📄 earth_science

Surface-to-subsurface inference via AlphaEarth: A shared foundation-model representation for diverse subsurface properties

This study demonstrates that the AlphaEarth foundation model's surface embeddings serve as a versatile shared geospatial representation for inferring diverse subsurface properties, acting either as augmenting features alongside traditional covariates or as primary descriptors depending on the target physics and prediction scale.

Original authors: Nori Nakata, Jingxiao Liu, Guodong Chen, Rie Nakata, Charuleka Varadharajan

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nori Nakata, Jingxiao Liu, Guodong Chen, Rie Nakata, Charuleka Varadharajan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess what the basement of a house looks like, but you aren't allowed to go inside. You can only walk around the outside, look at the roof, the garden, the driveway, and the neighborhood.

That is the challenge scientists face when trying to understand what lies deep underground. They can't drill holes everywhere (it's too expensive and slow), so they have to guess the underground conditions based on what they see on the surface.

This paper introduces a new "super-observer" tool called AlphaEarth to help solve this guessing game. Here is how it works, broken down simply:

The Problem: The "Blind" Basement

Scientists need to know two specific things about the ground:

How stiff the soil is near the surface (called VS30). This is crucial for knowing how much the ground will shake during an earthquake.
How hot it is deep underground (subsurface temperature). This is important for finding geothermal energy or understanding how the Earth's crust moves.

The problem is that we only have a few "drill holes" (measurements) scattered across the country. In some places, we have lots of data; in others, we have almost none. Traditional methods try to guess the underground by looking at simple things like "how steep is the hill?" But hills aren't the whole story. A steep hill could be solid rock or loose sand, and the surface doesn't always tell you the difference.

The Solution: The "All-Seeing" Camera (AlphaEarth)

The researchers used a tool called AlphaEarth. Think of this as a super-smart camera that has looked at the entire Earth from space for years. It doesn't just take a picture; it creates a "digital fingerprint" for every spot on the ground.

This fingerprint captures everything at once:

What the land looks like (topography).
What plants are growing there.
How wet the soil is.
What kind of rocks are likely underneath.
How humans have changed the land.

Instead of just looking at "steepness," AlphaEarth looks at the whole story of the landscape.

The Experiment: Two Different Tests

The researchers tested this "digital fingerprint" in two very different ways to see if it could predict what's underground.

Test 1: The Earthquake Shake (VS30)

The Goal: Predict how stiff the ground is near the surface.
The Analogy: Imagine trying to guess if a floor is made of solid wood or soft carpet just by looking at the room.
The Result:

Old Way: Just looking at how steep the roof (or hill) is. This works okay, but it's not perfect.
New Way: Using the AlphaEarth "fingerprint" plus the steepness.
Outcome: The new way was much better. It's like adding a magnifying glass to your eyes. The fingerprint helped the computer understand that a steep hill might be solid rock (good for earthquakes) or loose sand (bad for earthquakes), something the old "steepness only" method missed.
Limitation: If you take a model trained in California and try to use it in Washington without any local data, it gets a bit confused. It's good at the general patterns, but it struggles with the tiny, local details.

Test 2: The Deep Heat (Temperature)

The Goal: Predict how hot the ground is 1,000 to 6,000 meters down.
The Analogy: Imagine trying to guess how hot the oven is inside a house just by looking at the outside walls and the garden.
The Result:

Old Way: There wasn't really a good "old way" for this. Traditional methods struggled because the heat deep down is influenced by complex things like ancient tectonic plates and underground water flow, which are hard to summarize with simple rules.
New Way: Using only the AlphaEarth "fingerprint."
Outcome: This worked surprisingly well. The computer learned that certain "fingerprint" patterns (like rough, volcanic-looking landscapes) usually mean it's hot deep down, while smooth, ancient, flat landscapes usually mean it's cooler.
Outcome: It successfully mapped out the "big picture" of heat across the whole country. It correctly identified that the western US is generally hotter (due to tectonic activity) and the middle/east is cooler. It didn't get the exact temperature for every single drill hole, but it got the regional patterns right.

The Big Takeaway

The paper concludes that this "digital fingerprint" (AlphaEarth) is a powerful tool, but it plays different roles depending on what you are trying to find:

For Earthquakes (Shallow Ground): It acts as a helper. It works best when you combine it with the old, trusted rules (like "steepness"). It fills in the gaps the old rules miss.
For Deep Heat: It acts as the main detective. Since there are no simple rules for deep heat, the fingerprint becomes the primary way the computer understands the landscape.

The Catch

The researchers were honest about the limits.

It's not magic: If you train the computer on data from California and then ask it to predict the ground in a totally different geological area (like the middle of the country) without any local data, it gets less accurate.
It sees the forest, not the trees: It is excellent at seeing big regional patterns (like "this whole mountain range is hot" or "this whole valley is soft soil"), but it can't perfectly predict the exact conditions of a specific backyard without local measurements.

In short: AlphaEarth gives scientists a new, super-powered pair of glasses. It doesn't replace the need for drilling holes, but it helps them make much smarter guesses about what's underground based on what they can see from the surface.

