SWEEP (Seismic Wave Equation Exploration Platform) A Unified Solver Framework for Differentiable Wave Physics

Shaowen Wang, Tariq Alkhalifah · 2026

Read SWEEP (Seismic Wave Equation Exploration Platform) A Unified Solver Framework for Differentiable Wave Physics by Shaowen Wang, Tariq Alkhalifah (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-2604-14189.txt.

Introduction

Wang and Alkhalifah present SWEEP, a unified solver library for modeling seismic wavefields and running inversions. It supports acoustic, elastic, attenuative, VTI, and TTI engines plus their Born approximations. Built-in automatic differentiation enables full-waveform inversion, least-squares reverse-time migration, and other gradient methods. A plug-and-play design lets users combine custom losses, multi-GPU runs, and neural networks with ease, making SWEEP a flexible platform for advanced seismic inverse problems.

Catalog record

CREATORShaowen Wang, Tariq Alkhalifah
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31204
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.14189

Browse the full library · About the Tower of Babel

Full text

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

SWEEP (Seismic Wave Equation Exploration Platform) A Unified Solver Framework for Differentiable Wave Physics Shaowen Wang, Tariq Alkhalifah Abstract SWEEP (Seismic Wave Equation Exploration Platform) is a unified and extensible wave equation solver library designed for wavefield modeling and inversion. It supports a wide range of wave propagation engines, including acoustic, elastic, attenuative, VTI, TTI, and their Born approximations, among others. With a built-in support for automatic differentiation, the framework enables seamless implementation of full-waveform inversion (FWI), least-squares reverse time migration (LSRTM), and other gradient-based optimization methods. It also features a plug-and-play architecture, allowing easy integration and flexible combination of custom loss functions, multi-GPU computation, neural networks, and more. This makes Sweep a powerful and customizable platform for tackling advanced seismic inverse problems. 1 Introduction Wave-equation-based velocity model building and imaging are fundamental tasks in seismic exploration. These processes typically involve solving wave equations that govern the propagation of seismic waves through the Earth’s subsurface. For inverse problems, the corresponding adjoint wave equations must also be solved. Traditionally, this requires manually deriving the adjoint equations and associated gradient expressions using the adjoint-state method ( 14 ) . However, deriving and implementing these adjoint wave equations can be complex and demands careful attention to wave physics, including the temporal consistency between forward and backward wavefields (the dot product test, 3 ), boundary conditions, and source terms. This complexity often presents challenges in maintaining code quality, ensuring correctness, and achieving computational efficiency. Automatic differentiation (AD) provides a powerful alternative to the manual derivation and implementation of adjoint wave equations. By leveraging AD, we can compute gradients of loss functions with respect to model parameters—or wavefields—as long as the involved operations are differentiable. The current challenge of developing AD-based solvers lies in balancing implementation simplicity ( 16 ; 9 ) with computational performance ( 11 ) : PyTorch-based implementations offer high development efficiency but often suffer from lower runtime performance, while CUDA-based implementations provide high computational efficiency at the cost of increased complexity and reduced development flexibility. In this context, the emergence of JAX ( 2 ) presents a promising solution to this dilemma. JAX offers automatic differentiation capabilities similar to PyTorch but combines them with just-in-time (JIT) compilation to achieve performance comparable to hand-written CUDA codes. Built on the XLA compiler, JAX can optimize and compile Python functions into high-performance machine code, enabling efficient execution on GPUs and TPUs. Building on PyTorch and JAX libraries, we developed SWEEP (Seismic Wave Equation Exploration Platform), which supports both frameworks as backends for the same wave equation solver. It is also a repository with a clean and concise codebase—particularly in its wave equation implementation—that significantly enhances usability and flexibility. Its streamlined and efficient solver design allows users to easily implement higher-level operations such as wavefield separation, common image gather (CIG) computation, and other advanced processing tasks. Furthermore, it enables seamless integration with diverse workflows, including neural-network-based velocity representations, multi-scale inversion strategies, source encoding techniques, and optimized loss function designs. By reducing the complexity of the low-level complexity operations, SWEEP empowers users to concentrate on high-level workflow designs necessary to address the underlying scientific challenges. 