🔢 mathematics

ROMNet: a hybrid reduced order modeling and machine learning approach to waveform inversion

This paper introduces ROMNet, a hybrid approach that combines reduced order modeling with a neural network to efficiently map ROM matrices to wave speeds, thereby overcoming the computational challenges and cycle-skipping issues inherent in traditional full waveform inversion.

Original authors: Liliana Borcea, Alexander Mamonov, Kui Ren, Haizhao Yang, Chugang Yi

Published 2026-08-27

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Liliana Borcea, Alexander Mamonov, Kui Ren, Haizhao Yang, Chugang Yi

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 trying to see inside a mountain, a human body, or the deep earth without cutting it open. Scientists do this by sending waves—sound waves in the case of medical imaging or seismic waves for geology—into the ground and listening to how they bounce back. By analyzing these returning echoes, they can build a picture of the hidden structures inside. This process, known as waveform inversion, is a powerful tool for finding oil reserves, mapping underground faults, or diagnosing medical conditions. However, it is notoriously difficult. The waves travel through complex, uneven materials, and the math required to reverse-engineer the journey from the echo back to the source is full of traps. If the initial guess about the underground speed of sound is even slightly off, the calculation can get stuck in a false solution, mistaking a small ripple in the data for a major geological feature. This problem, often called "cycle skipping," has long been a major bottleneck, forcing researchers to rely on slow, expensive computer simulations or to accept blurry, inaccurate images.

A team of researchers has developed a new hybrid method called ROMNet that aims to bypass these traps and speed up the process significantly. Their approach combines two distinct fields: a mathematical technique that simplifies the raw data into a manageable form, and machine learning, which teaches a computer to recognize patterns in that simplified data. Instead of trying to solve the complex wave equations from scratch every time, the researchers first convert the messy wave measurements into a compact algebraic structure, which they call a reduced order model. Think of this as distilling a complex recipe down to its essential ingredients. While this step is well-understood, figuring out exactly what the underground speed of sound is based on this distilled data has remained a difficult, slow optimization problem.

The innovation in this study lies in how they solve that final step. The researchers trained a neural network—a type of artificial intelligence—to act as a translator. This network takes the distilled data structure and maps it to a new, simpler mathematical form that has a direct and predictable relationship with the wave speed. By doing this, they transformed a difficult, iterative guessing game into a much smoother path to the answer. To test their idea, they ran thousands of computer simulations using two very different types of underground models. The first set consisted of random, patchy variations in wave speed, while the second used realistic geological models featuring faults, folded rock layers, and massive salt domes, similar to those found in real-world oil exploration.

The results showed that this new method outperformed existing machine learning approaches that try to learn the entire process from scratch. In tests on random underground patterns, the new method produced the most accurate images among the learning-based techniques. More importantly, when faced with the complex, realistic geological structures, it maintained high accuracy where other learning-based methods struggled. The researchers found that while the new method could not perfectly reconstruct underground features that were completely unlike anything it had seen during training, it was exceptionally good at providing a high-quality starting point. In fact, using the new method's output as a starting guess allowed the traditional, slower methods to converge on the correct answer much faster, cutting the total computing time from over fourteen minutes down to just ten seconds for a single inversion.

This work suggests that the future of imaging the unseen may not lie in choosing between pure mathematics or pure artificial intelligence, but in weaving them together. The researchers demonstrated that by using machine learning to simplify the relationship between data and the physical world, they could avoid the common pitfalls that have plagued the field for decades. While the method relies on simulations and requires training on specific types of data, it offers a promising path forward for making high-resolution underground imaging faster, cheaper, and more reliable. The study confirms that with the right hybrid approach, it is possible to navigate the complex landscape of wave physics without getting lost in the noise.

Technical Summary: ROMNet – A Hybrid Reduced Order Modeling and Machine Learning Approach to Waveform Inversion

Problem Statement
Waveform inversion is an inverse problem aimed at estimating the wave speed $c(x)$ of a heterogeneous, inaccessible medium from time-resolved measurements of acoustic waves gathered by an active array of sensors. The forward map, which relates the wave speed to the measurements, is nonlinear and highly oscillatory. This oscillatory nature leads to "cycle skipping," a phenomenon where standard nonlinear least-squares formulations, known as Full Waveform Inversion (FWI), become trapped in spurious local minima. While recent data-driven approaches have introduced Reduced Order Models (ROMs) as algebraic surrogates of the wave operator to bypass direct data fitting, the subsequent mapping from the ROM matrix to the wave speed remains a computationally expensive, iterative optimization problem that is difficult to solve efficiently.

