Modeling of Large-Scale Dynamic Systems using Partially Connected Physics-Informed Neural Networks

Machine learning holds a significant potential for pattern recognition in large datasets. However, access to the necessary data is not always possible, particularly for physical systems. To address this, physics-informed neural networks (PINNs) have emerged as an innovative approach, incorporating physical principles into neural networks to enhance the model quality and data efficiency. Despite this, the performance of PINNs can be challenged by the complexity of the underlying physics. This paper introduces a novel framework aimed at improving both the accuracy and training efficiency of PINNs, particularly for large or complex systems. The proposed method outperforms traditional approaches, as demonstrated through various simulations presented in this paper.