Acoustic Full Waveform Inversion with Hamiltonian Monte Carlo Method

This paper proposes a novel mass matrix tuning strategy based on seismic acquisition geometry to enhance the efficiency and accuracy of Hamiltonian Monte Carlo sampling for solving the ill-posed acoustic Full-Waveform Inversion problem under noisy and limited data conditions.

The Gist

Imagine you are trying to figure out what's inside a giant, opaque chocolate cake without cutting it open. You can only tap on the top with a spoon and listen to the sound it makes. This is essentially what geophysicists do when they try to map the Earth's underground using Full Waveform Inversion (FWI). They send sound waves (seismic waves) into the ground and listen to how they bounce back to build a picture of the rocks, oil, and gas hidden deep below.

However, this is a notoriously difficult puzzle. The "cake" is huge, the data is noisy (like trying to hear a whisper in a hurricane), and there are often many different ways the cake could be arranged that would produce the same sound. This is called an "ill-posed problem."

Here is a simple breakdown of what this paper does to solve that puzzle:

1. The Old Way vs. The New Way

• The Deterministic Way (The "One Guess" Approach): Traditional methods try to find the single best answer. They adjust the model until the sound matches the data as closely as possible. But if you start with a slightly wrong guess, you might get stuck in a "local minimum"—like thinking you found the bottom of a valley, but actually, there's a deeper valley right next to it that you missed. Also, this method doesn't tell you how confident you should be in your answer.
• The Probabilistic Way (The "Many Guesses" Approach): This paper uses a method called Hamiltonian Monte Carlo (HMC). Instead of looking for one perfect answer, it takes thousands of different "guesses" (samples) to build a map of all the possible underground structures. This gives scientists a sense of uncertainty: "We are 90% sure there is oil here, but only 50% sure about that other spot."

2. The Problem with the "Many Guesses" Approach

The problem with taking thousands of guesses is that it's incredibly slow and computationally expensive. It's like trying to explore a massive, dark maze by stumbling around randomly. In high-dimensional spaces (where you have millions of variables to adjust), random stumbling is inefficient. You might spend hours walking in circles in the same room of the maze.

3. The Solution: The "Smart Hiker" (Hamiltonian Dynamics)

The authors use HMC, which is like giving the hiker a map and a momentum.

• Imagine a ball rolling down a hill. The shape of the hill represents the "error" in the model. The ball naturally rolls toward the bottom (the best solution).
• HMC uses physics equations to make the ball roll fast and smoothly across the landscape, rather than taking tiny, random steps. This allows it to explore the whole maze much faster.

4. The Secret Sauce: Tuning the "Mass"

The big innovation in this paper is how they tune the "mass" of these rolling balls.

• The Analogy: Think of the underground layers like a multi-story building.
  • The Top Floors (Shallow): These are easy to see. The data is clear. You can use a heavy, slow-moving ball here because you don't need to move fast to find the truth.
  • The Basement (Deep): This is dark and hard to see. The data is noisy and weak. If you use a heavy ball here, it gets stuck or moves too slowly to explore the possibilities.
• The Paper's Strategy: The authors realized that in seismic reflection (looking at echoes from the side), we know less about the deep parts of the Earth. So, they proposed a strategy where the "mass" of the ball changes based on depth.
  • As the simulation goes deeper, they make the "particles" (the balls) lighter.
  • Why? A lighter particle is more agile. It can bounce around more easily in the deep, dark, noisy regions, exploring different possibilities without getting stuck. It's like switching from a heavy boulder to a ping-pong ball when you enter a tricky, narrow cave.

5. The Results

By making the particles lighter as they go deeper, the method:

1. Finds better answers faster: It reconstructs the underground models much more efficiently.
2. Handles noise better: Even when the data is very noisy (like a bad phone connection), the method can still figure out the general shape of the underground.
3. Quantifies uncertainty: It doesn't just give a picture; it tells you, "This part of the picture is blurry because the data is weak," which is crucial for oil companies to decide where to drill.

Summary

This paper is about teaching a computer how to "feel" its way through a dark, noisy underground maze. Instead of dragging a heavy, slow boulder through the whole maze, the authors taught the computer to switch to a light, agile ping-pong ball when it gets to the deep, confusing parts. This allows it to explore the possibilities much faster and give a much clearer, more honest picture of what's hiding beneath our feet.

Technical Summary

1. Problem Statement

Full Waveform Inversion (FWI) is a high-resolution geophysical technique used to reconstruct subsurface physical parameters (e.g., acoustic velocity) by minimizing the difference between observed seismic data and modeled data. However, FWI presents significant challenges:

• Ill-posed Nature: The problem is nonlinear and non-unique; multiple models can fit the observed data, especially in the presence of noise and limited spatial/frequency coverage.
• Deterministic Limitations: Traditional deterministic FWI methods (gradient-based) provide a single "best-fit" solution but fail to quantify the uncertainty of the resulting parameters.
• Probabilistic Challenges: While probabilistic approaches (like Markov Chain Monte Carlo - MCMC) can estimate uncertainty distributions, standard MCMC methods suffer from the "curse of dimensionality." In high-dimensional seismic model spaces, they become computationally inefficient due to slow convergence and high autocorrelation between samples.
• HMC Tuning Difficulty: The Hamiltonian Monte Carlo (HMC) method, which combines gradient information with stochastic sampling, is a promising alternative. However, its application in complex seismic reflection setups has been limited by the difficulty of tuning the mass matrix. An inappropriate mass matrix leads to inefficient exploration of the phase space, causing the sampler to get stuck or move too slowly.

