Real-time probabilistic tsunami forecasting in Cascadia from sparse offshore pressure observations

This paper demonstrates that a hypothetical network of 175 seafloor pressure sensors, combined with a Bayesian inversion framework utilizing offline precomputation and online data assimilation, can enable real-time probabilistic tsunami forecasting for Cascadia earthquakes with high accuracy and sub-second latency despite sparse offshore observations.

The Gist

Imagine the Cascadia Subduction Zone as a massive, sleeping giant lying beneath the Pacific Ocean off the coast of the Pacific Northwest. Every few hundred years, this giant wakes up, stretches, and causes a massive earthquake that can trigger a tsunami. The problem is, we don't know exactly how the giant will wake up. Will it stretch just a little (a "partial rupture"), or will it stretch all the way from Canada to California (a "margin-wide rupture")?

This uncertainty makes predicting the tsunami incredibly hard. If we guess wrong about the size of the earthquake, our warning might be too weak (leaving people unprepared) or too strong (causing panic).

This paper presents a brilliant new "digital twin" system designed to solve this puzzle in real-time, using a hypothetical network of underwater microphones. Here is the breakdown in simple terms:

1. The Problem: The "Silent" Ocean

Usually, when an earthquake happens, we rely on sensors on land or buoys far out at sea. But by the time a buoy far away feels the wave, the tsunami is already on its way to the coast. We need to know what's happening right next to the earthquake, but the ocean is too big and too deep to put sensors everywhere. It's like trying to guess the shape of a giant, invisible cloud by only looking at a few raindrops.

2. The Solution: A "Digital Twin" of the Ocean

The researchers built a super-advanced computer model (a "digital twin") that simulates exactly how the ocean and the Earth's crust interact during an earthquake. They didn't just model the shaking ground; they modeled the water too.

Think of it like this:

• The Earthquake: When the giant stretches, it pushes the ocean floor up.
• The Sound: This push creates a loud "thump" in the water (acoustic waves) that travels super fast, like a shout.
• The Wave: It also creates a slow-moving, heavy wave (the tsunami) that travels like a slow-moving freight train.

The cool thing is that for the first two minutes, the "shout" (sound) and the "freight train" (tsunami) look almost identical, whether the earthquake is small or huge. But after two minutes, they start to look different. The system is designed to catch that split-second difference.

3. The Magic Trick: The "Offline" vs. "Online" Brain

The hardest part of this math is that it usually takes a supercomputer days to solve the equations for a tsunami. But you can't wait days for a warning; you need it in seconds.

The authors used a clever two-step trick:

• Step 1: The "Offline" Homework (Done beforehand): Before any earthquake ever happens, they use massive supercomputers to run millions of simulations. They calculate every possible way the ocean could react to an earthquake. They store these answers in a giant library. This takes a lot of time and power, but it only needs to be done once.
• Step 2: The "Online" Reaction (Done in real-time): When an earthquake actually happens, the system doesn't do the hard math again. Instead, it looks at the data coming in from the sensors, opens the "library" of pre-calculated answers, and finds the match.

The Analogy: Imagine you are a chef.

• The Old Way: Every time a customer orders soup, you have to grow the vegetables, chop them, and cook the broth from scratch. It takes hours.
• The New Way: You pre-cook thousands of different soups and freeze them (the "Offline" phase). When a customer orders, you just grab the right frozen soup, heat it up, and serve it in seconds (the "Online" phase).

4. The Sensor Network: The "S-net" Idea

The researchers tested this system using a hypothetical network of 175 underwater pressure sensors (like the real S-net network in Japan). They asked: "If we had this many sensors, could we tell the difference between a small and a huge earthquake?"

The Result: Yes! Even with only 175 sensors (which is sparse compared to the size of the ocean), the system could:

1. Figure out the earthquake: It could tell if the giant stretched a little or a lot.
2. Predict the wave: It could forecast the height of the tsunami at specific coastal towns.
3. Do it fast: It took less than one second to give the forecast.
4. Be accurate: The predictions were off by only about 20%, which is incredibly good for such a chaotic event.

5. Why This Matters

This isn't just about math; it's about saving lives.

• Speed: Because the hard work is done beforehand, the system can run on a simple laptop in an emergency center.
• Precision: It can distinguish between a "partial" earthquake (which might only hit Vancouver) and a "full" one (which hits the whole coast), allowing for more targeted warnings.
• Uncertainty: Instead of giving a single number like "The wave will be 3 meters," it gives a probability: "There is a 95% chance the wave will be between 2 and 4 meters." This helps emergency managers make better decisions.

The Bottom Line

This paper proves that if we build a network of underwater sensors in the Pacific Northwest, we can use a "pre-calculated" digital brain to listen to the ocean, instantly figure out how big the earthquake is, and tell coastal towns exactly how big the tsunami will be—all in the blink of an eye. It turns a chaotic, terrifying natural event into a manageable, predictable one.

Technical Summary

1. Problem Statement

The Cascadia Subduction Zone (CSZ) poses a significant tsunami threat, yet near-field early warning is severely limited by the sparsity of offshore observations. Current systems rely heavily on land-based seismic data and sparse Deep-ocean Assessment and Reporting of Tsunamis (DART) buoys, which often provide insufficient data to constrain tsunami generation close to the source.

