# Highly-scalable, physics-informed GANs for learning solutions of 2019-10-31آ Highly-scalable,...

date post

17-Jun-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Highly-scalable, physics-informed GANs for learning solutions of 2019-10-31آ Highly-scalable,...

Highly-scalable, physics-informed GANs for learning solutions of stochastic PDEs

Liu Yang∗, Sean Treichler†, Thorsten Kurth‡, Keno Fischer§, David Barajas-Solano¶, Josh Romero†, Valentin Churavy‖, Alexandre Tartakovsky¶,

Michael Houston†, Prabhat‡, George Karniadakis∗ ∗ Brown University, Providence, RI 02912, USA {liu yang,george karniadakis}@brown.edu

† NVIDIA, Santa Clara, CA 95051, USA {sean,joshr,mhouston}@nvidia.com ‡ Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA {tkurth,prabhat}@lbl.gov

§ Julia Computing, Cambridge, MA 02139, USA keno@juliacomputing.com ¶ Pacific Northwest National Lab, Richland, WA 99352, USA {david.barajas-solano,alexandre.tartakovsky}@pnnl.gov

‖ Massachusetts Institute of Technology, Cambridge, MA 02139, USA vchuravy@mit.edu

Abstract—Uncertainty quantification for forward and inverse problems is a central challenge across physical and biomedical disciplines. We address this challenge for the problem of modeling subsurface flow at the Hanford Site by combining stochastic com- putational models with observational data using physics-informed GAN models. The geographic extent, spatial heterogeneity, and multiple correlation length scales of the Hanford Site require training a computationally intensive GAN model to thousands of dimensions. We develop a hierarchical scheme for exploit- ing domain parallelism, map discriminators and generators to multiple GPUs, and employ efficient communication schemes to ensure training stability and convergence. We developed a highly optimized implementation of this scheme that scales to 27,500 NVIDIA Volta GPUs and 4584 nodes on the Summit supercomputer with a 93.1% scaling efficiency, achieving peak and sustained half-precision rates of 1228 PF/s and 1207 PF/s.

Index Terms—Stochastic PDEs, GANs, Deep Learning

I. OVERVIEW A. Parameter estimation and uncertainty quantification for subsurface flow models

Mathematical models of subsurface flow and transport are inherently uncertain because of the lack of data about the distribution of geological units, the distribution of hydrological properties (e.g., hydraulic conductivity) within each unit, and initial and boundary conditions. Here, we focus on parameter- ization and uncertainty quantification (UQ) in the subsurface flow model at the Department of Energy’s Hanford Site, one of the most contaminated sites in the western hemisphere. During the Hanford Site’s 60-plus years history, there have been more than 1000 individual sources of contaminants distributed over 200 square miles mostly along Columbia River [1]. Accurate subsurface flow models with rigorous UQ are necessary for assessing risks of the contaminants reaching the Columbia river and numerous wells used by agriculture and as sources of drinking water, as well as for the design of efficient remediation strategies.

B. UQ with Stochastic Partial Differential Equations Uncertain initial and boundary conditions and model pa-

rameters render the governing model equations stochastic. In

this context, UQ becomes equivalent to solving stochastic PDEs (SPDEs). Forward solution of SPDEs requires that all model parameters as well as the initial/boundary conditions are prescribed either deterministically or stochastically, which is not possible unless experimental data are available to provide additional information for critical parameters, e.g. the field conductivity. The Hanford Site has about 1,000 to 10,000 wells equipped with sensors that can provide diverse sources of additional information. To this end, we can formulate – instead of a forward problem – a mixed problem, wherein we fuse observational data and mathematical models into one unified formulation. The solution of the mixed problem will then yield both the quantity of interest (e.g., pressure head) as well as the conductivity fields in the form of stochastic processes.

C. UQ with physics-informed GANs (PI-GANs)

Generative adversarial networks (GANs) can learn prob- ability distributions from data, and hence they can be em- ployed for modeling the stochasticity in physical systems. In order to encode the physics into GANs, we follow the physics-informed neural network approach [2], [3], which was subsequently implemented using GANs for 1D problems in [4]. The authors of [4] demonstrated that WGANs-GP, in particular, can represent stochastic processes accurately and can approximate solutions of SPDEs in high dimensions. Specifically, it was found that when doubling the number of stochastic dimensions (from 30 to 60), the computational cost is approximately doubled, suggesting that the overall computational cost increases with a low-polynomial growth.

