Excellence in Research: Experiment Efficient Modeling Method of Dynamic Systems Based on Short-Term Dependency and Non-Recurrent Neural Networks

点击翻译按钮获取中文摘要

This grant will support research to contribute new knowledge related to dynamic systems, promoting a modeling method that is experiment efficient and accelerating the research and development processes. Current modeling methods either require comprehensive understanding, which is difficult for complex systems, or extensive experiment effort, which is impractical due to the expense and time. However, a limited number of experiments does not always mean a limited amount of data. With advanced sensing techniques, abundant in-situ data can be collected in every experiment. This award supports fundamental research to provide needed knowledge to extract abundant independent data from the in-situ data in a limited number of experiments. The new knowledge will provide enough data to train neural network models even with a limited number of experiments and help reduce the time and cost to model dynamic systems. As dynamic systems are common in aerospace, manufacturing, material science, and civil systems, the results from this research will benefit the U.S. economy and society. This grant will support women students and broaden the participation of women in science, technology, engineering, and math. A micro-scale supporting network for women engineering students will be built within the PI’s group, which will connect the PI, graduate students, undergraduate students, and K-6 children on a regular basis to encourage women to pursue an engineering career. Practitioners find that short-term memory feed-forward neural networks and infinite memory recursive neural networks have comparable performance in some dynamic systems. This project is to study this phenomenon based on the well-known observability property of dynamic systems, and will focus on two specific case studies, fiber orientation in additive manufacturing and nanotube network quality of continuous nanotube thin film. Different from the existing observability criteria that rely on the full system knowledge, the observability criteria built in this project only depend on the in-situ data and/or the partial system knowledge. Based on the short-term dependency study, abundant independent data will be extracted from a limited number of experiments to train feed-forward neural network models. Based on the partial knowledge of the dynamic systems, this project is for a customized feed-forward neural network structure to achieve further data efficiency.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.