Methodology: The ROMNet Approach
The paper proposes ROMNet, a hybrid methodology that combines data-driven reduced order modeling with machine learning to simplify the inversion process. The approach is built upon the following logical steps:

Data-Driven ROM Construction: The method first computes an operator ROM matrix, denoted as $A_{ROM}$ , directly from the array response data $M(t)$ . This matrix is a Galerkin projection of the wave operator $-c^2(x)\Delta$ onto a space spanned by wave field snapshots. Crucially, $A_{ROM}$ is computed in a non-iterative, data-driven manner using tools from numerical linear algebra (specifically, block Cholesky factorization of a data-derived mass matrix), avoiding the need to solve the wave equation during this stage.
Theoretical Motivation: The authors observe that while the mapping from wave speed $c(x)$ to $A_{ROM}$ is nonlinear, the dependence of the underlying orthonormal basis $V(x)$ on the wave speed is "smooth and slowly changing" compared to the raw wave snapshots. This suggests that the mapping from $A_{ROM}$ to a modified matrix $A_{ML}$ can be learned to have a simpler, explicit quadratic dependence on $c(x)$ .
Neural Network Mapping: A neural network, $G_\theta$ , is trained to map the input $A_{ROM}$ to a target matrix $A_{ML}$ . The target $A_{ML}$ is defined as:
$A_{ML} = -\int_\Omega dx \, V(x; c_0)^T c(x) \Delta [c(x) V(x; c_0)]$
where $V(x; c_0)$ is an orthonormal basis computed in a known reference medium with wave speed $c_0(x)$ . This formulation ensures that $A_{ML}$ depends explicitly and quadratically on the unknown wave speed $c(x)$ .
Inversion Process: The inversion is performed in two stages:
- Stage 1 (Prediction): The trained network $G_\theta$ processes the data-derived $A_{ROM}$ to predict the target matrix $A_{ML}$ .
- Stage 2 (Optimization): The wave speed is estimated by solving a minimization problem that fits a parametrized search speed $\hat{c}(x; \eta)$ to the predicted $A_{ML}$ . Because $A_{ML}$ has an explicit dependence on $c(x)$ , this optimization does not require solving the wave equation at every iteration, unlike standard ROM or FWI approaches.

Key Contributions

Hybrid Framework: The paper introduces ROMNet, which leverages the mathematical tractability of data-driven ROMs and the approximation power of neural networks to decouple the complex physics of wave propagation from the optimization of the wave speed.
Explicit Dependence: By training the network to map $A_{ROM}$ to a matrix with explicit quadratic dependence on $c(x)$ , the authors transform a difficult, implicit inverse problem into a more tractable optimization problem.
Computational Efficiency: The methodology eliminates the need for repeated wave equation solves during the optimization phase. The most expensive step (wave equation solving) is replaced by a single forward pass through the neural network and numerical quadrature.
Resolution Analysis: The paper provides a theoretical resolution analysis indicating that while micro-scale features ( $\ll \lambda_0$ ) cannot be resolved, variations on scales $\gtrsim \lambda_0$ can be estimated, consistent with the Abbe-Rayleigh limit.

Numerical Results
The authors evaluated ROMNet using two distinct training datasets:

Random Gaussians: Media with wave speeds modeled as superpositions of random Gaussian functions.
GeoFWI: A dataset containing geologically realistic structures (folds, faults, salt bodies) generated from sedimentary rules.

The performance was compared against:

Direct ROM Inversion: The standard iterative optimization of the ROM misfit.
Fourier-DeepONet: An operator network approach.
InversionNet: A standard end-to-end CNN baseline for FWI.
Oracle: An idealized scenario using the exact $A_{ML}$ (unavailable in practice).

Findings:

Accuracy: ROMNet achieved the lowest relative $L_2$ error among all learning-based methods (outperforming Fourier-DeepONet and InversionNet) on both in-distribution and out-of-distribution test sets. Its accuracy approached that of the direct ROM inversion and the Oracle.
Generalization: While ROMNet struggled with media significantly different from the training distribution (e.g., heavy salt structures in the BP-2004 model), it produced estimates that were sufficiently accurate to serve as excellent initial guesses for the direct ROM inversion.
Computational Cost: ROMNet demonstrated a massive reduction in computational cost. On a single NVIDIA A100 GPU, a full inversion took approximately 10.5 seconds for ROMNet compared to 14 minutes for the direct ROM inversion. This is a reduction of roughly two orders of magnitude, primarily because ROMNet avoids solving the wave equation during the iterative optimization.
Robustness: The method was shown to be robust to additive Gaussian noise when a specific regularization technique (boosting the zero-time data block) was applied.

Significance and Claims
The paper claims that ROMNet offers a significant advancement in waveform inversion by addressing the "cycle skipping" problem through a hybrid physics-informed machine learning approach. The authors emphasize that while the neural network cannot perfectly generalize to geologies far outside the training distribution, the method's primary value lies in its ability to provide high-quality initial guesses for traditional ROM-based inversions. This "warm start" capability allows the more expensive, physics-based ROM inversion to converge in fewer iterations, thereby retaining the accuracy of the pure ROM approach while drastically reducing the total computational burden. The work positions ROMNet not as a replacement for physics-based solvers in all scenarios, but as a powerful accelerator and initializer that bridges the gap between purely data-driven deep learning and rigorous PDE-constrained optimization.

More like this