⚡ electrical engineering

Survey of Deep Learning and Physics-Based Approaches in Computational Wave Imaging

This paper presents a comprehensive review and structured framework for integrating deep learning techniques with traditional physics-based methods to address the computational challenges and limitations of computational wave imaging across diverse scientific domains.

Original authors: Youzuo Lin, Shihang Feng, James Theiler, Yinpeng Chen, Umberto Villa, Jing Rao, John Greenhall, Cristian Pantea, Mark A. Anastasio, Brendt Wohlberg

Published 2026-04-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Youzuo Lin, Shihang Feng, James Theiler, Yinpeng Chen, Umberto Villa, Jing Rao, John Greenhall, Cristian Pantea, Mark A. Anastasio, Brendt Wohlberg

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 figure out what's inside a giant, opaque black box without opening it. You can't see inside, but you can throw different things at it (like sound waves) and listen to how they bounce back. This is the core idea of Computational Wave Imaging (CWI). It's used everywhere: from looking for oil deep underground, to checking if a bridge has cracks, to seeing if a patient has a tumor without using harmful radiation.

For a long time, scientists solved this puzzle using Physics-Based Methods. Think of this like a master detective who knows every law of nature perfectly. They calculate exactly how sound should travel through different materials. If the sound bounces back differently than expected, they deduce what's inside.

The Good: It's very accurate and trustworthy.
The Bad: It's incredibly slow and computationally heavy. It's like trying to solve a 1,000-piece puzzle by calculating the physics of every single piece's shape and weight. Also, if the starting guess is slightly off, the whole solution can collapse (a problem called "cycle-skipping").

Recently, Deep Learning (AI) has entered the chat. This is like hiring a super-fast, super-observant apprentice who has seen millions of puzzles before. Instead of calculating physics from scratch, the AI looks at the pattern of the bouncing sound and says, "I've seen this pattern before; it usually means there's a crack here."

The Good: It's lightning fast once trained.
The Bad: It's a "black box." If you show it a puzzle it has never seen before (like a weird new type of rock or a different kind of metal), it might guess wildly wrong because it's just memorizing patterns, not understanding the rules.

The Big Idea: The "Cyborg" Detective

This paper is a massive review of how we are now combining the Master Detective (Physics) and the Super Apprentice (AI) to create the ultimate "Cyborg Detective."

The authors argue that neither side works perfectly alone. Physics is too slow; AI is too brittle. The future lies in Hybrid Methods where the AI learns the rules of physics, or the physics uses AI to speed things up.

Here are the main ways they are mixing these two worlds, explained with analogies:

1. The "Supervised" Approach (The Flashcard Method)

How it works: You show the AI thousands of pairs of "Sound Bounce" and "What's Inside" (like flashcards). The AI memorizes the connection.
The Catch: You need a massive library of flashcards. In the real world, we often don't have the "answer key" (we don't know what's inside the ground or a human body for sure).
Analogy: It's like a student who memorizes the answer key for a practice test but fails the real exam because the questions are slightly different.

2. The "Self-Supervised" & "Unsupervised" Approach (The Puzzle Solver)

How it works: The AI doesn't need the answer key. Instead, it tries to solve the puzzle by ensuring its guess makes sense physically. If the AI guesses there's a rock inside, it simulates the sound bouncing off that rock. If the simulated sound doesn't match the real sound, the AI knows it's wrong and tries again.
The Catch: It can get stuck in a "local minimum," meaning it finds a solution that almost works but isn't the best one.
Analogy: It's like trying to find your way out of a dark maze by feeling the walls. You might get stuck in a small dead-end that feels like the exit, but it's not the real one.

3. The "Plug-and-Play" Approach (The Team-Up)

How it works: This is a clever dance between the two. The Physics side does a rough calculation to get close to the answer. Then, the AI side steps in to "clean up" the image, removing noise and fixing weird artifacts, acting like a professional photo editor. Then the Physics side checks it again. They take turns until the image is perfect.
Analogy: Imagine a physicist builds a rough sketch of a house, and an artist comes in to paint the details and fix the colors. Then the physicist checks the structural integrity, and the artist fixes the shading. They keep passing the drawing back and forth until it's perfect.

