1: \begin{abstract}

2: Approximation of solutions to partial differential equations (PDE) is an important problem in computational science and engineering. 

3: Using neural networks as an ansatz for the solution has proven a challenge in terms of training time and approximation accuracy. 

4: %

5: In this contribution, we discuss how sampling the hidden weights and biases of the ansatz network from data-agnostic and data-dependent probability distributions allows us to progress on both challenges. 

6: In most examples, the random sampling schemes outperform iterative, gradient-based optimization of physics-informed neural networks regarding training time and accuracy by several orders of magnitude.

7: % 

8: For time-dependent PDE, we construct neural basis functions only in the spatial domain and then solve the associated ordinary differential equation with classical methods from scientific computing over a long time horizon. 

9: %

10: This alleviates one of the greatest challenges for neural PDE solvers because it does not require us to parameterize the solution in time.

11: %

12: For second-order elliptic PDE in Barron spaces, we prove the existence of sampled networks with \(L^2\) convergence to the solution. 

13: %

14: We demonstrate our approach on several time-dependent and static PDEs.

15: We also illustrate how sampled networks can effectively solve inverse problems in this setting.

16: %

17: Benefits compared to common numerical schemes include spectral convergence and mesh-free construction of basis functions.

18: \end{abstract}

19: