AI-Powered Subsurface Imaging: Understanding and Advancing Continuous Representation Full-Waveform Inversion

Peering Beneath the Surface with AI: The Power of Full-Waveform Inversion

      Understanding what lies beneath surfaces, whether it's Earth's crust, human tissue, or industrial materials, is critical across numerous fields. Full-Waveform Inversion (FWI) is a sophisticated technique designed to do just this. It’s essentially a high-resolution imaging method that estimates the physical properties of a medium by analyzing how waves—like seismic waves, ultrasound, or electromagnetic pulses—travel through it. FWI has found extensive application in geophysical exploration for oil and gas, advanced medical imaging, and non-destructive testing (NDT) to inspect materials without causing damage.

      While FWI promises unparalleled resolution, its practical deployment has faced a significant hurdle: an extreme sensitivity to the "initial model." This initial model is essentially an educated guess about the subsurface properties. If this guess isn't accurate enough, the FWI process can struggle to converge, leading to unreliable results or outright failure, a phenomenon often called "cycle-skipping." This challenge has prompted researchers to explore innovative solutions, with Artificial Intelligence (AI) emerging as a powerful ally.

The Challenges of Conventional FWI

      Conventional FWI methods, while theoretically powerful, operate as a highly nonlinear inverse problem. Imagine trying to precisely map the internal structure of a complex, layered cake just by sending vibrations through it and listening to the echoes. If your initial idea of the cake's layers, density, and ingredients is significantly off, your analysis might completely misinterpret the echoes. Similarly, in FWI, complex geological structures (like salt bodies or faults), sparse data acquisition, missing low-frequency components, and noise can all compound the problem.

      Historically, mitigating these issues involved meticulous initial model generation, complex regularization techniques, and advanced optimization algorithms. However, these approaches still often required an accurate, smooth starting model—a challenging prerequisite to obtain, especially in areas with complex near-surface conditions or highly heterogeneous subsurface layers. The inability to start with a sufficiently accurate initial model has remained a bottleneck, preventing FWI from reaching its full potential in many real-world scenarios.

Continuous Representation FWI: A New Approach

      Recent advancements have introduced a transformative concept: Continuous Representation FWI (CR-FWI). This innovative framework re-imagines how physical parameter models are represented. Instead of using discrete grids to map subsurface properties, CR-FWI employs coordinate-based neural networks, such as Implicit Neural Representation (INR). Think of INR as a neural network that learns to describe a continuous landscape—like a mountain range—from individual coordinates, rather than storing a pixelated image. This allows for a smooth, high-fidelity representation of the physical properties.

      The shift to CR-FWI has demonstrated significant advantages, particularly in terms of robustness. Unlike conventional FWI, CR-FWI can achieve satisfactory inversion results even when starting with a very simple, constant initial model and working with lower quality seismic data. This greatly reduces the dependence on expensive and time-consuming initial model generation. However, this robustness has come with a trade-off: CR-FWI typically exhibits a slower convergence rate, especially for high-frequency components, meaning it takes more iterations to achieve highly detailed and precise results compared to traditional methods that start with excellent initial models. This discrepancy is a key area of ongoing research, as highlighted by a recent conference paper by Ruihua Chen et al. (ICLR 2026).

Unveiling the Mechanism: The Wave-Based Neural Tangent Kernel

      To understand why CR-FWI behaves the way it does—offering robustness but slower high-frequency convergence—the aforementioned research developed a unified theoretical framework centered around an extended concept: the wave-based Neural Tangent Kernel (NTK). In simple terms, the Neural Tangent Kernel is a mathematical tool that helps predict how a neural network learns and converges during training. It provides insights into the network's behavior, often without needing to run a full training simulation.

      The researchers' analysis of this wave-based NTK for FWI yielded two profound insights:

• Dynamic Nature: Unlike the standard NTK used in other AI applications, the wave-based NTK for FWI is not constant; it changes dynamically throughout the training process. This is primarily due to the inherent nonlinearity of the FWI problem itself. This dynamic behavior opens new avenues for theoretical research into PDE-constrained nonlinear inverse problems.
• Eigenvalue Decay and Learning Speed: The way the "eigenvalues" of this wave-based NTK decay provides a powerful explanation for CR-FWI's performance characteristics. Rapid eigenvalue decay indicates that the network prioritizes learning smooth, low-frequency information first, before slowly picking up on intricate, high-frequency details. This phenomenon, often referred to as the "frequency principle" or "spectral bias" in neural networks, explains why CR-FWI is robust (it gets the big picture right even with poor initial models) but is slower to capture fine details (high-frequency components). This smooth learning process reduces reliance on precise initial models, making it a more forgiving and adaptable solution.

