naPINN: Noise-Adaptive Physics-Informed Neural Networks for Recovering Physics from Corrupted Measurement
Researchers have developed a new method called naPINN, designed to improve the accuracy of Physics-Informed Neural Networks (PINNs) when dealing with corrupted measurement data. This novel approach embeds an energy-based model to learn residual distributions, enabling adaptive filtering of unreliable data points. naPINN demonstrates superior performance over existing robust PINN methods in reconstructing physical dynamics from data with non-Gaussian noise and outliers. AI
IMPACT Enhances the robustness of AI models in scientific discovery from noisy real-world data.