2. Methodology

The authors propose a probabilistic FWI framework using the Hamiltonian Monte Carlo (HMC) method within an acoustic wave approximation.

• Physical Model: The subsurface is modeled as a 2D acoustic medium. The forward problem is solved using the acoustic wave equation with a Ricker wavelet source. The misfit function (energy potential) is defined as the squared difference between observed and modeled data, assuming Gaussian noise.
• HMC Formulation:
  • The model parameters are treated as particles in a phase space with position ($m$) and momentum ($p$).
  • The system evolves according to Hamiltonian dynamics, where the potential energy corresponds to the FWI misfit function.
  • The Leapfrog integrator is used to simulate the dynamics, ensuring symplectic properties (time reversibility and volume preservation).
  • Gradients are computed efficiently using the adjoint state method, which avoids explicit Jacobian calculations by cross-correlating the forward and adjoint wavefields.
• Key Innovation: Depth-Dependent Mass Tuning Strategy:
  • The authors identify that the standard choice of a fixed identity mass matrix ($M=I$) is suboptimal for reflection seismic data, where information content decreases with depth.
  • They propose a new strategy where the particle mass ($\mu$) is dynamically adjusted based on depth ($z$) and the number of HMC steps ($k$).
  • Mechanism: Initially, particles have a uniform mass. As the simulation progresses, the mass for parameters at a specific depth is reduced according to a monotonically increasing function $\gamma(z)$.
  • Physical Interpretation: Reducing the mass makes particles "lighter" as the system approaches a potential energy minimum. This allows for finer exploration of the phase space in deeper regions where data illumination is poor, effectively adapting to the acquisition geometry's lack of information at depth.

3. Key Contributions

1. Feasibility Demonstration: The paper successfully demonstrates the application of HMC to acoustic FWI in a reflection setup, a scenario previously underdeveloped due to tuning difficulties.
2. Novel Mass Matrix Tuning: The primary contribution is the proposal of a depth- and iteration-dependent mass tuning strategy. This overcomes the inefficiency of standard HMC in high-dimensional seismic problems by adapting the sampling dynamics to the specific information content of the seismic survey geometry.
3. Uncertainty Quantification: The study moves beyond finding a single solution to providing a full probabilistic characterization of the subsurface, including mean, variance, and skewness of the velocity models.

4. Results

The method was tested on a cropped version of the Marmousi model (a standard benchmark) under three noise scenarios (low, medium, and high variance).

• Convergence: The proposed tuning strategy significantly improved HMC convergence compared to the standard fixed-mass approach. The normalized potential energy stabilized faster, and the acceptance rate remained high ($\sim63\%$).
• Reconstruction Quality:
  • The tuned HMC reconstructed the main features of the target model faster than the conventional approach, particularly in deep regions ($z > 1$ km) where data constraints are weak.
  • The conventional method required significantly more gradient calculations (computational cost) to achieve similar results, making it impractical for large-scale problems without the tuning strategy.
• Uncertainty Analysis:
  • Variance: The variance maps correctly identified discontinuities in the Marmousi model and showed that uncertainty increases with depth, consistent with the acquisition geometry.
  • Skewness: The analysis revealed non-Gaussian behavior (significant skewness) in the posterior distributions, particularly at depth. This indicates that the "most probable value" (mode) is not always sufficient to characterize the inversion, and relying solely on Gaussian assumptions is inadequate for reflection FWI.
• Noise Trade-off: The study highlighted a trade-off between image sharpness and accuracy. Low noise variance ($\sigma^2$) produced sharp images but with potential velocity inaccuracies, while high variance produced blurred images but mean values closer to the true velocity.

5. Significance and Conclusion

This work bridges the gap between advanced probabilistic sampling methods and practical geophysical inversion.

• Computational Efficiency: By introducing a physics-informed mass tuning strategy, the authors make HMC viable for high-dimensional FWI problems that were previously too computationally expensive for probabilistic methods.
• Robustness: The method provides a rigorous way to quantify uncertainty, which is crucial for risk assessment in oil and gas exploration.
• Future Outlook: The authors suggest that this approach opens the door to applying HMC to large-scale, real-world seismic problems. Future work aims to generalize the strategy using different prior geological information and varying mass updates across iterations.

In summary, the paper establishes that HMC, when coupled with a depth-adaptive mass tuning strategy, is a powerful and efficient tool for solving acoustic FWI, offering superior convergence and comprehensive uncertainty quantification compared to traditional deterministic or standard stochastic methods.