A Deep Learning Approach to Describing the Plasma Sheath

E. Webb, Yuzhi Li, Christopher J. McDevitt · 2026

Read A Deep Learning Approach to Describing the Plasma Sheath by E. Webb, Yuzhi Li, Christopher J. McDevitt (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-22566.txt.

Introduction

Webb, Li, and McDevitt use a physics-informed neural network to describe the plasma sheath, the boundary layer where plasma meets a surface. The sheath's rich physics has blocked a general description. Unlike ordinary deep learning, physics-informed networks constrain predictions with governing equations, needing no experimental or simulation data to train. The team uses the network to find parametric solutions for fluid sheath models of varied fidelity. Offline training often takes longer than a traditional solver, but once trained the network predicts sheath profiles across a broad parameter range.

Catalog record

CREATORE. Webb, Yuzhi Li, Christopher J. McDevitt
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54881
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.22566

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Deep Learning Approach to Describing the Plasma Sheath E. Webb, Yuzhi Li, Christopher J. McDevitt Abstract Despite their ubiquity, the rich physics present in a plasma sheath has inhibited the development of a generally applicable description of this critical region. The present study utilizes a physics-informed neural network (PINN) to evaluate a hierarchy of models of the plasma sheath. Unlike traditional deep learning methods, PINNs use the governing PDEs to constrain the predictions of a neural network, and thus do not require any experimental or simulation data to train. In this work, we utilize a PINN to identify the parametric solution to fluid models of different physics fidelity of the plasma sheath. While the offline training time of the PINN is often longer than a traditional solver, once trained, the PINN is able to efficiently predict the sheath profiles across a broad range of parameter regimes, thus yielding an effective surrogate of the plasma sheath. I Introduction The plasma-sheath transition strongly influences a diverse range of applications including magnetic fusion 44 , electric propulsion 22 , plasma diagnostics 20 , and plasma processing 27 . Despite extensive research and the attention obtained from a broad range of applications, this region still presents significant challenges for modeling and simulation to overcome. This is largely due to the rich physics present in the sheath, where fundamental quantities, such as the Bohm criterion, are strongly impacted by a diverse range of plasma phenomena, including collisions 16 ; 40 , an oblique-angle magnetic field 13 , and plasma instabilities 18 ; 6 . The present work seeks to demonstrate the potential of physics-constrained machine learning methods 23 to provide a novel path through which this classic problem can be treated. This is accomplished by considering a hierarchy of fluid models of the plasma sheath, where it is demonstrated that a physics-informed neural network (PINN) 38 ; 51 is able to provide an accurate solution of the sheath profile across a broad range of plasma conditions. Specifically, unlike the majority of machine learning (ML) techniques that have, to date, been applied to low-temperature plasmas 48 , PINNs seek to embed physical constraints into the training of a neural network. Broadly, imposing physical constraints during the training of a neural network serves three purposes. For the case of sparse datasets 2 , physical constraints can be used to regularize predictions of the neural network, thus providing a robust mechanism for interpolating across regions that would otherwise be left largely unconstrained. For the case where the physics of the system is completely specified, including both the governing equations along with initial and boundary conditions, this approach can be used to fully constrain the system even in the absence of data, such that the PINN can be used as an ordinary differential equation (ODE) or partial differential equation (PDE) solver. In contrast, for cases where a large dataset is present, but the physics of the problem is incompletely understood, the same PINN framework can be employed to solve an inverse problem, where synthetic or experimental data are used to close the incomplete set of imposed physics 38 ; 30 , an approach often referred to as physics discovery 10 ; 43 ; 4 . In this paper, we will assume the limit of no synthetic or experimental data, and develop a series of PINNs that solve the system of ODEs characterizing a one-dimensional sheath. For this approach, PINNs solve an optimization problem, where the PINN seeks to minimize the loss associated with the residuals of a system of ODEs together with boundary conditions. If successful, this results in the identification of a solution to the sheath profiles. This approach is applied to several fluid models of the sheath, some of which can be straightforwardly solved with an ODE solver such as Runge-Kutta 7 ; 41 ; 42 ; 19 , but others that require a more involved approach to obtain a steady state solution 3 ; 14 . A distinguishing feature of the PINN approach is that while the offline training time of the PINN is often longer than a traditional ODE or PDE solver, a single PINN is able to learn the parametric solution to the ODE or PDE system of interest 46 ; 32 ; 21 ; 34 . Noting that the online inference time of a PINN is typically on the order of microseconds per prediction, once trained, a PINN can be used to predict sheath profiles across a broad range of physics parameters. Physics-informed neural networks thus provide a means through which efficient surrogates for the plasma sheath can be constructed. In this paper, we build three surrogates of the plasma sheath, trained on fluid models of varying physics fidelity. The resulting models provide an efficient means of investigating fundamental quantities such as the Bohm criterion, sheath width, and potential drop and how they depend on parameters such as the ion-neutral mean-free-path, ion-electron temperature ratio, and ion mass, as well as demonstrate the ability of PINNs to accurately treat sheath physics. The remainder of this paper is organized as follows. The physics-informed machine learning approach employed is briefly described in Sec. II . In Sec. III , three fluid models of the plasma sheath with increasing physics fidelity are defined and the results are included in each section. Section IV describes conclusions and future directions. II Physics-Constrained Machine Learning Neural networks (NN) are powerful tools that are well known for their ability to identify patterns in data 8 . Data-driven approaches are successful when making predictions within the scope of their training data, but require sufficient data to converge well. When data is scarce or expensive to obtain, a data-driven approach becomes prohibitive. Due to the difficulty of obtaining plasma sheath data, we find that physics-constrained approaches 23 provide a path to overcome this relative scarcity of data. Here, physical constraints are embedded in the training of a neural network (NN), providing a natural regularization term and mechanism for extrapolating beyond available data. This can be accomplished by the identification of NN architectures that enforce select properties of the solution 11 , or by enforcing physical properties as soft constraints in the loss function. Regarding the latter, PINNs 38 ; 15 ; 52 ; 47 include the residuals, which is simply an ODE or PDE rearranged to equal zero, of the governing equations into the loss function along with the data, yielding a loss function of the form: \text{Loss}=L_{Data}+L_{ODE}, (1a) L_{Data}=\frac{1}{N_{Data}}\sum^{N_{Data}}_{i}\left[\phi\left(x_{i};\bm{a}_{i}\right)-\phi_{i}\right]^{2}, (1b) L_{ODE}=\frac{w_{ODE}}{N_{ODE}}\sum^{N_{ODE}}_{i}\mathcal{R}^{2}\left(x_{i};\bm{a}_{i}\right), (1c) where \phi is a field predicted by the PINN, x_{i} is the spatial coordinate, \bm{a} are the physical parameters, \mathcal{R} is the residual of an ODE, and w_{ODE} is the relative weight applied to the ODE component of the loss function. For a system of ODEs multiple data and ODE loss terms will be present, but for simplicity we only list one each here. The first term corresponds to losses against available data, which could include initial or boundary conditions, and the second term represents losses associated with the residual of an ODE. By minimizing the loss, a PINN seeks to find a solution that obeys both the available data set together with the governing equations. For the case

Read the complete text