1 of 22

Neural Monte Carlo PDE Solvers

Presenter: Guandao Yang

2 of 22

Partial Differential Equation Solvers are Useful !

  • Partial differential equations are equations that involves partial derivatives.

E.g. Laplace equation:

  • PDE solvers are useful of graphics, vision, and physics simulations.

(Poisson eq) Image Editing

(Perez, Gangnet, and Blake, 2012)

(Biharmonic eq) Deformation

(Jacobson et. al, 2011)

(Navier-Stokes) Fluid Simulation

(Rioux-Lavoie et. al, 2022)

3 of 22

Solving PDEs - Finite-element method

Figure Credit: Keenan Crane

4 of 22

Solving PDEs - Finite-element method

Can we solve PDEs without discretization?

Figure Credit: Keenan Crane

Discretization can be difficult!

5 of 22

Can we solve PDEs without discretization?

Graphics: Monte Carlo Method

Learning: Neural fields / PINNs

(Shawney and Crane, 2020)

Neural network represent the mapping from spatial coordinate to the PDE solutions; train with losses to enforce PDE constraints.

Derive an integral solution for the PDE; estimate the integral by Monte Carlo method.

(Raissi et. al., 2019, Sitzmann et. al., 2020)

6 of 22

The “Bias and Variance Tradeoff” between MC and NF

Graphics: Monte Carlo Method

Learning: Neural fields / PINNs

Unbiased (accurate)

High variance (slow)

Low-variance (fast)

Biased (inaccurate)

7 of 22

Can we combine the advantages of these two methods?

Graphics: Monte Carlo Method

Learning: Neural fields / PINNs

Unbiased (accurate)

High-variance (slow)

Low-variance (fast)

Biased (inaccurate)

Hybrid Solver

Fast

Accurate

8 of 22

Monte Carlo Solver for Laplace

Figure Credit: Keenan Crane

9 of 22

Monte Carlo Solver for Laplace

10 of 22

Monte Carlo Solver for Laplace

Potentially Long Walk!!!

11 of 22

Ours: Hybrid Solver for Laplace

12 of 22

How do we obtain a Neural Field solution?

Supervise directly with noisy estimate of the MC solver

13 of 22

Hybrid is faster than Monte Carlo methods!

14 of 22

We are more accurate under the same compute!

15 of 22

Limitation: Hybrid solver is still Biased

16 of 22

Solution: Use network as Control Variates

Zilu

Li

Guandao

Yang

Xi

Deng

Bharath

Hariharan

Gordon

Wetzstein

Leonidas

Guibas

17 of 22

Solution: Use network as Control Variates

18 of 22

Achieve Lower Error with Equal Number of Samples

19 of 22

About Two Times Faster

20 of 22

Accurate Hybrid PDE Solvers can be fast and accurate!

NF as MC caching

NF as MC control variate

NF as MC sampling guidance?

SIGGRAPH Asia Conf 2023

(this talk)

On going…

Future direction?

21 of 22

Thanks for listening!

Happy to chat more during poster!

Zilu Li

Guandao Yang

Xi Deng

Chris

De Sa

Bharath

Hariharan

Steve

Marschner

Gordon

Wetzstein

Leonidas

Guibas

22 of 22

Neural Monte Carlo Method - Solving PDE without Discretization

Presenter: Guandao

Collaborators: Zilu Li, Xi Deng,

Bharath Hariharan, Chris De Sa, Steve Marschner, Leo Guibas, Gordon Wetzstein