TLDR: The Transparent Earth is a new AI foundation model that uses a transformer-based architecture to reconstruct Earth’s subsurface properties. It integrates diverse and heterogeneous data types (modalities like stress angle, mantle temperature, fault type) by using positional and modality encodings. The model can learn from sparse observations, scale to new data types, and significantly improve prediction accuracy for various geophysical fields, aiming to predict any subsurface property globally.
Scientists at Los Alamos National Laboratory have introduced a groundbreaking new artificial intelligence model called “The Transparent Earth.” This innovative system aims to revolutionize our understanding of the Earth’s subsurface by integrating diverse geological and geophysical data into a single, unified framework. Traditionally, models in Earth science have been highly specialized, focusing on narrow subdisciplines like weather forecasting or seismic modeling. This specialization often struggles with the vast heterogeneity of Earth systems and the wide variety of data types involved.
The Transparent Earth addresses this challenge by employing a transformer-based architecture, a type of AI model known for its ability to process complex, varied inputs. What makes this model particularly powerful is its capacity to learn from “heterogeneous datasets” – meaning data that varies significantly in how much information it contains, its level of detail, and its type. Each “modality” represents a distinct kind of observation, such as the angle of stress in rocks, the temperature deep within the Earth’s mantle, or the type of tectonic plate in a given region.
A core design principle of The Transparent Earth is its flexibility. It uses positional encodings to understand the geographic location of each observation and “modality encodings” derived from text descriptions of each data type. This clever approach allows the model to easily incorporate new types of observations without needing a complete redesign, making it highly scalable. Currently, the model works with eight different modalities, covering everything from directional angles (like stress and strain) to categorical classes (like fault and basin types) and continuous properties (like temperature and sediment thickness).
This advanced architecture enables “in-context learning,” meaning the model can generate predictions even with no initial inputs, relying on its learned global understanding. More importantly, it can significantly improve its predictions by incorporating an arbitrary number of additional observations from any subset of its supported modalities. For instance, on validation data, the model reduced errors in predicting stress angle by more than a factor of three. The researchers have shown that the model’s performance consistently improves as its complexity (number of parameters) increases, demonstrating its robust scalability.
The Transparent Earth represents a significant step towards creating a true “foundation model” for the Earth’s subsurface. Its ultimate goal is to predict any subsurface property, anywhere on Earth. By fusing sparse, multimodal observations from around the world, it can reconstruct geophysical fields with high confidence, even in data-sparse regions where direct observations are scarce or expensive.
The model’s design includes several innovative features. It uses a unified attention-based architecture that allows different types of data – directional, categorical, and continuous – to interact during the learning process. A unique training strategy helps it handle incomplete and spatially separated observations, mimicking real-world data challenges. Furthermore, a “query-driven decoder” allows users to ask for specific predictions at any desired location, without the model relying on simple memorization.
Also Read:
- New Method Uncovers Hidden Earthquake Warning Signals in Seismic Noise
- Unlocking Geographic Insights: Interpretable AI for Location Prediction
This research highlights the immense potential of integrating diverse data types to improve the resolution and reliability of subsurface property maps. The Transparent Earth is not just a technical achievement; it’s a vision for a more comprehensive understanding of our planet’s hidden depths, paving the way for new discoveries in geoscientific applications. You can read the full research paper here.