Real-World Examples from the Paper

The paper looks at three main places where this tech is changing the world:

Seismic Imaging (Looking for Oil or Carbon Storage):
- The Problem: We need to see underground to store CO2 safely.
- The Fix: AI helps detect tiny leaks of gas that traditional methods miss. It's like using a metal detector that can also tell you the exact shape of the buried coin, not just that something is there.
Industrial Testing (Checking for Cracks):
- The Problem: Checking airplane wings or pipelines for cracks without breaking them.
- The Fix: AI can look at the sound waves bouncing off a wing and instantly spot a hairline crack that a human might miss. It's like having X-ray vision for metal.
Medical Imaging (Ultrasound):
- The Problem: Traditional ultrasound can be blurry or take a long time to process.
- The Fix: AI can turn a blurry ultrasound into a crystal-clear image of a tumor, helping doctors diagnose cancer earlier. It's like turning a low-resolution photo into 4K HD instantly.

The Future: What's Next?

The paper concludes with a few important lessons:

Don't trust the AI blindly: If the AI sees something it hasn't been trained on, it might hallucinate. We need to know how sure the AI is (Uncertainty Quantification).
Physics is still the boss: The best AI models are the ones that respect the laws of physics. An AI that ignores physics is just guessing.
The "Black Box" problem: We need to make AI more explainable so doctors and engineers can trust why it made a decision.

In a nutshell: This paper is a roadmap for the future of seeing the invisible. It tells us that the future isn't about choosing between "old-school physics" or "new-school AI." It's about building a super-tool that uses the speed of AI and the reliability of physics to see the world more clearly than ever before.

1. Problem Definition

Computational Wave Imaging (CWI) is the process of reconstructing the internal physical properties (e.g., density, bulk modulus, wave velocity) of a material volume by analyzing wave signals that traverse it. The core challenge is an inverse problem: inferring the medium's properties ( $m$ ) from boundary measurements ( $d$ ) governed by a forward wave equation ( $d = f(m) + \eta$ ).

The paper identifies three primary application domains:

Seismic Full-Waveform Inversion (FWI): Subsurface exploration for energy and geology.
Acoustic/Industrial Ultrasound: Non-Destructive Testing (NDT) for material defects.
Medical Ultrasound Computed Tomography (USCT): Imaging for breast cancer and other medical diagnostics.

Key Challenges in CWI:

Ill-posedness: Limited data coverage (2D surface measurements for 3D volumes) leads to non-unique solutions.
Non-convexity & Cycle-Skipping: The optimization landscape contains many local minima. If the initial guess is poor, the solution may converge to an incorrect state where the modeled waveform is shifted by more than half a cycle relative to the observed data.
Computational Cost: Traditional physics-based methods require iterative solving of Partial Differential Equations (PDEs), which is computationally prohibitive for large-scale or real-time applications.
Data Scarcity: High-fidelity ground-truth labels (e.g., exact velocity maps) are often unavailable or expensive to obtain in real-world scenarios.

2. Methodology

The paper provides a structured taxonomy of methods, categorizing them by Supervision Strategy and Learning Strategy.

A. Supervision Strategies

Fully Supervised: Trains deep neural networks (e.g., InversionNet, U-Net) on paired datasets of waveforms and ground-truth properties. While fast at inference, these models struggle with generalization to out-of-distribution data and require massive labeled datasets.
Semi-Supervised: Leverages abundant unlabeled data alongside scarce labeled data. Techniques include adaptive data augmentation, where an initial model generates "pseudo-labels" for unlabeled waveforms, which are then used to refine the training set iteratively.
Self-Supervised: Generates supervision signals from the data itself without external labels. Examples include CycleGANs for low-frequency extrapolation (translating between high-frequency field data and synthetic low-frequency data) and contrastive learning.
Unsupervised/Unlabeled: Uses the physical wave equation as a hard constraint. Methods like SiameseFWI and UPFWI minimize the mismatch between observed and simulated waveforms directly, often using differentiable physics simulators within the network.

