Reproduce “Evolutionary Optimization of Physics-Informed Neural Networks: Advancing Generalizability by the Baldwin Effect” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jian Cheng Wong, Chin Chun Ooi, Abhishek Gupta, Pao-Hsiung Chiu, Joshua Shao Zheng Low, My Ha Dao, Yew-Soon Ong
A Baldwinian neuroevolution framework imbues Physics-Informed Neural Networks (PINNs) with enhanced generalizability and computational efficiency, enabling rapid and accurate solutions across diverse linear and nonlinear physics problems. This approach yields approximately an order of magnitude lower relative error and up to two orders of magnitude faster computation for new tasks compared to existing meta-learning PINNs, without requiring explicit task parameterization.
Key insights
- Evolving the 'genetic makeup' (hidden layer weight distributions and regularization) of PINNs enables the models to be inherently biased towards rapid and accurate physics learning for diverse tasks.
- The closed-form Tikhonov regularization for the final layer weights allows for extremely fast (milliseconds) and robust lifetime learning for new tasks, even for nonlinear PDEs via an iterative approach.
- The framework effectively generalizes across broad, unparameterized task distributions (e.g., varying PDE parameters, initial conditions, boundary conditions) without relying on explicit interpolation between tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.