By most standards, obtaining a physical core sample from the deep ocean below is an amazing task. While a coring device descends several kilometers in water that gets colder, darker, and more pressured with each meter, a research vessel maintains position over an abyssal plain. The entire process may require more than a day. On a seafloor that makes up about 70% of the earth, the resulting core, a cylinder of sediment that may be a meter long, represents a tiny geographic pinpoint. The resulting science is genuine and worthwhile. In comparison to the space between sampling locations, the coverage it offers is almost absurdly limited.
This is the real-world issue that machine learning has begun to tackle in deep-sea geology, yet the development has been so quiet that most people outside of a somewhat specialized academic community haven’t yet seen it. Millions of years of Earth’s history, including changes in ocean chemistry, tectonic plate movement, mass extinction events, and climate variations, are encoded in the sediment that covers the ocean floor. Understanding sediment characteristics across large, largely unsampled regions is necessary to effectively read that record.
Physical samples and geophysical surveys, which require years and significant financial resources to gather, are the foundation of traditional geological mapping techniques. By identifying the hidden correlations between variables in the scant data that already exists and using those associations to create predictions across the gaps, machine learning takes a different approach to the problem.
When researchers test various algorithms against deep-sea sediment datasets, XGBoost, a gradient boosting algorithm that has become a standard tool in competitive machine learning due to its ability to handle complex, non-linear relationships in tabular data, consistently outperforms other approaches. Particularly in regions with low sampling density, it accurately predicts characteristics like porosity, grain density, and calcite concentration that conventional interpolation techniques could not. Because those characteristics are precisely what geologists need to reconstruct historical ocean conditions or evaluate subsurface resources in frontier basins, this is more practically valuable than academically fascinating.
Because machine learning in earth sciences has occasionally been criticized for yielding data that researchers cannot use to develop physical intuition or explain to colleagues, the interpretability aspect of recent work is noteworthy. Researchers have discovered that the two variables performing the majority of the predictive work in these models are sediment depth and compressional wave velocity thanks to SHAP analysis, a method that links each prediction to particular input properties. Once stated, that conclusion is not surprising, but having it verified quantitatively and consistently across datasets provides geologists with a solid foundation upon which to develop. Not only does the model produce accurate predictions, but it does so for geologically plausible reasons.
In order to categorize sediment limits and seafloor types using bathymetric and sonar data, convolutional neural networks have found a parallel application. This is effectively teaching the algorithm to detect patterns in sound imagery in the same way that image classifiers learn to recognize objects. For anyone attempting to estimate the environmental impact of deep-sea mining activities, the ability of LSTM networks, which were designed for sequential data, to handle temporal issues like sediment plume migration is crucial. As interest in bottom mineral resources increases and environmental regulators pose more challenging issues about what industrial activity at those depths actually disrupts, that final application is becoming more and more pertinent.
