About the Competition

Beneath the Earth’s surface lie vital resources, hidden hazards, and clues to our planet’s history. To access and understand these, geophysicists use seismic waves. When these waves travel through the Earth, they carry information about underground structures. 

Full Waveform Inversion (FWI) is a technique that does just that. It analyzes entire seismic waveforms to build a detailed image of the subsurface. However, traditional methods for FWI are slow and often unreliable due to weak from noisy data. 

In this competition, we aim to develop an AI system that enhances FWI by combining machine learning with the physics of wave propagation. The goal is to improve the speed, accuracy, and generalization of subsurface imaging.

Competition Page:
https://www.kaggle.com/competitions/waveform-inversion/overview

Relevance

Seismic data is useful in many areas, including studying earthquakes, evaluating soil and rock stability for construction, resource exploration and monitoring underground mining operations.

New technology makes these tasks more accurate and less harmful to the environment. For example, we can explore for energy without as much drilling, and we can better predict earthquakes and plan safer buildings. These improvements help protect people, save money, and reduce damage from natural disasters.

Using machine learning in seismic work helps us create clearer and faster images of what’s underground. This makes it easier to understand the Earth and make better choices in science and engineering.

“Transforming early literacy through innovative teaching tools unlocks children’s potential.”

Yale

The competition is hosted by Yale University’s Department of Statistics and Data Science. They are a leader in developing innovative data-driven methods for solving complex scientific problems. They are committed to furthering education and research in the rapidly growing field of data science.

Technical Details

The competition’s main challenge is to reconstruct a high-resolution underground velocity map on a 70 × 70 grid from a multishot seismic waveform recorded at surface receivers. Training data come from the OpenFWI benchmark and are grouped into three families – Vel, Fault, and Style – that progress from simple flat layers to curved layers with faults and finally texture-rich geological styles. Performance is scored with Mean Absolute Error (MAE) averaged over every grid node.

Because the OpenFWI simulator can generate practically endless training examples, overfitting is less of a worry than raw computing power, which means even very large neural networks – think billion-parameter transformers or diffusion models – can be trained. Approaches therefore span a spectrum: fully data-driven networks that map waveform cubes directly to velocity grids, traditional physics-based solvers like adjoint-state FWI that march wave equations for maximum accuracy, and hybrid methods that embed physical laws or use learned priors to guide faster physics loops, all aiming to balance speed, fidelity, and the ability to generalise across every geological style in the competition.

UN Sustainable Development Goals

The competition aligns with UN Sustainable Development Goal #9: Industry, Innovation, and Infrastructure. By improving seismic imaging with AI, we can learn more about the Earth and make informed decisions in science and engineering, such as enabling safer construction and more efficient resource use.