From the article:

“Accurately estimating future flood risk at every property in a single city or watershed—let alone the entire United States—is fundamentally not possible given current knowledge,” said James Doss-Gollin, an assistant professor of engineering and a climate-risks specialist at Rice University in Houston.

While flood risk models serve an important role in informing homebuyers and leveling the playing field between buyers, lenders, and insurers, we must be skeptical of claims that national, building-scale flood risk models are empirically “validated.” Probabilistic flood risk estimation involves not just flood modeling, but also estimates of the probability of extreme weather events based on imperfect models and short records—which cannot be validated in the traditional sense.

For more on this topic, see our recent PNAS paper led by Adam Pollack on how uncertainty in flood risk estimates affects flood insurance policy.

Most engineering products can be tested directly: a new material can be stress-tested regardless of whether its formula is secret. Climate risk estimates can’t. There’s no instrument that measures the probability of a 100-year flood, so when two models disagree—as they do for 76% of Los Angeles properties—we have no independent way to settle the question. Scrutiny of methods and assumptions is the only path to trust, yet only 4% of highly-cited climate-risk studies share both data and code.

The paper charts a course forward: climate-risk tools should be explainable, benchmarked, built on open foundations, and comparable across providers. Getting there will take more than good intentions from individual researchers—it requires institutional commitment to open science infrastructure.

See the full paper or the Dartmouth press release.

The talk surveyed the lab’s recent work on probabilistic models for nonstationary precipitation extremes — including the Bayesian spatiotemporal framework led by Yuchen Lu — and discussed open methodological challenges in extreme value theory, regionalization, and integrating climate-model projections with short observational records.

Thanks to the Rice Statistics department for the invitation and for a great discussion afterward.

Many assessments of rainfall and flood hazard require high-resolution precipitation data (e.g., 1-4 km). While coarse global datasets from reanalysis products exist, and fine-grained radar data is available for recent years, there is a critical gap in producing long-term, high-resolution historical datasets needed for robust climate and hydrological modeling. Existing deep learning methods for this downscaling or super-resolution task often fail to accurately capture the statistical properties of the most extreme, and therefore most dangerous, events.

We developed Wasserstein Regularized Diffusion (WassDiff), a novel generative AI framework. Unlike other diffusion models that can replicate the general distribution of rainfall, WassDiff is specifically designed to get the tails of the distribution right. It integrates a Wasserstein distribution-matching regularizer directly into the denoising process. This forces the model to pay close attention to the frequency and magnitude of extreme values, reducing the biases commonly seen in generative models for climate data.

Experiments show that WassDiff generates high-resolution precipitation fields that are more physically realistic and statistically accurate than those from competing models. It more faithfully reproduces the spatial patterns and, most importantly, the statistical characteristics of extreme rainfall, making it a more reliable tool for climate impact studies.

This work provides a significant step forward for both the machine learning and climate science communities. By creating more reliable high-resolution precipitation data, WassDiff can help scientists and engineers better assess flood risk, design more resilient infrastructure, and ultimately improve our understanding of how climate change impacts local communities.

The Consortium for Enhancing Resilience and Catastrophe Modeling held its inaugural Industry Advisory Board Meeting on August 7-8, 2025, where our Industry Advisory Board played a critical role in providing project feedback and helping prioritize key projects for the consortium’s first year.

This collaborative project aims to develop innovative approaches for assessing tropical cyclone hazards under changing climate conditions, addressing the critical need for nonstationary methods that can account for evolving risk patterns.

We’re excited to be part of this collaborative effort alongside other groundbreaking projects focused on multi-hazard resilience, wildfire fragility curves, and AI-driven damage assessment.

For more information about CERCAT and the full portfolio of funded projects, visit the CERCAT website.

This research represents the first step in Yuchen’s work, funded by Rice and the Texas Water Development Board, on incorporating robust estimates of nonstationarity into precipitation frequency estimates. She is currently working with colleagues at TAMU and TWDB to incorporate these insights and methods into Intensity-Duration-Frequency (IDF) curves for Texas, and we will be presenting updates at the upcoming AGU Fall meeting.

• Historic Texas Flooding — KEYE-AUS (CBS), July 25, 2025
• Houston’s Morning Show — Fox26 Houston, July 10, 2025
• Breaking down the force of water in the Texas floods — AP News (Michael Phillis), July 10, 2025
• Here are some things you can do to be better prepared for major flooding — Associated Press (Caleigh Wells), July 10, 2025
• ABC13 interview on flooding in Houston vs. Central Texas — ABC13 (Luke Jones), July 9, 2025
• After deadly flooding in Central Texas, state lawmakers look to prevent similar tragedies — KVUE (Daniel Perreault), July 9, 2025
• New Flood Warning System Greenlit Shortly Before Deadly Texas Disaster — The Wall Street Journal (Joseph De Avila), July 9, 2025
• Questions arise on how emergency warning systems work after Central Texas flood — ABC13 (Tom Abrahams), July 8, 2025

For James’s longer-form perspective on the event, see the original post.