This observation is illustrated in Figure 1. The green line shows how stochastic dimension increases rapidly as the correlation distance of the behavior of interest shrinks. (Some modeling problems related to the Hanford Site have relative correlation distances on the order of one part per million.) A brute-force attempt to model the stochastic processes in a 2-D system without knowledge of the physics would require (at least) a quadratically-increasing number of sensors (the red

ar X

iv :1

91 0.

13 44

4v 1

[ ph

ys ic

s. co

m p-

ph ]

2 9

O ct

2 01

9

Fig. 1: Visualization of relationships between stochastic di- mension, correlation length, and the rough sensor counts required for brute-force modeling of k/u fields vs. learning distributions using a PI-GAN. The three networks considered in this paper are indicated with blue circles.

line in the figure), whereas PI-GANs can, in principle, feasibly tackle very high dimensional stochastic problems. The blue line in Figure 1 shows their expected growth rate, and in this work, we study three specific PI-GAN network configurations indicated by the blue circles.

D. Contributions

Tackling a realistic dataset from the Hanford Site (cor- responding to a large stochastic dimensionality and num- ber of sensors) requires development of a scalable PI-GAN framework that can obtain high levels of performance on computational nodes, and scale across massive HPC systems. Towards accomplishing this goal, our paper makes the follow- ing contributions: • Development and validation of physically-informed GAN

architecture for modeling subsurface flow. First demon- stration of framework to problem with unprecedented stochastic dimensionality (1,000).

• A novel domain decomposition algorithm for GANs involving one generator and hundreds of discriminators.

• Highly optimized generator and discriminator architec- ture implementation in TensorFlow of the proposed do- main decomposition algorithm that obtains 59.3 TF/s and 38.6 TF/s on Volta GPUs

• Highly scalable architecture that obtains 93.1% scaling efficiency up to 4584 Summit nodes.

• Demonstrated 1207 PF/s sustained and 1228 PF/s peak half-precision (FP16) performance for distributed train- ing of PI-GAN architecture on 27,500 Volta GPUs.

II. STATE OF THE ART A. State-of-the-art in SPDE solvers

The full solution of a system of SPDEs is the joint prob- ability density function (PDF) of the state variables. The

most common method for UQ is Monte Carlo simulation (MCS) [5], a straightforward and robust but computation- ally demanding sampling technique. Generalized Polynomial Chaos (PC) [6], [7] methods have emerged in the past two decades as an effective method for solving SPDEs in low to moderate dimensions. All PC methods share the same foundational principles: (i) approximating infinite-dimensional random parameter fields with finite-term expansions (where the number of terms in the finite-term expansion defines the dimensionality of the probability space); and (ii) adopting stochastic Galerkin method or choosing collocation points in the probability space to compute moments of the PDFs [8]. As a result, they all suffer from the so-called “curse of dimen- sionality”: their computational cost increases exponentially with the number of stochastic dimensions. The state-of-the-art PC methods, including enhancements like ANOVA, become less computationally efficient than MCS for approximately 100 random dimensions [9]–[11]. Based on available data, we estimate that the heterogeneity and multiscale dynamics of the Hanford Site requires at least 1,000 random dimensions for adequate representation, which is out of reach of the PC methods. In addition, the PC methods are most efficient for computing the leading moments of the joint PDFs, i.e., the mean describing the average behavior of the system, and the variance measuring the scale of uncertainty. While mean and variance are sufficient for describing uncertainty in some systems, many applications require full PDFs, especially those involving risk analysis or those monitoring rare but possibly catastrophic events. We hypothesize that PI-GANs have the potential to address both challenges; they can mitigate the curse of dimensionality and provide an accurate estimate of the entire PDF of the parameters and states rather than leading moments.

Here it must be noted that in comparison to PC methods, MCS does not suffer from the curse of dimensionality and thus can be employed for UQ in highly complex stochastic problems if large-scale computational resources are available. On the other hand, MCS cannot be used to assimilate measure- ments (which is one of the main objectives of this work) and must be coupled with other methods such Ensemble Kalman filter [12]. The comparative advantages and disadvantages of physics-informed learning and PDE-based methods for parameter estimation are discussed in t

*View more*