Technical Summary: Surface-to-subsurface inference via AlphaEarth

Problem Statement
Inferring subsurface properties from surface observations is a fundamental challenge in geoscience due to the sparsity, cost, and uneven distribution of direct subsurface measurements (e.g., boreholes, seismic surveys). Conversely, surface observations are dense and temporally repeated. While surface expressions (geomorphology, lithology, vegetation, moisture) often correlate with subsurface conditions, these links are indirect, non-unique, and scale-dependent. Traditional mapping often relies on hand-crafted proxies (e.g., topographic slope for seismic velocity) which may fail to capture complex, coupled landscape contexts. This study investigates whether a general-purpose Earth observation foundation model embedding can serve as a shared, transferable geospatial representation to improve surface-to-subsurface inference across distinct physical domains.

Methodology
The authors utilize AlphaEarth, a geospatial foundation model developed by Google DeepMind. AlphaEarth produces a global 64-dimensional embedding field at 10 m resolution, learned from multi-source Earth observation data (optical imagery, radar, topography, LiDAR, climate reanalysis, and gravity-derived hydrologic signals). The model is not retrained; instead, fixed embedding vectors are used as input features for downstream task-specific regression models.

The study evaluates this approach across two distinct applications in the conterminous United States (CONUS):

Shallow Seismic Velocity ( $V_{S30}$ ): Predicting the time-averaged shear-wave velocity in the top 30 meters.
- Features: AlphaEarth embeddings used alone, and in hybrid sets with conventional predictors (log-slope and a binary tectonic status indicator).
- Models: Linear Regression, Random Forest (RF), and XGBoost.
- Validation: Random 80/20 train-test splits and geographically withheld "leave-one-state-out" extrapolation tests (California, Oregon, Washington).
Subsurface Temperature: Reconstructing temperature fields at depths of 1,000–6,000 meters.
- Features: AlphaEarth embeddings as the primary geospatial descriptor (no equivalent standard low-dimensional proxy set exists for deep temperature).
- Models: A Multilayer Perceptron (MLP) neural network.
- Validation: Random splits and geographically withheld tests for Arizona, Missouri, and New York.

Key Results

$V_{S30}$ Prediction:
- Hybrid models combining AlphaEarth embeddings with log-slope significantly outperformed conventional baselines. The best model (XGBoost with embeddings + log-slope) achieved a test RMSE of 126.1 m/s and $R^2$ of 0.53, a 24.3% reduction in RMSE compared to a baseline using only log-slope and tectonic status.
- Embeddings alone performed better than the conventional terrain proxy baseline, indicating they encode richer information regarding lithology, vegetation, and hydrologic organization.
- Extrapolation: In leave-one-state-out tests, predictive skill decreased significantly (e.g., $R^2_{log}$ dropped to ~0.06 for California), reflecting the local nature of $V_{S30}$ and the influence of specific sediment thickness and stratigraphy not captured by regional surface patterns. However, the models still recovered physically plausible first-order regional patterns (e.g., basin vs. upland contrasts).
Subsurface Temperature:
- The embedding-informed neural network achieved strong performance ( $R^2 = 0.919$ , RMSE = 6.0°C) compared to a non-embedding nonlinear baseline, demonstrating that the embedding captures complementary information beyond model flexibility.
- The model successfully recovered broad thermotectonic structures, distinguishing between the active, warmer western US (Basin and Range, Yellowstone) and the stable, cooler cratonic interior.
- Extrapolation: State-level extrapolation tests showed variable skill depending on the region's thermal heterogeneity (e.g., higher error in geologically complex Arizona vs. more coherent New York). The model preserved broad thermal gradients but smoothed local anomalies, consistent with the expectation that deep temperature is more regionally organized than shallow velocity.

Key Contributions

Shared Representation: The study demonstrates that a single, general-purpose foundation model embedding can serve as a transferable surface representation for diverse subsurface targets (mechanical stiffness and thermal state) without task-specific retraining of the embedding itself.
Role of Embeddings: The research clarifies the conditional utility of such embeddings:
- For $V_{S30}$ , where strong physics-guided proxies (slope) exist, the embedding acts as an augmenting feature, adding non-redundant landscape context to improve accuracy.
- For subsurface temperature, where controls are diffuse and lack simple conventional proxies, the embedding serves as the primary geospatial descriptor.
Validation Framework: By pairing random held-out tests with geographically withheld extrapolation tests, the study distinguishes between predictive skill within sampled regions and the limits of spatial transferability. This highlights that while embeddings support regional inference, extrapolation performance is constrained by sampling density, geologic heterogeneity, and the scale of the target property.

Significance and Claims
The paper claims that foundation-model surface embeddings are not merely performance-boosting features but act as flexible, shared representations that encode coupled geomorphic, geologic, hydrologic, and land-use contexts. These contexts provide indirect but statistically significant constraints on subsurface structure.

The authors position this work as a study of how transferable representations encode geologically meaningful context, rather than a definitive solution for all subsurface mapping. They emphasize that the value of the representation depends on the target physics, the prediction scale, and the strength of existing domain knowledge. The results suggest that such embeddings can bridge Earth observation and subsurface science for regional screening and first-pass mapping, particularly in data-sparse regions, but they do not replace the need for local calibration or site-specific investigations for engineering design. The study concludes that these embeddings are most valuable when used to study how landscape-scale surface signals constrain the subsurface, rather than as a direct physical measurement of subsurface properties.