2 Features In Sweep, wave equations are primarily implemented using the time-domain finite-difference (FDTD) method, with a focus on 2D simulations. A central component of the framework is the time-stepping scheme, which updates the wavefield at each time step based on its previous values and the spatial derivatives defined by the underlying physical model. While the implementation follows standard finite-difference techniques, the time-stepping process is abstracted in a way that is structurally similar to a recurrent neural network (RNN) ( 7 ; 12 ) . This analogy serves as a conceptual abstraction rather than a functional equivalence—the propagation is not carried out using neural networks, but the recurrent formulation offers a unified and extensible interface for incorporating various wave equations. Importantly, this abstraction is not limited to supporting different physical models (e.g., acoustic, elastic, viscoacoustic); it also provides standardized interfaces for common operations such as source injection and data recording. This modular design enhances code reusability, extensibility, and compatibility with automatic differentiation frameworks, enabling flexible integration into workflows such as full-waveform inversion (FWI), least-squares reverse time migration (LSRTM), and learning-based approaches. 2.1 Flexible implementation There are three main modules in the Sweep wave equation implementation: Propagator , Source , and Receiver . The Propagator module is responsible for the time-stepping process, handling memory allocation and wavefield updates at each time step. The Source module manages the injection of wavefields into the simulation, supporting flexible source configurations. The Receiver module records wavefields at specified locations and for specified components, enabling data collection for analysis or inversion tasks. To implement a specific wave equation, users only need to define the wavefield update behavior for a single time step. There is no need to manage model shapes or memory allocation—these are automatically inferred from the input model dimensions and configurations. The Propagator module orchestrates the full time-stepping process based on the provided parameters, and, in combination with the Source and Receiver modules, automatically executes the complete simulation. The following code snippet demonstrates how we can numerically solve a simple acoustic wave equation in Sweep : ⬇ def step ( u_now , u_pre , vp , dt , h , b , lap_u_now ): u_next = 2 * u_now - u_pre + vp **2 * dt **2 * lap_u_now return u_next , u_now class Acoustic : def __init__ ( self , ...): ... @property def models ( self ): return [ ’vp’ ] @property def wavefields ( self ): return [ ’h1’ , ’h2’ ] def func ( self , * args , ** kwargs ): lap_u_now = laplace ( args [0], args [4], self . kernel ) return step (* args , lap_u_now ) The step function defines the wavefield update for a single time step, while the Acoustic class encapsulates the entire wave equation implementation. The models property specifies the required physical model parameters (e.g., velocity), and the wavefields property defines symbolic names for the wavefields involved in the simulation. These symbolic names indicate how many wavefields are needed and serve as keys for injecting sources and recording data. The func method implements the time-stepping logic, including the computation of spatial derivatives using a Laplacian operator. 2.2 Batch modeling To maximize computational efficiency, a batch simulation approach is provided. This allows multiple shots to be computed simultaneously, reducing the number of kernel launches compared to processing individual shots sequentially. Assume that we need to launch the forward modeling kernel k times for a single source. If we model n shots sequentially, the kernel must be launched n\times k times. However, with batch modeling, we only need to launch the kernel k times, once for all shots, significantly reducing kernel launch overhead and improving overall performance. In addition, this approach enables simulations with multiple different velocity models simultaneously, supporting more diverse and flexible simulation scenarios. 2.3 High performance computing Sweep is designed to leverage high-performance computing (HPC) resources, including GPUs and TPUs. Although the underlying wave equation implementation targets single-device execution, it can be easily extended to multiple GPUs using PyTorch’s DistributedDataParallel wrapper or JAX’s pmap , enabling users to effectively exploit shot-based parallelism.

Read the complete text