Bayesian Time-Lapse Full Waveform Inversion using Hamiltonian Monte Carlo

This paper proposes a Bayesian sequential approach for time-lapse Full Waveform Inversion using Hamiltonian Monte Carlo to effectively quantify uncertainties in high-dimensional seismic problems by integrating baseline survey data as prior knowledge, demonstrating accuracy comparable to parallel schemes while managing computational costs.

The Gist

Imagine you are trying to figure out what's happening inside a giant, dark, underground cave system (the Earth's crust) without ever digging a hole. You can't see inside, but you can throw rocks (seismic waves) at the cave walls and listen to the echoes. By analyzing how those echoes bounce back, you can build a 3D map of the caves, the rocks, and the hidden treasures (like oil or gas) inside.

This process is called Full Waveform Inversion (FWI). It's like trying to solve a giant, complex jigsaw puzzle where the pieces are sound waves.

Now, imagine you want to see how that cave changes over time. Maybe a gas pocket is slowly turning into water, or oil is being pumped out. This is called Time-Lapse Imaging. You take a "snapshot" of the cave today (the Baseline), and then another snapshot a year later (the Monitor). The difference between the two tells you what changed.

The Problem: The "Foggy Mirror"

Here's the tricky part: The changes underground are often tiny and very specific to one spot. But the data we get is noisy and imperfect. It's like trying to see a tiny scratch on a mirror while someone is shaking the mirror and the room is full of fog.

If you just use standard math to solve this, you get one answer. But you don't know if that answer is right or if it's just a lucky guess. You need to know: "How confident are we in this result?"

This is where Bayesian Inversion comes in. Instead of giving you one single map, it gives you a cloud of possibilities. It says, "There's a 90% chance the gas is here, and a 10% chance it's slightly to the left." It quantifies the uncertainty, which is crucial for making safe decisions about drilling or storing carbon.

The Challenge: Too Many Possibilities

The problem is that the "cloud of possibilities" is huge. It's like trying to find a specific grain of sand on a beach that keeps changing shape. Traditional methods of exploring these possibilities are slow and inefficient; they wander around aimlessly (like a drunk person stumbling in the dark).

The Solution: The "Smart Hiker" (Hamiltonian Monte Carlo)

The authors of this paper use a clever technique called Hamiltonian Monte Carlo (HMC).

Think of the traditional method as a hiker who takes random steps in the dark, hoping to find the treasure.
The HMC method is like a smart hiker with a map and momentum.

• Momentum: Instead of stopping and starting randomly, the hiker builds up speed. If they are moving toward the treasure, they keep going. If they hit a wall (a bad solution), they bounce off efficiently.
• Geometry: This hiker understands the shape of the landscape. They know where the "valleys" (good solutions) are and can slide down into them quickly, rather than stumbling around.

This allows them to explore the massive "cloud of possibilities" much faster and more accurately than before.

The Big Idea: The "Smart Handoff" (Sequential Strategy)

The paper compares two ways to do this time-lapse study:

1. The Parallel Approach (The Twin Strangers): You take the first photo (Baseline) and the second photo (Monitor) and try to solve them completely separately, starting from scratch each time. It's like two strangers trying to solve the same puzzle independently. They might both get it right, but they don't help each other.
2. The Sequential Approach (The Smart Handoff): This is the authors' new proposal.
   • First, you solve the Baseline puzzle. You get a good idea of what the cave looks like.
   • Instead of starting the Monitor puzzle from scratch, you use your knowledge of the Baseline as a starting point.
   • Imagine you are painting a portrait. You paint the first one (Baseline). When you paint the second one (Monitor), you don't start with a blank canvas. You start with a faint sketch of the first painting and just adjust the parts that changed.

Why This Matters

The researchers tested this on a famous computer model called "Marmousi" (a fake underground world). They found:

• Accuracy: The "Smart Handoff" (Sequential) method was just as accurate as solving them separately.
• Robustness: When the data was messy (like if the sensors moved slightly between shots, which happens in real life), the "Smart Handoff" method was much better at ignoring the noise and finding the real changes.
• Uncertainty: Both methods gave similar levels of confidence, but the Sequential method was better at avoiding "ghosts" (fake changes that look real but aren't).

The Takeaway

In simple terms, this paper says: "Don't forget what you learned yesterday when you're trying to figure out what's happening today."

By using the results from the first survey to guide the second survey, and by using a "smart hiker" algorithm to explore the possibilities, we can get a clearer, more reliable picture of how our underground resources are changing, even when the data is noisy or imperfect. This helps oil companies, carbon storage projects, and geologists make safer, better decisions.

Technical Summary

1. Problem Statement

Time-lapse (4D) seismic inversion is critical for monitoring dynamic changes in the Earth's interior, such as reservoir management and CO2 storage. However, these changes are often minute and spatially localized, making them difficult to distinguish from noise and acquisition artifacts.

• The Challenge: Traditional deterministic Full Waveform Inversion (FWI) methods struggle to quantify uncertainties. The FWI problem is ill-posed and high-dimensional, meaning small errors in data or model parametrization can lead to significant, non-trivial errors in the solution.
• The Gap: While Bayesian approaches offer a framework for uncertainty quantification by treating the solution as a probability density function (PDF), standard sampling methods like Markov Chain Monte Carlo (MCMC) are computationally inefficient in high-dimensional spaces due to the "curse of dimensionality."
• Specific Issue: Existing Bayesian time-lapse methods often assume repeatable acquisition geometries or rely on joint inversions that may not fully leverage prior knowledge from baseline surveys to constrain monitor surveys.