• Uncertainty: The extent of future CSZ megathrust earthquakes (partial rupture vs. margin-wide rupture) is uncertain, making it difficult to pre-select the correct scenario for forecasting.
• Computational Challenge: Real-time forecasting requires solving complex, fully-coupled physics models (solid Earth rupture + ocean acoustic/gravity waves) within seconds. High-resolution 3D simulations are computationally prohibitive for real-time use.
• Data Scarcity: Existing methods often rely on low-dimensional parameterizations (e.g., moment magnitude, hypocenter) or precomputed databases, which may not capture the full spatiotemporal complexity of the seafloor motion required for accurate near-field tsunami prediction.

2. Methodology

The authors propose a physics-based Bayesian inversion framework that separates computation into an offline precomputation phase and an online real-time inference phase.

A. Forward Modeling: Fully-Coupled Simulations

• Scenarios: Two fully-coupled dynamic rupture–tsunami scenarios were simulated using the open-source code SeisSol:
  1. Partial Rupture: Magnitude 8.6, arresting at ~46°N.
  2. Margin-wide Rupture: Magnitude 8.7, arresting at ~42°N.
• Physics: The models extend previous solid-earth-only simulations by adding a 3D acoustic ocean layer. This captures the interaction between the solid Earth and the water column, including:
  • Seismic P-waves and S-waves.
  • Oceanic Rayleigh waves.
  • Acoustic waves (compressible ocean).
  • Tsunami gravity waves (incompressible limit).
• Data Generation: Synthetic seafloor pressure data were generated at 1 Hz sampling rates for hypothetical sensor networks (600 sensors and a sparse 175-sensor network).

B. Inversion Framework: Acoustic-Gravity Model

Instead of inverting the full coupled system in real-time, the authors use a simplified acoustic-gravity wave model for the inversion.

• Linear Time-Invariant (LTI) Structure: The governing equations are linearized around a hydrostatic background, allowing the system to be treated as LTI.
• Offline/Online Decomposition:
  • Offline: Computationally expensive adjoint PDE solutions are precomputed for each sensor and forecast location using high-order finite element methods (MFEM library) on supercomputers (512 GPUs, ~1 hour per solution).
  • Online: Real-time inference uses Fast Fourier Transform (FFT)-based mappings derived from the offline data. This allows the inversion to run on a standard laptop in less than one second.
• Bayesian Approach: The framework treats the seafloor velocity field as a random field. It computes the posterior probability distribution of the seafloor motion given the observed pressure data, rather than a single deterministic solution. This provides quantified uncertainty estimates (credible intervals) for both the inferred motion and the resulting tsunami forecasts.

3. Key Contributions

1. Real-Time Probabilistic Forecasting: Demonstrated a framework capable of inferring full spatiotemporal seafloor velocity fields and forecasting tsunami wave heights in <1 second, even with sparse data.
2. Sparse Network Feasibility: Showed that a hypothetical network of 175 offshore pressure sensors (comparable to Japan's S-net) is sufficient to achieve accurate probabilistic forecasts, reducing the required infrastructure significantly compared to previous idealized studies using 600 sensors.
3. Scenario Discrimination: Proved that the system can distinguish between partial and margin-wide rupture scenarios within minutes of rupture onset by analyzing the divergence in seismo-acoustic wavefields (specifically the attenuation of oceanic Rayleigh waves in the partial rupture case).
4. Physics Consistency: Unlike methods that assume static slip or simplified source geometries, this approach inverts directly for the full spatiotemporal seafloor velocity field without assuming a specific fault geometry, capturing complex near-field physics.

4. Results

• Forecast Accuracy:
  • Margin-wide Rupture: Tsunami forecast error was 22.1% with 175 sensors (vs. 18.6% with 600 sensors).
  • Partial Rupture: Tsunami forecast error was 19.6% with 175 sensors (vs. 18.1% with 600 sensors).
  • Observation: The degradation in forecast accuracy when moving from 600 to 175 sensors is modest, indicating robustness to sensor sparsity.
• Seafloor Motion Inference:
  • While the tsunami forecasts remained highly accurate, the inference of the full seafloor displacement field had higher errors (36.1% for margin-wide, 31.7% for partial with 175 sensors).
  • Interpretation: Sparse data cannot resolve fine-scale spatial features of the uplift, but it successfully recovers the "smooth," tsunami-relevant components of the motion that drive gravity waves.
• Wavefield Evolution:
  • For the first two minutes, acoustic, Rayleigh, and tsunami wavefields are similar for both rupture scenarios.
  • After rupture arrest (approx. 2 minutes), the partial rupture scenario shows rapid attenuation of oceanic Rayleigh waves, while the margin-wide rupture continues to generate strong signals. This divergence allows the inversion to distinguish the scenarios.
• Uncertainty Quantification: The Bayesian framework successfully provided 95% credible intervals around forecasts, which widened appropriately as sensor density decreased, accurately reflecting the increased uncertainty.

5. Significance

• Operational Viability: This study provides a concrete pathway for implementing real-time, probabilistic tsunami early warning in Cascadia. By shifting heavy computation to an offline stage, the system becomes deployable on modest hardware (e.g., laptops) during an emergency.
• Sensor Deployment Strategy: The results justify the deployment of a ~175-sensor offshore network (similar to Japan's S-net) as a cost-effective solution for near-field tsunami warning, rather than requiring denser, more expensive arrays.
• Beyond Deterministic Models: By providing probabilistic forecasts with uncertainty quantification, the system offers decision-makers a more nuanced view of risk, crucial for evacuation planning where false alarms or missed warnings have high societal costs.
• Generalizability: The methodology (offline precomputation + online FFT-based Bayesian inference) is applicable to other subduction zones and could be extended to other tsunami sources (e.g., landslides, volcanic explosions) since the inversion targets seafloor motion rather than specific fault mechanics.