Extreme weather events can be important to the oil and gas industry. Hurricanes dodging offshore platforms dotted about in the Gulf of Mexico. Droughts that limit water supplies in high-volume fracking for shale oil plays.
Robert Webber is a postdoctoral scholar in the Department of Computing and Mathematical Sciences at CalTech. His group are working on a genetic model to lower the uncertainty of the predictions of, say, a hurricane category, track, and landfall in the Gulf of Mexico. This could save a lot of money if fewer offshore platforms have to be abandoned.
Webber’s research, which is featured in SIAM News, currently focuses on predicting the likelihood that a category five hurricane will hit a major city. However, it can be applied to any weather extremes for which weather simulation models already exist, including floods, heat waves, wildfires, and extreme cold temperatures caused by the polar vortex. It is the first genetic algorithm – a high-level math procedure inspired by the process of natural selection – to deliver accurate weather predictions by using smaller sample sizes and fewer computing resources.
Robert was interviewed in early July 2023:
Q1: How did you come up with the idea of using genetic algorithms traditionally used in biochemistry and finance to solve complex medical issues/predict stock market vulnerabilities?
I’m an expert in rare event sampling methods, so I know the history of these methods and how they’ve been applied successfully, e.g., to understand diseases in the human body or how stock prices can catastrophically collapse. Despite the success of rare event sampling methods in several scientific areas, I noticed they’re not widely used in geophysics.
When geophysicists ask climate change questions, like, “Will we see more Category 5 hurricanes in the future?”, they rely on an inefficient method based on brute-force sampling. This can lead to vague answers with large error bars. The geophysicists are missing the opportunity to accelerate their calculations, to ask more penetrating questions, and to obtain more precise answers.
Looking at the research landscape a few years back, I saw there was an opportunity to improve the computations of extreme events in geophysics, so I started developing the genetic method and applying to study hurricanes.
Q2: Can you talk about one practical application of your genetic model?
One simple, practical application of my genetic method is real-time weather prediction for extreme events. Suppose there is a hurricane that is starting to develop and moving in the direction of the East Coast of the United States. We want to know the probability the hurricane will make landfall, where will it make landfall, and how likely is the storm to reach Category 4 or 5 status.
Typically, the answers given by the weather centers are vague, because the hurricane models require a large sample size of 10,000 simulations to give precise answers, and we can’t afford to run the hurricane models enough times because they’re very expensive models. My technique allows us to reduce the sample size to 100-1000 simulations, which is more manageable. We can start to give a precise answer, like there is a 10% chance of a Category 4 storm hitting Boston, which enables policymakers to plan effectively.
This is just one example, but it’s a general-purpose method that can be applied to any extreme weather or climate event that builds up incrementally.
Q3: How does your genetic model reduce the uncertainty of predicted results from an underlying model, such as one that predicts hurricane strength? Why is this important? Can your model predict new trends, or does it just refine the trends already predicted by the underlying stimulation?
The genetic method is based on a fundamental change in the way we run our computer simulations. Typical computer simulations of an extreme event, such as a hurricane, are wasteful: we spend a long time simulating hurricanes that are not very intense and we only see a few examples of truly intense hurricanes that are the most impactful and critical for policymakers.
A better way to run our simulations is the genetic method, which is based on applying the survival-of-fittest principle, where the “fittest” is defined as the most extreme version of a weather or climate event. We periodically stop our simulations, identify the “fittest” members, and devote extra resources toward those simulations. In this way, we can obtain many more simulations of an intense hurricane occurring (the extreme), and we can better understand the probability of occurrence, the location of landfall, the storm surge, and the rainfall pattern of the extreme hurricane, with an eye toward future damages.
It’s important to acknowledge the genetic method gives the same answer that we would obtain from the underlying weather or climate model, if we could afford to run the model 10,000 times —- it just does so more quickly and efficiently. Therefore, if the weather or climate model is giving incorrect predictions of extreme weather events, we can see this more clearly with the genetic method. We can more quickly identify any deficiencies of our models and try to address them. However, the genetic model cannot fix these underlying modeling problems. We need to pass the models back to the developers and ask the developers to fix them.
Q4: Oil and gas producers in the Gulf of Mexico are always concerned about the next hurricane, especially now that its hurricane season. Can your genetic model improve the prediction, and how does it do that?
The genetic method can help to predict the most extreme hurricanes occurring in the 2023 hurricane season and their pathways and landfalls. These predictions are already possible using a high-resolution weather model like the Weather Research & Forecasting model (from NOAA), but they are too expensive to run because we would need tens of thousands of high-performance computers running for weeks.
The genetic method reduces the computational cost by 1-2 orders of magnitude, making it much more practical to get a better hurricane category, and track to see what offshore production platforms might be hit.
Q5: Droughts are important to the oil and gas industry, because up to 70% of shale oil is produced from water-challenged regions of the US (shale oil is over 60% of total US production). The shale revolution depends on fracking and fracking uses a lot of water — just one well uses as much water as would cover the grassed area of a football stadium to a height of 40 feet. Would it be feasible for your genetic model to better predict future droughts in the Permian basin of West Texas, for example, so oil and gas companies could plan their water needs better?
Predicting droughts is a perfect application for the genetic method because droughts are extreme events that build up gradually over time. If we apply the genetic method to this problem, we would be allocating extra resources to simulate water-starved years over West Texas, while ignoring the years that are characterized by typical or high rainfall. The genetic method would give the best probabilities of future droughts occurring, while also identifying the geographic patterns of droughts over West Texas, should they occur.
Q6: Have you thought about working with the oil and gas industry to use the genetic model to run more reservoir simulations in a shorter time to improve predictions of oil and gas production in a particular field? Or is the genetic model only used to improve statistics of extreme events?
The genetic method is useful for extreme events, not regular events. Regular events are already accessible using direct simulations with a relatively small sample size, say 100 simulations. The genetic method is important because it makes it nearly as easy to simulate extreme events as regular events. It reduces the sample size requirements by a factor of 10x – 100x, making extreme events more accessible.
Q7: How have various owners of underlying simulation models responded to your approach in the genetic model?
I’ve often received a warm reception for my work, since the major climate and weather centers understand how rare event sampling is useful and likely to become important in the future. I’ve even received offers of partnerships e.g., with the national labs, to develop my genetic method further.
Nonetheless, this is a cutting-edge method which has not been widely adapted yet. The owners of the simulation models are accustomed to their standard way of running the models, by using brute-force sampling. They want strong evidence for the validity of a new method before they start using it. To address these concerns and convince the owners of the simulation models, my research group and another competing research group in France have been actively developing rare event sampling and applying this idea to hard problems, such as heat waves, rogue waves, flooding events, and hurricanes.
While I’m 100% confident that my method or another similar method will be adopted in the mainstream, we’ll have to wait and see how quickly this happens. I expect it’s a matter of years, not decades. The momentum behind these methods is building, and the need is great.
Read the full article here