What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to solve the problem of system parameter identification using Physics - Informed Neural Networks (PINNs) in the presence of unknown non - Gaussian noise. Specifically, the paper focuses on how to improve the performance of PINNs when the measurement data is contaminated by non - Gaussian noise with a non - zero mean and an unknown form. ### Background and motivation The traditional PINNs method assumes that the data is noise - free or only slightly contaminated by Gaussian noise. However, in practical applications, the measurement data may be affected by various complex noises, which may be non - Gaussian and have a non - zero mean. In this case, the standard PINNs method will fail because it cannot correctly handle this type of noise. ### Solutions To meet this challenge, the paper proposes a method combined with an Energy - Based Model (EBM), called PINN - EBM. The specific steps are as follows: 1. **PINN model**: - Use PINN to approximate the solution of the partial differential equation (PDE). - Calculate the residual between the PINN prediction and the measurement data as a noise estimate. 2. **EBM model**: - Use EBM to learn the probability density function (PDF) of the noise. - Use the learned noise PDF to calculate the log - likelihood of the measurement data as a data loss term. 3. **Joint training**: - Jointly train PINN and EBM, and optimize the model parameters by minimizing the total loss function \( L_{\text{tot}} \). - The total loss function \( L_{\text{tot}} \) includes the data loss \( L_{\text{data}} \) and the PDE loss \( L_{\text{PDE}} \). ### Formula representation - **Data loss**: \[ L_{\text{data}}=\frac{1}{N_d'}\sum_{i = 1}^{N_d'}(\hat{x}(t_i^d)-y_i^d)^2 \] where \( N_d' \) is the number of data points in the current mini - batch, \( \hat{x}(t_i^d) \) is the PINN prediction, and \( y_i^d \) is the measurement data. - **PDE loss**: \[ L_{\text{PDE}}=\frac{1}{N_c'}\sum_{i = 1}^{N_c'}f(t_i^c)^2 \] where \( N_c' \) is the number of configuration points in the current mini - batch, and \( f(t_i^c)=F(\lambda)\hat{x}(t_i^c) \) is the PDE residual. - **Total loss**: \[ L_{\text{tot}}=L_{\text{data}}+\omega L_{\text{PDE}} \] where \( \omega \) is a weighting factor. ### Experimental results The paper verifies the effectiveness of PINN - EBM through multiple experiments, including simple exponential equations and complex Navier - Stokes equations. The experimental results show that PINN - EBM is significantly superior to the standard PINN and PINN - off (PINN with offset parameters) when dealing with non - Gaussian noise. ### Conclusion The method proposed in the paper successfully solves the problem of PINNs' performance degradation in the presence of unknown non - Gaussian noise. By jointly training PINN and EBM, it can not only identify the PDE parameters more accurately, but also effectively estimate the noise distribution, thus providing new insights into the measurement process. Future work can further explore the possibility of combining PINN - EBM with other PINN improvement methods.

Physics-informed Neural Networks with Unknown Measurement Noise

Recipes for when physics fails: recovering robust learning of physics informed neural networks

Robustness of Physics-Informed Neural Networks to Noise in Sensor Data

Error-Aware B-PINNs: Improving Uncertainty Quantification in Bayesian Physics-Informed Neural Networks

B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data

Robust Regression with Highly Corrupted Data via Physics Informed Neural Networks

Machine Learning Through Physics–Informed Neural Networks: Progress and Challenges

Computation of the Time-Dependent Dirac Equation with Physics-Informed Neural Networks

Randomized Physics-Informed Neural Networks for Bayesian Data Assimilation

Physics-informed Neural Networks for Encoding Dynamics in Real Physical Systems

A principled distance-aware uncertainty quantification approach for enhancing the reliability of physics-informed neural network

Physics-informed attention-based neural network for solving non-linear partial differential equations

Quantifying total uncertainty in physics-informed neural networks for solving forward and inverse stochastic problems

Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What's Next

Modelling parametric uncertainty in PDEs models via Physics-Informed Neural Networks

Conformalized Physics-Informed Neural Networks

Physics-informed neural nets for control of dynamical systems

Estimates on the generalization error of Physics Informed Neural Networks (PINNs) for approximating PDEs

Estimates on the generalization error of Physics Informed Neural Networks (PINNs) for approximating a class of inverse problems for PDEs

Physics-Informed Neural Networks and Extensions

Understanding Physics-Informed Neural Networks: Techniques, Applications, Trends, and Challenges