💻 computer science

Evaluating Operators for Acoustic Wave Simulation Correction

This paper evaluates twelve correction architectures, ranging from linear regression to Fourier Neural Operators, within a Deep Finite Difference framework to mitigate numerical dispersion in two-dimensional anisotropic acoustic wave simulations across 27,000 heterogeneous velocity fields.

Original authors: Pascal Tribel, Gianluca Bontempi

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pascal Tribel, Gianluca Bontempi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how sound waves travel through a complex, rocky underground. Scientists use powerful computer programs (solvers) to simulate this. However, these programs have a flaw: because they break time and space into tiny, chunky steps, they introduce a "fuzziness" or distortion called numerical dispersion. It's like trying to draw a smooth, perfect circle using only square Lego bricks; the result is a jagged, slightly wrong shape.

This paper is about finding the best "eraser" to fix those jagged edges.

The Setup: A Race Between Two Solvers

The researchers set up a race between two different computer programs:

The "Fast but Flawed" Solver (FD): This is a standard, fast method that uses a 4th-order grid. It's quick but produces those jagged, distorted wave patterns.
The "Slow but Perfect" Solver (PS): This is a high-precision method that acts as the "gold standard" or the reference. It produces the smooth, correct wave patterns.

The team generated 27,000 different scenarios with varying underground speeds (heterogeneous velocity fields). For every scenario, they had a "bad" picture from the Fast solver and a "perfect" picture from the Slow solver.

The Goal: Teaching a Machine to Fix the Mistakes

The goal was to train a machine learning model to look at the "bad" picture and instantly predict what the "perfect" picture should look like. Think of it as teaching a student to look at a blurry, pixelated photo and mentally reconstruct the sharp, high-definition version.

The researchers didn't just try one type of student; they tested 12 different "architectures" (learning methods), ranging from simple math tricks to complex deep learning networks.

The Big Discovery: The "Two-Step" Strategy

The most important finding wasn't just which model was best, but how they should be used.

The researchers found that the best strategy was a two-step process:

Step 1 (The Linear Fix): First, use a simple, basic math tool (Principal Components Regression) to remove the obvious, predictable errors. This is like using a basic filter to remove the biggest blur from a photo.
Step 2 (The Smart Fix): Then, feed that "mostly fixed" image into a complex neural network to clean up the remaining, tricky details.

The Analogy: Imagine you are painting a wall.

Method A (Direct): You try to paint the final masterpiece directly over the dirty, cracked wall. It's hard, and the result is messy.
Method B (The Paper's Winner): You first sand down the wall and apply a primer (the simple math fix) to make it smooth. Then, you paint the masterpiece (the neural network) on top.
Result: The paper shows that skipping the "sanding and priming" step and trying to paint directly on the rough wall leads to much worse results.

The Winners and Losers

After testing all 12 methods, here is how they ranked:

The Champion: The Fourier Neural Operator (FNO) took the top spot.
- Why? The errors in the sound waves are "spectral" (related to waves and frequencies). The FNO is like a musician who understands the entire song at once, rather than just listening to one note at a time. This matches the nature of the error perfectly.
The Runner-Up: 2D Convolutional Neural Networks (CNNs) came in second. These are standard image-processing AI. They did well but were slightly less effective than the FNO.
The Surprising Underperformers: Simpler methods like Linear Regression and K-Nearest Neighbors (which just looks for similar past examples) did okay, but they couldn't capture the complex, non-linear distortions as well as the deep learning models.

The Takeaway

This paper is the first to systematically test these different "erasers" specifically for acoustic waves (sound) in 2D.

Don't skip the basics: You must remove the simple, linear errors first before asking a complex AI to fix the rest.
Global vs. Local: The best AI (FNO) works by looking at the whole picture globally (like understanding the whole melody), which is better for fixing wave errors than AI that only looks at small, local patches (like looking at individual notes).

In short, if you want to fix computer simulations of sound waves, use a complex AI that understands the whole wave, but make sure you clean up the easy mistakes before you let the AI do its magic.

