Deep recurrent neural network model improves stable physical system predictions
Data-driven discrete-time deep recurrent neural network-based modeling for dissipative systems
Machine LearningRobotics
Summary
Predicting how physical systems behave over time is important but difficult, especially when you want to keep certain physical properties like stability and energy loss. The authors designed a special type of neural network, called DissipNet, that learns to predict these systems while making sure they behave stably and lose energy as they should. Unlike some other AI methods that guess rules from data but don't guarantee these properties, this new approach uses math to keep them intact. They tested DissipNet on various problems and showed it works better than common neural networks and some physics-based methods.
What this means in practice
- •For control engineers: Build more reliable control systems by modeling physical processes that maintain stability and energy loss properties through data-driven neural networks.
- •For robotics developers: Develop robotic systems that predict and manage dynamic interactions with the environment while preserving stable behaviors using dissipative recurrent neural networks.
Authors
Tuan Luong, Hyungpil Moon
Abstract
Physical AI has gained increasing attention for its role in developing AI systems that better understand, predict, and control real-world dynamics. Achieving this requires AI models that not only achieve high prediction accuracy but also preserve fundamental physical properties of dynamical systems. In this paper, we propose a deep discrete-time dissipative recurrent neural network (DissipNet) that explicitly enforces dissipativity, a key property related to stability and energy dissipation, through structural weight constraints and a dedicated training algorithm. By construction, the proposed network is capable of learning dissipative dynamics while preserving their inherent stability, which is formally analyzed using Lyapunov theory. In contrast to Physics-Informed Neural Networks (PINNs), which incorporate governing equations into the training loss but do not guarantee preservation of internal analytical properties such as dissipativity or passivity, our approach provides explicit guarantees on stability at the model level. We demonstrate the effectiveness of the proposed method through several modeling applications, and compare its performance with a naive recurrent neural network (RNN) and a PINN-based model.