Discover more from Startup Pirate by Alex Alexakis
AI for Flood Resilience
Natural hazards and physics-based neural networks, giving robots the sense of touch, from founder to investor, jobs, events, and more
Friends, this is Startup Pirate issue #84, and today, we touch on a timely and absolutely important topic for many: Floods. We look at it from a technology lens and how AI can help communities become more resilient against them. Right after, you can read the usual roundup of Greek Tech funding rounds, jobs, new products, and events. And don’t forget to subscribe here:
AI for Flood Resilience
The most frequent and deadly natural hazards worldwide are floods. Severe, devastating floods have recently posed significant challenges for communities worldwide: Greece, US, Spain, Turkey, India, and more, causing extensive loss of life and property. Enter REOR20, a startup that brings together AI and physics to help individuals, governments, and insurers understand flood risk and, for the first time, protect against floods based on accurate and complete data. Danny Chatziprodromou, CEO & co-founder, joins us to discuss the limitations of traditional flood simulation models, physics-based neural networks, and more.
Let’s get to it.
Danny, great to have you on the newsletter to discuss how technology can help more people stay safe from floods.
DC: Thanks for having me, Alex. Looking forward to our discussion.
How do we know the risk of flood in a particular area today? And how is this information used for civil protection?
DC: Floods occur in three different ways: fluvial (rivers and streams break out of their banks and water flows out onto the adjacent low-lying areas), pluvial (extreme rainfall event creates a flood independent of an overflowing water body — flash floods are usually falling under this category), and coastal (inundation of land areas along the coast by seawater due to intense windstorm events happening at the same time as high tide and tsunamis). In many countries, we have static flood maps showing a community’s flood zone, floodplain boundaries, and base flood elevation using data from aerial images (drones, airplanes), satellites or topography. Civil protection agencies use these maps to understand and mitigate flood risks and, as new information arises e.g. weather forecasts, to determine the potential risk of an event (e.g. rainstorm) and, if necessary, take actions to alleviate any effects.
Flood maps are pre-calculated based on computational fluid dynamics, which is the process of mathematically describing the physics of fluid flow. This advanced scientific field has existed for over half a century and focuses on numerically approximating a set of equations called Navier–Stokes that describe how the velocity, pressure, temperature, and density of a moving fluid are related. We use computers to perform such calculations and simulate flooding events. Now, the problem is when you do that for a very large area, say at a municipality level, and include many details (drainage systems, protective infrastructure, buildings, etc.) in high resolution, it takes A LOT of computational time and cost. It could take months on supercomputers.
Hence, we have to “water down” the mathematics (pun intended) behind the calculations or simplify the equations, and this leads to inaccuracies in the output of the models, e.g. maximum depth around the building and, most importantly, incomplete information since the simplified models cannot provide estimates for the momentum of the water i.e. dynamic force and the duration that the properties will be under water. It is very different to be in a situation where the water accumulates slowly up to, let’s say, 1m and then dries off fast or being in the wake of a 1m tall wave that leads to stagnated water for a long period. The status quo used for understanding the flood risk does perceive those two situations as equally risky. Unfortunately, this is not good enough for anyone who wants to make informed decisions, especially in the context of civil protection.
We have had many recent flood events globally, so this is quite timely. Can technological advances mitigate the limitations of traditional simulation models?
DC: Artificial intelligence nowadays lets us process vast amounts of data, such as weather observations, hydrological data, and remote sensing imagery, to extract patterns and relationships that can improve the accuracy of the inputs in flood models. And this is where the technology we build at REOR20 comes in. Our team brings years of research in computational fluid dynamics from ETH Zurich, MIT and other prominent institutions, along with seasoned insurance business experts. We had experienced first-hand the uncertainty traditional hazard flood models were introducing and how it would be impossible for civil protection applications to properly use them. It all started in 2019 with a crazy idea to train AI to understand physics. The field is still so new that there is very little to no academic literature to refer to.
Fast forward four years later, we have built physics-based neural network technology, which, in essence, is artificial intelligence that understands how to do physics. This is a new way of carrying out scientific research, emphasising learning from data and now learning from physics. By incorporating physical principles into machine learning, we can create much more powerful models that learn from data and build upon existing scientific knowledge, enabling the algorithm to instantaneously identify the right solution. As a result, our technology can use very high-fidelity ground data from the landscape and buildings (how elevated the entrance is, the existence of backflow preventer valves, etc.) even down to centimeters and at a country level, and can take into account real-time information from the event, including meteorology, as it unfolds. This is unprecedented if you think we make such calculations in seconds, 100,000 times faster than traditional flood modeling, at a significantly lower cost.