2. Methodology

The authors propose a probabilistic Bayesian sequential approach for time-lapse FWI, utilizing Hamiltonian Monte Carlo (HMC) as the sampling engine.

A. Bayesian Formulation

Instead of seeking a single deterministic model, the goal is to sample the posterior distribution $\rho(m|d)$.

• Parallel Strategy: Baseline ($m_B$) and Monitor ($m_M$) models are inverted independently using the same uninformative prior.
• Sequential Strategy (Proposed): The posterior distribution of the baseline survey, $\rho_B(m|d)$, is used to construct an informative prior for the monitor survey. Specifically, the monitor prior is approximated as a Gaussian distribution centered on the baseline mean ($m_{mean}^B$) with the baseline covariance ($\Sigma_B$).
  • $\rho_M(m) \approx \mathcal{N}(m_{mean}^B, \Sigma_B)$
  • This leverages the fact that time-lapse changes are small and localized, allowing the baseline inversion results to geometrically constrain the monitor inversion.

B. Hamiltonian Monte Carlo (HMC)

To overcome the inefficiency of standard MCMC in high dimensions ($\sim 10^4$ parameters in this study), the authors use HMC.

• Mechanism: HMC treats model parameters as particles in a physical system with "momentum." It uses Hamiltonian dynamics to propose new states, allowing the sampler to traverse the parameter space more efficiently than random walks.
• Mass Matrix Tuning: A critical component of HMC is the mass matrix ($M$). The authors employ a depth-dependent tuning strategy (based on water layer depth and maximum model depth) to adjust particle masses. This improves convergence and reduces the number of forward calculations required compared to using a standard identity matrix.

C. Experimental Setup

• Model: A subset of the Marmousi model (3.0 km offset, 1.5 km depth) with two gas reservoirs.
• Time-Lapse Scenario: A velocity variation of $\sim 0.1$ km/s (approx. 5%) representing gas-to-water replacement.
• Scenarios Tested:
  1. Perfect Geometry: Repeatable source/receiver locations.
  2. Perturbed Geometry: Source locations in the monitor survey shifted horizontally by 100 m (simulating non-repeatable acquisition).
• Comparison: The proposed sequential Bayesian approach is compared against:
  • Parallel Bayesian inversion.
  • Deterministic FWI (using L-BFGS optimization).

3. Key Contributions

1. Sequential Bayesian Framework: The paper introduces a method where the baseline posterior is explicitly converted into a prior for the monitor survey. This differs from previous works that either used the last baseline samples as starting points or imposed priors directly on the time-lapse difference.
2. HMC Implementation for 4D FWI: Demonstrates the effective application of HMC with a tuned mass matrix to handle the high dimensionality of time-lapse FWI, providing a computationally feasible way to sample uncertainties.
3. Analysis of Correlation and Uncertainty: The study rigorously analyzes the correlation between baseline and monitor samples. It shows that while parallel strategies induce strong correlations (due to shared priors), the sequential strategy produces decorrelated samples between surveys, which is crucial for accurate error propagation.
4. Robustness to Non-Repeatable Geometry: The authors demonstrate that the sequential approach is more robust to acquisition geometry variations (non-repeatability) than the parallel approach, reducing artifacts in the final time-lapse image.

4. Results

• Accuracy: Both sequential and parallel Bayesian strategies produced time-lapse estimates with errors of similar magnitude. However, the sequential strategy provided clearer identification of the reservoirs, particularly in the non-repeatable geometry scenario.
• Artifacts: The parallel strategy (both Bayesian and deterministic) exhibited "decoherent" small-scale structures (artifacts) in the time-lapse difference, likely due to imperfect cancellation of baseline and monitor errors. The sequential strategy suppressed these artifacts significantly.
• Uncertainty Quantification:
  • Parallel: Showed lower uncertainty in deep regions due to strong correlation with baseline samples, but this correlation was artificial (stemming from the shared prior) rather than physical.
  • Sequential: Showed slightly higher individual uncertainties but resulted in time-lapse uncertainties comparable to the parallel method because the correlation term in the error propagation formula ($\sigma_{TL} = \sqrt{\sigma_M^2 + \sigma_B^2 - 2\sigma_{MB}}$) was handled more naturally.
• Non-Repeatable Geometry: When sources were shifted, the parallel strategy produced significant negative anomalies and artifacts. The sequential strategy remained robust, correctly identifying the reservoir changes, confirming that using the baseline posterior as a prior helps mitigate non-repeatability issues.

5. Significance and Conclusion

This work establishes that integrating baseline posterior information as a prior for monitor surveys is a superior strategy for time-lapse FWI, especially when dealing with non-repeatable acquisition geometries.

• Decision Making: By providing a probabilistic framework, the method allows geoscientists to distinguish between true geomechanical changes and inversion artifacts, which is vital for reservoir management and CO2 monitoring.
• Computational Efficiency: The use of HMC with a tuned mass matrix makes high-dimensional Bayesian time-lapse inversion feasible, overcoming the limitations of traditional MCMC.
• Future Outlook: The authors note that while the sequential strategy is robust, its success depends on the quality of the baseline estimation. Future improvements could involve incorporating well-log data to further constrain vertical resolution or using burn-in periods to reduce sample dispersion.

In summary, the paper successfully demonstrates that a Bayesian sequential approach using HMC offers a more reliable, artifact-free, and robust method for time-lapse seismic inversion compared to traditional parallel or deterministic methods.