Reproduce “Weak Adversarial Neural Pushforward Method for the Wigner Transport Equation” on OpenResearch

About this paper

Andrew Qing He, Wei Cai, Sihong Shao

Researchers developed a Weak Adversarial Neural Pushforward Method (WANPM) to solve the Wigner transport equation, which exactly reformulates the nonlocal quantum potential into a pointwise finite difference and employs a signed neural pushforward architecture to manage negative quasi-probability distributions. This approach offers a mesh-free, Jacobian-free framework for high-dimensional quantum transport simulations without truncating the Moyal series or encountering the numerical sign problem.

Key insights

• Plane-wave test functions can precisely invert the Fourier transform embedded in the Wigner potential operator's definition, allowing its exact reduction to a local finite difference within a weak formulation.
• Decomposing the Wigner function into separately learned positive and negative distributions, guided by a trainable mixing weight, provides a robust way to handle quasi-probability negativity without the variance issues of stochastic signed-particle methods.
• An adversarial training framework, where test functions are learned to maximize the residual, ensures that the neural networks accurately satisfy the Wigner transport equation in a weak sense.

Summary by alphaXiv.