Inspiration
The devastating 2023 earthquake in Afghanistan was a sobering reminder of a painful truth: it is rarely the earthquake itself that kills, it is the buildings. Despite the quake registering a relatively moderate magnitude, thousands of lives were lost because the structures people called home simply weren't built to withstand the shaking. We also live along an active fault line and have experienced a multitude of a earthquakes ourselves. For us, Earthquake safety is something that's both deeply personal and globally relevant, and that's why we decided to start Seismo.
What It Does
Seismo lets you click anywhere on a map, place a building, and watch what happens when the earth moves beneath it. Using real geographic and seismic data pulled for that specific location, we simulate how an earthquake would propagate through the terrain and interact with the structure. From there, a Reinforcement Learning agent analyzes the simulation and automatically optimizes the building's design, reinforcing the right joints, adjusting materials, redistributing load, all while keeping cost at the forefront. The result is a structurally safer building, tailored to its environment, at the lowest viable price.
How We Built It
We started with the physics. Before anything else, we needed confidence that our earthquake simulator was grounded in reality, so we spent considerable time validating our model against known seismic events and cross-referencing with geological data. Accuracy wasn't optional — a beautiful tool built on a flawed simulation would be worse than useless. Once the simulator was solid, we tackled the rendering pipeline: visualizing the building geometry, layering in real terrain data, and animating the ground motion in a way that felt physically believable. Getting the shaking to look right — not just move, but move the way an earthquake actually moves — turned out to be one of the hardest parts of the project. With the simulation and rendering in place, we turned to the intelligence layer. We implemented an actor-critic reinforcement learning algorithm to serve as our structural optimizer. The actor proposes design modifications; the critic evaluates their long-term impact on both safety and cost. Through thousands of simulated earthquake scenarios, the agent learns which structural choices matter most for a given location.
Challenges We Ran Into
The RL training loop was where we hit the most friction. Reward hacking, where the agent finds clever but meaningless shortcuts to maximize its score, was a persistent problem. Solving it required careful, iterative design of the reward function, balancing structural integrity metrics against cost penalties in a way that closed off unintended loopholes. Rendering the earthquake shaking realistically was its own beast. Earthquake motion is complex: it involves multiple wave types (PP P-waves, SS S-waves, surface waves) arriving at different times and frequencies. Capturing even an approximation of that in a real-time visualization required significant experimentation.
Finally, sourcing accurate terrain data and computing meaningful, location-specific seismic risk, accounting for soil composition, fault proximity, and historical activity, proved to be a substantial data engineering challenge in its own right.
Accomplishments We're Proud Of
We built something that works, and works intuitively. Anyone can drop a building onto a map, run a simulation, and receive an optimized structural recommendation, no seismology PhD required. We're proud that the underlying science is rigorous: the simulations are physically grounded, and the RL agent uses state-of-the-art techniques to produce recommendations that are genuinely better and genuinely cheaper. Making those two goals coexist was hard. We're glad we didn't compromise on either.
What We Learned
We came in knowing reinforcement learning in theory. We left knowing it in practice — including all the messy, unintuitive ways real training runs can go wrong. On the geoscience side, we developed a deep appreciation for how seismic waves propagate, how soil amplification can turn a moderate quake into a catastrophic one, and how relatively small architectural decisions, the right shear wall here, the right foundation depth there, can be the difference between a building that stands and one that doesn't.
What's Next for Seismo
We plan to open-source Seismo and build integrations with existing structural engineering tools, so it can slot into real workflows rather than exist as a standalone prototype. We also plan on making it available for significantly more locations. The goal is to make earthquake viability testing fast, cheap, and accessible, especially for the communities that need it most.
Log in or sign up for Devpost to join the conversation.