Technical Summary: Evaluating Operators for Acoustic Wave Simulation

Problem Statement
Numerical simulations of acoustic wave propagation are essential for applications such as seismic imaging, fault detection, and non-destructive testing. However, these simulations suffer from two primary sources of mismatch with reality: physical simplifications (e.g., neglecting attenuation) and numerical discretization errors. While machine learning (ML) has been applied to correct numerical dispersion artifacts in elastic wave equations, this framework has not been systematically applied to the acoustic wave equation. Furthermore, existing studies typically evaluate a single architecture in isolation, leaving a gap in understanding which learning method is best suited for bridging the gap between a suboptimal numerical solver and a high-fidelity reference.

Methodology
The authors instantiate the "Deep Finite Difference" framework for 2D anisotropic acoustic wave propagation. The core task is formulated as operator learning: mapping spatiotemporal traces generated by a low-fidelity solver to those of a high-fidelity reference.

Data Generation: The study generates 27,000 trace pairs using two solvers on a 2D periodic uniform grid ( $N_t = 256$ $N_{t} = 256$ time steps):
- Reference Solver: Fourier Pseudo-Spectral (PS) method with second-order leapfrog time integration.
- Proxy Solver: Fourth-order centered Finite Difference (FD) method with second-order leapfrog time integration.
- Inputs: Heterogeneous velocity fields are drawn from Negahdari et al. (2025), and the source is a Ricker wavelet modulating a 2D Gaussian envelope. Periodic boundary conditions are used to isolate FD dispersion errors.
Correction Pipeline: The correction task is treated as a residual learning problem. An initial Principal Components Regression (PCR) correction is applied to the FD traces to remove the linear component of the error. Subsequent models then predict the residual to refine the trace.
Architectures Evaluated: Twelve correction architectures are benchmarked under a unified 10-fold cross-validation protocol. The suite includes:
- Classical ML: Linear Regression (LR), K-Nearest Neighbors (KNN), Extra Trees (ET), and PCR.
- Deep Learning: Multi-Layer Perceptrons (MLP), 1D and 2D Convolutional Neural Networks (CNN1d, CNN2d), U-Nets, and Fourier Neural Operators (FNO).
Metrics: Performance is measured using Root Normalized Mean Squared Error (RNMSE). Statistical significance is assessed via the Wilcoxon signed-rank test with Holm correction.

Key Results

Impact of Linear Pre-processing: All pipelines incorporating an initial PCR guess substantially outperform those operating directly on raw FD traces. This confirms that removing the linear dispersion component first facilitates effective non-linear correction.
Architecture Performance:
- FNO: The Fourier Neural Operator achieved the lowest mean RNMSE (0.0790), followed closely by CNN2d (0.0804).
- CNN2d vs. CNN1d: 2D convolutions marginally outperformed 1D convolutions. While dispersion error manifests as a temporal phase shift, the spatial dependence of the acoustic wave suggests that 2D kernels capture both spatial and temporal correction axes more effectively.
- Statistical Significance: While FNOs showed the best performance, the absolute margin over CNN2d was modest. The Critical Differences diagram indicated no statistically significant difference between the top-performing architectures (FNO and CNN2d) given the sample size ( $n=10$ folds).
Baseline Comparison: All ML pipelines reduced the dispersion mismatch below the baseline PCR error (0.0903), demonstrating that a non-linear relationship exists between the FD and PS traces that ML can exploit.

Significance and Contributions
The paper claims two primary contributions:

First Systematic Benchmark for Acoustic Correction: This work presents the first systematic benchmark for correcting numerical dispersion in 2D anisotropic acoustic wave simulations, a setting previously unaddressed in the literature. It introduces a publicly benchmarked dataset of 27,000 trace pairs.
Unified Evaluation of Diverse Architectures: Unlike prior works that test single models, this study evaluates a wide range of methods—from classical regression to advanced Neural Operators—under a single protocol. It establishes that while FNOs theoretically align well with the spectral nature of dispersion errors, their practical advantage over CNNs in this specific acoustic setting is modest and requires further investigation to confirm statistical significance.

The authors conclude that while FNOs currently yield the best performance, the lack of a statistically significant difference over CNN2d suggests that practitioners should weigh performance against computational cost and parameter count. Future work is suggested to extend this framework to 3D simulations and investigate alternative Neural Operator architectures.