On fast simulation of dynamical system with neural vector enhanced numerical solver.

Area of Science:

• Computational Science
• Applied Mathematics
• Machine Learning

Background:

• Large-scale simulations of dynamical systems are crucial across science and engineering.
• Traditional numerical solvers face accuracy-computational efficiency trade-offs due to fixed step sizes.
• Existing methods struggle with error accumulation and require small time steps.

Purpose of the Study:

• To introduce Neural Vector (NeurVec), a deep learning-based corrector for dynamical system simulations.
• To enable larger time step sizes and improve computational efficiency without sacrificing accuracy.
• To demonstrate NeurVec's generalization capabilities and potential to establish a new paradigm in solving differential equations.

Main Methods:

• Development of Neural Vector (NeurVec), a deep learning model designed to correct integration errors.
• Training NeurVec using limited and discrete data from various complex dynamical system benchmarks.
• Extensive experimental validation on diverse simulation scenarios to assess performance.

Main Results:

• NeurVec demonstrated remarkable generalization across a continuous phase space, even with sparse training data.
• The deep learning corrector significantly accelerated traditional solvers, achieving speedups of 10x–100x.
• High levels of accuracy and stability were maintained despite the increased time step sizes.

Conclusions:

• NeurVec effectively compensates for integration errors, enabling faster and more accurate simulations of dynamical systems.
• The model's ability to generalize and its ease of implementation suggest a paradigm shift towards deep learning for differential equation solving.
• NeurVec offers a promising approach to overcome the limitations of traditional numerical solvers in scientific and engineering applications.