Time-embedded convolutional neural networks for modeling plasma heat transport.

Area of Science:

• Plasma Physics
• Computational Physics
• Machine Learning

Background:

• Accurate modeling of heat transport in plasmas is crucial for understanding fusion energy and astrophysical phenomena.
• Previous methods like the Luciani-Mora-Virmont (LMV) Informed Neural Network (LINN) showed promise but struggled with strongly time-dependent and nonlocal conditions.
• Existing models often rely on quasistationary assumptions that fail in dynamic plasma regimes.

Purpose of the Study:

• To develop a novel neural network architecture capable of modeling spatiotemporal heat transport in plasmas under strongly nonlocal and time-dependent conditions.
• To overcome the limitations of previous methods that produced physically inconsistent results in dynamic regimes.
• To create a data-driven yet physically consistent framework for multiscale plasma transport.

Main Methods:

• Introduction of a time-embedded convolutional neural network (TCNN).
• The TCNN architecture is informed by physical principles to capture coupled evolution of heat flux and nonlocality.
• Training and validation using fully kinetic Particle-in-Cell (PIC) simulations across various collisionalities.

Main Results:

• TCNN accurately reproduces nonlocal plasma heat transport dynamics.
• The model demonstrates enhanced predictive performance compared to previous approaches, particularly in time-dependent regimes.
• Physically consistent predictions are achieved even under strongly nonlocal conditions.

Conclusions:

• The TCNN provides a robust and physically consistent data-driven framework for modeling complex plasma transport.
• Time modulation, coupled prediction, and convolutional depth are key features enhancing TCNN's predictive power.
• This approach advances the capability to simulate multiscale plasma phenomena accurately.