Hybrid AI method speeds up multibody simulation

By AI, Created 11:41 AM UTC, June 03, 2026, /AGP/ – Researchers at Dalian University of Technology developed a physics-informed neural network that combines mechanical equations with training data to improve simulation of moving systems. The approach is designed to make predictions faster and more stable for robots, vehicles and deployable aerospace structures without losing key physical constraints.

Why it matters: - Fast simulation is critical for machines that move under constraints, including robots, vehicles and deployable aerospace structures. - Traditional solvers can be accurate but slow. - Pure neural networks can be faster, but they may drift from real mechanical behavior when systems become highly nonlinear. - The new hybrid approach aims to close that gap by keeping predictions tied to the governing physics.

What happened: - A research team from Dalian University of Technology published a study in Acta Mechanica Sinica on a physics-informed neural network method for nonlinear multibody systems. - The work combines differential-algebraic equations with neural-network training to improve prediction accuracy, robustness and generalization as system parameters change. - The study focused on multibody systems such as connected components with joints, rods, sliders or panels.

The details: - The method trains a neural network to match both data and the system’s mechanical equations. - The model uses scaling coefficients in the loss function so the dynamics guide learning. - The framework splits input into a larger fitting dataset and a smaller mechanism-informed dataset that corrects the model against physical laws. - Automatic differentiation is used to derive velocity and acceleration from predicted positions. - The team tested the approach on a simple pendulum, a slider-crank mechanism and a satellite panel system. - The method was compared with an artificial neural network, tested with noisy data and evaluated across different driving velocities, external forces, time steps, hidden layers and neuron counts. - The reported constraint violations were generally kept between 10−2 and 10−4. - The results showed better constraint preservation, stronger stability and faster prediction than the baseline network. - The study’s DOI is 10.1007/s10409-024-24159-x.

Between the lines: - The research reflects a broader shift toward physics-guided AI, where engineering models are not treated as black boxes. - That matters most for nonlinear systems, where small errors can quickly turn into large violations of the governing equations. - The hybrid setup is meant to give neural networks a stronger grasp of real motion constraints without giving up speed. - Funding came from the National Natural Science Foundation of China and the China Postdoctoral Science Foundation.

What’s next: - The authors said the method could support future real-time analysis and design for robotic systems, vehicle mechanisms and aerospace deployable structures. - The framework may also apply to other scientific computing tasks that need both data efficiency and physical interpretability. - Future work may extend the strategy to larger, more flexible and more strongly coupled systems. - The broader goal is to move AI-assisted mechanical simulation closer to practical engineering use.