Breakthroughs in Solving Partial Differential Equations

The field of partial differential equations (PDEs) is experiencing a significant shift with the integration of deep learning techniques, enabling the discovery of hidden PDE models and the advancement of physics-informed neural networks. Researchers are exploring innovative methods to solve high-dimensional PDEs, such as the use of backward stochastic differential equations and the development of novel integration schemes. Transfer learning and physics-preserved methods are being proposed to improve the generalizability and accuracy of PDE solvers. Furthermore, the application of PDEs to real-world problems, including traffic dynamics and population forecasting, is demonstrating the potential of these advances to drive meaningful impact. Noteworthy papers include:

  • A novel deep learning framework, designed to discover hidden PDE models of traffic network dynamics, which has demonstrated effectiveness in predicting traffic evolution.
  • A Stratonovich-based BSDE formulation that eliminates bias issues in existing BSDE-based solvers, achieving competitive results with PINNs.
  • A physics-preserved transfer learning method that adaptively corrects domain shift and preserves physical information, showing superior performance and generalizability in solving differential equations.

Sources

Neural Networks Enabled Discovery On the Higher-Order Nonlinear Partial Differential Equation of Traffic Dynamics

Integration Matters for Learning PDEs with Backwards SDEs

A Physics-preserved Transfer Learning Method for Differential Equations

Learning and Transferring Physical Models through Derivatives

Advances in Particle Flow Filters with Taylor Expansion Series

Dynamical Update Maps for Particle Flow with Differential Algebra

An LSTM-PINN Hybrid Method to the specific problem of population forecasting

Predicting the Dynamics of Complex System via Multiscale Diffusion Autoencoder

Physics-Informed DeepONets for drift-diffusion on metric graphs: simulation and parameter identification

Built with on top of