DeepParticle: Learning invariant measure by a deep neural network minimizing Wasserstein distance on data generated from an interacting particle method

Zhongjian Wang, Jack Xin, Zhiwen Zhang

TL;DR

This paper introduces DeepParticle, a novel deep learning method for learning invariant measures of stochastic dynamical systems. By minimizing the Wasserstein distance between data distributions, DeepParticle can efficiently generate samples from complex, unknown target distributions, significantly accelerating simulations of physical phenomena like reaction-diffusion fronts.

What Problem Does This Paper Address?

Simulating complex stochastic dynamical systems, particularly those involving reaction-diffusion fronts in chaotic flows, often relies on interacting particle methods (IPMs). While effective, these methods can be computationally expensive, especially when advection dominates (high Péclet numbers). A key bottleneck is the need to evolve particle distributions from a simple initial state (like uniform distribution) to a complex invariant measure. This paper addresses the challenge of efficiently learning these invariant measures, aiming to provide a better starting point for IPM simulations and thus accelerate the overall computation.

Methods

The core of the DeepParticle method is a deep neural network (DNN) designed to learn a mapping from a simple source distribution (e.g., uniform) to a complex target invariant measure. The DNN is trained by minimizing the 2-Wasserstein distance between the output of the network and samples from the target distribution.

Key methodological contributions include:

  • Physically Parameterized DNN Architecture: The network incorporates a "parameter net" that allows it to learn the dependency of the invariant measure on physical parameters (like the Péclet number, denoted by $\kappa$). This network includes both general linear layers and specialized layers whose weights are generated from a shallow network conditioned on the physical parameter. Skip connections are also used to improve information flow.
  • Wasserstein Distance Minimization: The training objective is to minimize the Wasserstein distance, a metric that measures the "cost" of transforming one probability distribution into another. This is particularly useful when the target distribution is unknown or complex.
  • Iterative Divide-and-Conquer Algorithm for Transition Matrix: Directly computing the Wasserstein distance involves solving a linear program to find an optimal transport plan (a doubly stochastic matrix). This can be computationally prohibitive for large numbers of particles. DeepParticle proposes an efficient mini-batch interior point method to approximate this transition matrix, significantly reducing computational cost.
  • Data-Driven Approach: The method does not require the invariant measure or the mapping to be invertible or known in closed form, making it applicable to a wide range of problems where only data samples are available.

Datasets

The paper uses synthetic datasets and data generated from simulations of interacting particle methods (IPMs).

  • Synthetic Data: A 1D uniform distribution mapped to a 1D normal distribution with varying standard deviations.
  • IPM-Generated Data:
    • For 2D cellular flows: Samples of invariant measures for reaction-diffusion particles in advecting flows with varying Péclet numbers ($\kappa$).
  • For 3D time-dependent Kolmogorov flows: Samples of invariant measures for reaction-diffusion particles in chaotic flows.

The paper does not introduce new benchmark datasets but rather uses data generated from established IPM simulations.

Key Insights and Ideas

  • Learning Complex Distributions from Data: The success of DeepParticle demonstrates that DNNs can effectively learn to generate samples from complex, unknown probability distributions, which are crucial for understanding and simulating physical systems.
  • Parameter-Dependent Learning: The ability to learn how physical parameters influence the invariant measure is a significant advantage, allowing for efficient prediction across a range of parameter values without retraining from scratch.
  • Wasserstein Distance as a Powerful Training Objective: Using the Wasserstein distance provides a robust way to compare and match distributions, especially when analytical forms are unavailable.
  • Computational Efficiency through Algorithmic Innovation: The mini-batch approach for approximating the transition matrix is key to making the Wasserstein distance minimization computationally feasible for large-scale problems.
  • "Warm Start" for Simulations: The learned invariant measure can serve as an excellent initial condition for IPM simulations, leading to substantial speedups (e.g., 4x to 8x) compared to starting from a uniform distribution.

Key Results

  • Accurate Invariant Measure Generation: The DeepParticle method successfully reproduces the sharpening effect of the invariant measure as the Péclet number ($\kappa$) decreases in both 2D and 3D chaotic flow simulations.
  • Efficient Prediction: The trained network can predict invariant measures for parameter values not seen during training, demonstrating good interpolation capabilities.
  • Significant Speedup: In a 2D front speed computation, using the DeepParticle-generated invariant measure as a "warm start" for IPM accelerated the convergence to the principal eigenvalue approximation by 4x to 8x compared to a uniform start.
  • Qualitative Agreement: Histograms of generated samples closely match the target distributions (ground truth PDFs) in the numerical examples.

Limitations and Future Work

  • Computational Cost of Transition Matrix: While the mini-batch approach is efficient, finding the optimal transition matrix can still be costly as the number of particles increases. The paper mentions that the cost of finding optimal $\gamma$ increases with $N$.
  • Comparison with Other Methods: The authors note that a detailed comparison with other distribution mapping techniques like Normalizing Flows (NF) and Generative Adversarial Networks (GANs) is left for future work.
  • Future Applications: The authors plan to apply DeepParticle to learn and generate invariant measures for other stochastic particle systems. They also suggest that the method is applicable to other interacting particle systems beyond KPP front speeds.

Summary updated

Abstract

We introduce the so called DeepParticle method to learn and generate invariant measures of stochastic dynamical systems with physical parameters based on data computed from an interacting particle method (IPM). We utilize the expressiveness of deep neural networks (DNNs) to represent the transform of samples from a given input (source) distribution to an arbitrary target distribution, neither assuming distribution functions in closed form nor a finite state space for the samples. In training, we update the network weights to minimize a discrete Wasserstein distance between the input and target samples. To reduce computational cost, we propose an iterative divide-and-conquer (a mini-batch interior point) algorithm, to find the optimal transition matrix in the Wasserstein distance. We present numerical results to demonstrate the performance of our method for accelerating IPM computation of invariant measures of stochastic dynamical systems arising in computing reaction-diffusion front speeds through chaotic flows. The physical parameter is a large Peclét number reflecting the advection dominated regime of our interest.
BibTeX
@article{wang2022deepparticle,
  author = {Zhongjian Wang and Jack Xin and Zhiwen Zhang},
  title = {{DeepParticle: Learning invariant measure by a deep neural network minimizing Wasserstein distance on data generated from an interacting particle method}},
  journal = {Journal of Computational Physics},
  year = {2022},
  doi = {10.1016/j.jcp.2022.111309},
}
Figure