From Theory to Breakthrough: New CR-FWI Methods

      Armed with these theoretical insights into the wave-based NTK and its eigenvalue decay, the researchers proposed and developed several novel CR-FWI methods designed to optimize the trade-off between robustness and convergence speed.

• LR-FWI (Low-Rank Tensor Function Representation): This method incorporates low-rank tensor function representation into the CR-FWI framework. By encoding the intrinsic smoothness and low-rank properties of physical parameter models, LR-FWI improves inversion robustness while also accelerating the convergence of high-frequency components. This is achieved through carefully tailored eigenvalue decay properties that balance learning speed across different frequencies.
• MPE-FWI (Multi-Grid Parametric Encoding): Multigrid parametric encoding has been shown to yield a more favorable eigenvalue spectrum distribution in its NTK. Consequently, MPE-FWI exhibits a slower eigenvalue decay compared to standard INR-based FWI. This property allows MPE-FWI to achieve faster convergence for high-frequency components, leading to higher inversion accuracy, especially when smooth initial models are available.
• IG-FWI (Hybrid INR and Multi-Resolution Grid): Perhaps the most significant development, IG-FWI combines the strengths of Implicit Neural Representation (INR) with a multi-resolution grid approach. This hybrid representation strikes a superior balance, offering both the enhanced robustness of INR-based methods and the faster high-frequency convergence rates seen in conventional or MPE-FWI approaches. This method effectively merges the best of both worlds, promising both adaptability and precision in challenging inversion problems.

      The efficacy of these proposed methods, particularly IG-FWI, has been demonstrated through extensive applications in geophysical exploration across various complex models, including Marmousi, 2D SEG/EAGE Salt and Overthrust, 2004 BP, and the more realistic 2014 Chevron models. These results consistently show superior performance compared to both conventional FWI and existing INR-based FWI methods.

Real-World Impact: Enhancing Critical Operations

      The implications of these advancements in CR-FWI are far-reaching and highly impactful across various industries:

• Geophysical Exploration: For energy companies and geologists, more robust and accurate subsurface imaging means better identification of oil and gas reservoirs, improved risk assessment for drilling operations, and more efficient resource management. This translates directly into substantial cost savings and increased operational success rates.
• Medical Imaging: In healthcare, enhanced FWI can lead to clearer and more reliable diagnostics for complex biological tissues. Imagine more precise tumor detection, better monitoring of organ health, or improved surgical planning, all contributing to better patient outcomes.
• Non-Destructive Testing (NDT): Industries relying on NDT, such as aerospace, infrastructure, and manufacturing, can benefit from superior material defect detection. This leads to increased safety, reduced maintenance costs, and prolonged asset lifespans.

      Companies like ARSA Technology leverage advanced AI and IoT solutions to help enterprises tackle complex challenges, from AI video analytics for operational intelligence to comprehensive custom AI solutions designed for unique industrial demands. The principles of robustness, efficient convergence, and privacy by design, central to CR-FWI advancements, resonate with ARSA’s focus on deploying practical, proven, and profitable AI systems in real-world, mission-critical environments. Our expertise, honed since 2018, lies in transforming complex theoretical breakthroughs into deployable solutions that generate measurable ROI and enhance security across various industries.

Conclusion: The Future of High-Precision Inversion

      The work on continuous representation Full-Waveform Inversion, supported by the wave-based Neural Tangent Kernel framework, marks a significant step forward in the field of inverse problems. By providing a deeper theoretical understanding of why AI-driven representations enhance robustness while simultaneously addressing the challenge of high-frequency convergence, this research paves the way for a new generation of FWI methods. The development of hybrid approaches like IG-FWI, which intelligently balance the strengths of different representation techniques, promises to unlock FWI's full potential. These innovations will enable industries to obtain more accurate, reliable, and detailed subsurface insights, ultimately driving better decision-making and operational efficiency.

      To learn more about how advanced AI and IoT solutions can transform your enterprise operations, we invite you to explore ARSA Technology's offerings and contact ARSA for a free consultation.

Source:

      Ruihua Chen, Yisi Luo, Bangyu Wu, and Deyu Meng. "Unveiling the Mechanism of Continuous Representation Full-Waveform Inversion: A Wave Based Neural Tangent Kernel Framework." Published as a conference paper at ICLR 2026. https://arxiv.org/abs/2603.22362