1: \begin{abstract}

2:   We introduce the Optimizing a Discrete Loss (ODIL)  framework for the

3:   numerical solution of Partial Differential Equations (PDE) using machine

4:   learning tools.   The framework formulates numerical methods as a minimization

5:   of discrete residuals that are solved using gradient descent and Newton's

6:   methods. We demonstrate the value of this approach on  equations that may have

7:   missing parameters or where no sufficient data is available to form a

8:   well-posed initial-value problem.

9:   The framework is presented for mesh based discretizations of PDEs and inherits

10:   their accuracy, convergence, and conservation properties. It  preserves the

11:   sparsity of the solutions and  is readily applicable to inverse and ill-posed

12:   problems.   It is applied to  PDE-constrained optimization, optical flow,

13:   system identification, and data assimilation using gradient descent algorithms

14:   including those often deployed in machine learning.  We compare ODIL with

15:   related approach that represents  the solution with neural networks.  We

16:   compare the two methodologies and demonstrate advantages of ODIL that include

17:   significantly higher convergence rates and several orders of magnitude lower

18:   computational  cost.  We evaluate the method on various linear and nonlinear

19:   partial differential equations including the Navier-Stokes equations  for flow

20:   reconstruction problems.

21: \end{abstract}

22: