Sensor Placement for Tsunami Early Warning via Large-Scale Bayesian Optimal Experimental Design
New multi-GPU algorithm places 175 sensors to forecast tsunamis with unprecedented accuracy.
A team of researchers has published a groundbreaking paper detailing a new computational framework that uses high-performance AI to design optimal sensor networks for tsunami early warning. The work, led by Sreeram Venkat, Stefan Henneking, and Omar Ghattas, tackles the exceptionally challenging problem of Bayesian Optimal Experimental Design (OED) for systems governed by complex physics, like tsunami propagation. Their key innovation is a reformulation that transforms the OED problem into a dense matrix subset selection task, making it computationally tractable at an unprecedented scale.
They developed a novel, multi-GPU algorithm based on Schur-complement updates that uses a pipelined, greedy approach. This method fully overlaps input/output operations with GPU computations, allowing it to achieve near-perfect weak and strong scaling across hundreds of GPUs on supercomputers like Perlmutter and Frontier. The framework's power was demonstrated by applying it to a state-of-the-art digital twin of the Cascadia Subduction Zone—a model that previously won the 2025 Gordon Bell Prize.
In this real-world application, the AI framework successfully optimized the placement of 175 sensors. Its goal was to minimize the predictive uncertainty for a parameter field with over one billion degrees of freedom, which is critical for accurately inferring seismic sources and seafloor uplift that cause tsunamis. This represents a massive leap in our ability to strategically design observation networks that maximize information gain from limited, expensive sensor deployments in the ocean.
- Scalable AI framework solves Bayesian OED for hyperbolic PDE systems, a previously intractable problem for tsunami forecasting.
- Multi-GPU algorithm achieves near-perfect scaling on hundreds of GPUs, optimizing a 175-sensor network for a billion-parameter model.
- Successfully applied to the Gordon Bell-winning Cascadia Subduction Zone digital twin, directly enhancing real-world early warning capabilities.
Why It Matters
This AI-driven approach enables more effective, lifesaving tsunami warning systems by optimally placing expensive ocean sensors to maximize forecast accuracy.