B. Learning Strategies (Physics Integration)

Simulations/Measurements-Driven: Purely data-driven approaches where physics is implicitly learned from large volumes of synthetic or experimental data (e.g., Fourier-DeepONet).
Physics-Aware: Explicitly incorporates physical laws (e.g., the wave equation) into the loss function or network architecture. Physics-Informed Neural Networks (PINNs) are a prime example, where the PDE residual is minimized alongside data fidelity.
Parameterized Solutions: Uses neural networks to parameterize the solution space (e.g., representing the velocity map as a function of a latent vector). This acts as an implicit regularizer, constraining the solution to a lower-dimensional, physically plausible manifold.
Plug-and-Play (PnP) Priors: Decouples the data-fidelity step (physics-based update) from the regularization step. A pre-trained denoiser (e.g., DnCNN) acts as a "prior" within an iterative optimization loop (like ADMM), allowing for flexible integration of learned priors without retraining the physics solver.

C. Emerging Trends

Uncertainty Quantification (UQ): Moving beyond point estimates to predictive distributions. Methods like InvNet_UQ use quantile regression to estimate confidence intervals, crucial for risk-aware decision-making.
Generative AI (GenAI): Utilizing Diffusion Models and GANs to learn the prior distribution of velocity models. These models can generate realistic synthetic data for augmentation or solve inverse problems by sampling from a learned high-probability manifold, effectively reducing ill-posedness.

3. Key Contributions

Unified Taxonomy: The paper consolidates over 200 papers from diverse fields (geophysics, medicine, NDT) into a single framework, distinguishing methods by supervision and learning strategies.
Critical Analysis of Trade-offs: It provides a comparative analysis of computational costs, generalization capabilities, and robustness. For instance, it highlights that while fully supervised methods are fast, they fail on unseen data, whereas physics-aware methods offer better generalization at higher computational costs.
Application-Specific Case Studies:
- Seismic: Demonstrates how semi-supervised adaptive augmentation improves CO2 leakage detection in the Kimberlina dataset, outperforming traditional FWI in resolving small-scale features.
- Industrial: Shows CNN-based reconstruction of high-contrast defects in multi-layered structures and non-invasive temperature imaging in high explosives.
- Medical: Evaluates task-informed learning for breast USCT, showing that optimizing for specific clinical tasks (tumor detection) yields better diagnostic performance than standard pixel-wise error minimization.
Identification of Limitations: The authors explicitly address the "Generalization Gap" caused by discrepancies between idealized training simulations and noisy, complex real-world measurements (Input, Output, and Forward Model discrepancies).

4. Results

Performance: Hybrid methods (combining physics and ML) consistently outperform pure data-driven or pure physics-based approaches in terms of robustness to noise and generalization.
Efficiency: ML-based inference is orders of magnitude faster than traditional iterative FWI (e.g., seconds vs. hours), making real-time monitoring feasible.
Accuracy: In the Kimberlina dataset, semi-supervised methods successfully localized CO2 leakage plumes that traditional FWI missed due to resolution limits, while avoiding the shape distortions seen in fully supervised models.
Uncertainty: UQ-enabled models (InvNet_UQ) maintained accuracy in noisy conditions (SNR=10) where standard networks failed, providing spatial confidence maps that guide further data acquisition.

5. Significance and Future Outlook

This survey is significant because it bridges the gap between traditional computational physics and modern deep learning, offering a roadmap for the next generation of wave imaging.

Scientific Impact: It shifts the paradigm from "black-box" deep learning to Scientific Machine Learning (SciML), where physical laws are embedded to ensure interpretability and physical consistency.
Engineering Impact: By reducing computational costs and improving robustness, these methods enable real-time monitoring in critical infrastructure (e.g., carbon sequestration, pipeline safety) and faster, radiation-free medical diagnostics.
Future Directions: The paper identifies key hurdles for future research:
- Generalization: Developing methods that bridge the gap between synthetic training data and real-world field data (domain adaptation).
- Interpretability: Moving beyond "black boxes" to understand why a model makes specific predictions.
- Task-Oriented Metrics: Shifting evaluation from standard image metrics (MSE, SSIM) to task-specific performance (e.g., tumor detection accuracy).
- Co-design: Jointly optimizing data acquisition hardware and inversion algorithms.

In conclusion, the paper argues that the future of CWI lies in hybrid frameworks that leverage the speed and pattern recognition of deep learning while strictly adhering to the fundamental laws of wave physics to ensure reliability and generalizability.