This was, first and foremost, a horrific tragedy. The pain of families who can’t locate their loved ones is unfathomable. The current focus is correctly on finding anyone who may still be alive, on supporting grieving families and communities, and on providing aid to those who have lost everything. However, this event also highlights systemic failures that must be addressed to prevent future tragedies.

1. Building in the floodway is dangerous. The floodway is the area where water flows fastest and deepest during a flood. It appears that Camp Mystic was built in the floodway of the Guadalupe River, which is a profoundly dangerous place to build a camp for children. In most communities, likely including Kerr County (I have not verified local regulations beyond scattered reporting), new construction or major renovations would trigger some scrutiny of flood risk, but building codes are not often applied retroactively to existing structures. While it is socially and financially costly to abandon or existing structures, safety must come first.

2. Need for flood alert and response systems. The SSPEED Center, of which I am a member, has built real-time flood forecasts for Houston, and we (along with other groups) have the capability to build a similar system for the Guadalupe River, given time, resources, and collaborators. As my colleague Phil Bedient has eloquently argued, we need to build a real-time flood forecasting system for the Guadalupe River.

   • A critical component of this is enhancing real-time data collection, namely stream and rainfall gauges, to monitor floods in real time. The Upper Guadalupe River has five stream gauges along its roughly 230 mile length, and some of the gauges were not functioning during the flood. Better data collection is essential to understanding and responding to floods.
   • The weather forecasts and warnings were generally good. The National Weather Service issued appropriate warnings as sources below describe in detail. However, predicting the exact magnitude and location of extreme rainfall is inherently uncertain, especially in complex terrain like the Texas Hill Country. The National Weather Service correctly identified the risk of heavy rainfall and flash flooding, and issued appropriate and timely warnings, with certainty and accuracy increasing as the event approached.
   • Alerts without trust and plans do not save lives. People need to get the alerts, understand what they mean, and know what steps to take. Building a flood prediction model is a necessary, but not sufficient, step to saving
   • There is a trade-off between certainty and lead time. Individuals and communities can take actions to prepare for flash floods, but these actions require time to implement. Unfortunately, the hydrology of the Guadalupe River means that the time from rain to flood can be very short. This means that if you wait until the rain has already fallen, or worse until the rivers are already rising, you can only take a very limited set of actions. On the other hand, if you take action based on forecasts, you have more time to prepare, but the forecasts are inherently uncertain and you will sometimes take action that turns out to be unnecessary.

I am not an expert on disaster response, communication, or the many agencies involved in flood response in this region, so I am in no position to point fingers or assign blame. However, it is unacceptable that an advanced society can place so little value on human life that we allow events like this to occur. There must be constructive accountability to ensure that we avoid preventable tragedies like this in the future.

You can watch the segment here.

Dongwook joined the lab in Fall 2024 and is co-advised with Avi Gori. His research uses synthetic aperture radar (SAR) and deep learning to map flood extents in urban and rural environments, contributing to ongoing work on flood hazard characterization in rural Texas.

Well-deserved, Dongwook!

Yuchen presented her research with confidence and expertise, demonstrating a deep understanding of nonstationary hydroclimate extremes and her innovative approaches to probabilistic analysis. Her work promises to make significant contributions to our understanding of changing extreme weather patterns and their implications for infrastructure design and risk management. Her first paper is already under review (Lu et al. 2025), and future work is forthcoming.

We would like to express our sincere gratitude to her committee members, Dr. Phil Bedient, Dr. Katherine Ensor, and Dr. Avantika Gori, for their thoughtful feedback and guidance throughout this process.

We are excited to invite you to the Texas Climate Conference 2025, a 1.5-day event at Rice University on April 10-11, 2025. This conference will bring together scientists, policymakers, and community stakeholders to discuss science and solutions on four critical Texas climate themes:

• Extreme Heat
• Flooding and Water Resources
• Sea Level Rise and Coastal Hazards
• Human Health

The project aims to create a performance-based framework that bridges critical gaps in flood hazard assessment and mitigation strategies for rural communities. Working in the rural communities of Fort Hancock and Premont, Texas, we will develop new methods to characterize flood risks and co-design mitigation strategies that reflect local priorities. We’re particularly focused on identifying adaptations that address both frequent and extreme flood events, and that are scalable to communities with similar characteristics.

For more details, see the press release from IDRT or the NSF award).

The initiative has since evolved into the AI for Urban Resilience cluster at the Ken Kennedy Institute.

Given a geographical region and precipitation data, our model predicts water depths in an auto-regressive fashion using a message-passing framework inspired by the conservation of momentum and mass as expressed in the shallow-water equations. This allows the model to effectively capture the propagation of water flow, particularly during the early stages of flooding when the water is scarce.

Our empirical results, based on a dataset covering 9 regions and 7 historical precipitation events, demonstrate that our model outperforms existing baselines, providing accurate simulations even in complex topographies. This solution achieves accurate results using real ground elevation data. Additionally, our model excels in predicting flood dynamics for unseen storms and watersheds, showcasing its robustness and generalization capabilities.

For more details, see here. This is an exciting space, and we are looking forward to sharing more updates on our ongoing research in this critical area!