This is a leap forward for studying a number of physical phenomena and, consequently, how we protect against natural hazards.
DC: Absolutely. The basis of the technology we build can be used to study other physical phenomena with much higher accuracy and lower cost. As you said, this is a leap forward, not an incremental improvement. Whatever phenomenon can be described theoretically through a set of partial differential equations (PDEs) can be attacked by AI in the same way that we do for floods. These are equations made up of functions and their derivatives, which are ubiquitous in mathematically oriented scientific fields, such as physics and engineering, from studying wind and atmospheric pollution to financial markets, etc.
Everyone can benefit from this technology, but since it touches civil protection, who is the right user for it?
DC: This information should be readily available to everyone. Residents and owners of apartments, commercial buildings, governments and insurers alike. And this is where we’re heading in the future. Floods’ economic and societal implications are massive — they are the most frequent and deadly natural hazards worldwide — and affect everyone. There’s the possibility of flooding even in areas not close to rivers or seas, the so-called flash floods (what we recently saw in NYC or in drylands such as Saudi Arabia in the past). These pluvial floods are not even included in the traditional flood maps, which are predominantly based on water bodies’ location. So, we need to raise awareness of how technology can help more people stay safe through better flood understanding, mitigation plans, and, finally, forecasting. Improving our flood resilience is of utmost importance for societies in the years to come.
At REOR20, we’ve worked with the city of Lucerne and an energy utility company in Switzerland. We’ve decided, though, that the right first adopters are insurers, and at the moment, that is our focus. We are helping them figure out the risk of commercial properties, design and test efficient mitigation measures, and thus de-risk them, price fairly, build their reinsurance strategy, etc. Moreover, we’re working on a project with a large insurer to introduce our offering for residential property insurance business.
Thank you so much for taking the time, Danny!
DC: Appreciate it, Alex.
Check out job openings here from startups hiring in Greece, abroad, and remotely.
Robots can swim, fly, walk and crawl, but one fundamental aspect where they still lag behind is the sense of touch. Bota Systems (ETH Robotics Lab spin-off) fixes that by bringing together force torque sensors, electronics and AI. At Marathon, we led their $2.5m Seed round. (link)
Rated Labs raised $12.888m Series A to instil greater transparency and rich context in crypto infrastructure data. (link)
Unblocked raised $8.3m Seed to improve developer workflow. (link)
Generative AI startup that uses gamified quizzes to engage learners, Gizmo announced $3.5m Seed. (link)
Edtech platform Morphoses raised €2.1m pre seed for its gamified learning experience for soft skills development led by UniFund. (link)
LANGaware raised €2m led by Eleven Ventures with participation from Metavallon VC to predict and monitor cognitive and mental health conditions with AI. (link)
Infrastructure modeling platform that uses AI to create 3D models for simulations and analytics, Neara, announced a $24m Series B. (link)
Voda AI, a provider of AI-powered asset management for water and wastewater utilities, raised funding. (link)
Genesis Ventures accepting applications for its DayOne founders program in London until later today. (link)
The journey from entrepreneur to investor with Apostolos Apostolakis, Partner at Venture Friends. (link)
A deep-dive into DarkFi smart contracts with Odysseas Lamtzidis. (link)
Artificial intelligence, data privacy, academic research, and more with Timos Sellis, Scientific Director of the Archimedes Research Center. (link)
SaaS Synergy: the problem vs. solution dynamics from Mike Giannakopoulos, Staff Product Manager at HackTheBox. (link)
Marily Nika, AI Product Lead at Meta's Reality Labs, on types of AI product management roles. (link)
Lefteris Statharas, Researcher at Welfare Matters, on AI-powered surgery simulations. (link)
Resilient robotic autonomy with Kostas Alexis, Professor at Norwegian University of Science and Technology. (link)
If you’re in Athens next week, join us at “Open Coffee Athens #114” on Oct 12
“19th Meetup + NGXS Workshop” by Angular Athens on Oct 10 & 11
“Laravel Queues” by Athens Laravel Meetup on Oct 19
“Cross-functional collaboration in product development” by Product Community Greece on Oct 24
“Friends of Figma Athens” on Oct 25
“AI Summit” by Product-Led Hub on Nov 2