlow remains one of the most prevalent platforms for implementing and experimenting with neural network architectures. TensorFlow abstracts the process of constructing and training neural networks, and supports several parallel computing interfaces, including Compute Unified Device Architecture (CUDA) and Direct Machine Learning (DirectML). However, the structure and training process may not be able to take full advantage of the underlying hardware due to how TensorFlow parallelizes computations. In this paper, we investigate the impact of the TensorFlow library on the performance of a sample Convolutional Neural Network (CNN) by varying two training parameters: batch size and the number of epochs. We use CNN to classify images from the Modified National Institute of Standards and Technology (MNIST) database of handwritten digits. We run the CNN on a desktop with an eight-core central processing unit (CPU) plus 4608-core graphics processing unit (GPU) and a high-performance computing (HPC) cluster with 2x18 CPU cores plus 5120 GPU cores. According to the experimental results, a larger batch size may reduce training time with negligible losses in accuracy for certain CNN parameters. Results suggest that optimization should be taken on a case-by-case basis to determine the best parameters. © 2026 IEEE.