Machine learning technique could speed up glacier modeling a thousand times
A new model of glaciers has been developed that simulates ice dynamics and the interaction of ice with climate a thousand times faster than previous models. The model can be used to predict the evolution of glaciers and ice sheets under different scenarios. Because meltwater from glaciers and ice sheets is a major component of sea level rise, such models are valuable tools for assessing their potential future contributions.
The new model uses machine learning methods to make glacier modeling faster while maintaining high fidelity (the degree to which a simulation or model accurately reproduces the object or process it was designed to represent). As a result, more model simulations with different inputs and assumptions can be performed to investigate a wider range of questions.
state-of-the-art technology Guide to glacier models Compared to mature simulation tools, it is very efficient. It implements an artificial neural network, a computer system that mimics the neural networks in our brains. They “trained” the neural network by feeding it data from an ice sheet model, allowing it to simulate ice dynamics. This training process is called machine learning, and it is considered part of the field of artificial intelligence. Whereas previous modeling approaches to AI required a lot of human input, supervision, and decision-making, with machine learning, computer systems can navigate the manual process of updating models on their own.
lead developer, Guillaume JuveyA senior researcher at the University of Zurich explained: “[there is] A new trend in machine learning learning from data generated by physical models. “Physically-based models (also known as physical models) have long been used to understand the physical processes taking place in the Earth system without relying on any artificial intelligence.
Physically modeling ice sheets and glaciers at high spatial resolution is a huge challenge even today. Over the past two decades, extraordinary efforts have been made to develop models to simulate ice flow, its associated physical processes, and its interaction with climate.Increasing the complexity of the model increases the computational cost of the simulation, so most models usually use approximations Stokes equation, which most faithfully describes ice flow, requires a compromise between accuracy and computational cost. Jouvet describes the main motivation for transitioning to machine learning as “in a way, you’re shortening your physical modeling, making the computational gain cheaper.”
GlacierHub spoke with Laura Sandoval of the University of Colorado at Boulder, who leads review Enter artificial intelligence in the earth sciences. “Over the past decade, artificial intelligence [and] Machine learning activity has greatly increased in the geosciences, [but] Most AI work in geoscience research groups is still in its infancy,” she said. “Currently, researchers are actively exploring many AI models and prototype solutions to solve challenging problems in their fields. However, compared to traditional physics-based models, artificial intelligence and machine learning products have not yet had a major breakthrough. Sandoval added: “The implementation of artificial intelligence is still a work in progress. “
Instructed Glacier models replace the most computationally demanding model components by using neural networks trained from large datasets. Leverage the vast amounts of modeling data available to train neural networks to provide high-fidelity solutions at a much lower computational cost. It predicts ice flow from given variables and a simplified process for use in global glacier modeling and for studying past glacier environments.
An artificial neural network diagram that reflects the diagram of the human brain. Credit Liam Huang /frick.
“The most expensive part is computational dynamics because it involves heavy physics, [but] Machine learning accelerates this part of the model. The result is that we can model glaciers with the same accuracy faster than before.We can use it to explore more parameters and [conduct] More refined simulations,” Jouvet said. The study of the model took Jouvet and his team more than a year. He added: “I had to learn this new technology — all the tools I used were very new. . “
The researchers are delighted that they have machine learning up and running, and Jouvet now looks forward to using his model to reconstruct the evolution of Alpine glaciers during the last glacial cycle of 100,000 years. “The benefit of this approach is that you can speed up modeling so you can afford to work long hours. [Where] What might have taken weeks for a traditional model can now take an hour. “
The research team will now use a glacier model to reconstruct the past history of glaciers like this one, the Monarch Glacier in Switzerland’s Bernese Alps. Credit John Lillis/frick.
The implementation of AI and machine learning is not without challenges and doubts, similar to what has been seen in high profile cases biology and project“Ethics are really one of the main issues,” Sandoval explained. “However, since we’re still in early prototyping, the main arguments against AI at the moment are uncertainty, interpretability, and repeatability.” Ethical issues include the loss of human jobs, the uneven distribution of wealth created by AI machines, the security of AI data, and the ability for malicious intent. As AI implementations increase, more concerns are emerging, such as the environmental problems of using large amounts of energy to run computer models. Similar arguments against other web services such as cryptocurrencies and electronic transactions have been widely seen.
Scientists have been studying big questions about our climate and Earth system for years and amassing vast amounts of data that will be used to train artificial intelligence models. “Given the recent massive public and private sector investments in AI, we expect to see a boom in the application of data-centric AI research in geosciences over the next few years,” Sandoval said.
Despite the shift, not all geoscience problems can be solved by AI, and some are not well suited for classical machine learning techniques. “Some Earth phenomena are extreme events, and their patterns cannot be learned from historical data. Finding the right problem is a critical first step in developing successful AI applications,” Sandoval concluded.
The novel Instructed Glacier Model is a successful example of how new techniques for modeling glaciers can replace traditionally known physics-based methods. Much of the uncertainty surrounding AI remains, and whether there will be large-scale progress in the field is the question of the next decade. For now, old and new technologies will be implemented to provide answers to some of our biggest questions about ice sheets and glaciers.
