New geoscientific modeling tool yields more holistic results in forecasting

Chaopeng Shen, associate professor of civil and environmental engineering at Penn State (seated, right), and postdoctoral researcher and graduate students in his lab study a map that displays information derived from their new geoscience modeling tool. Credit: Kelby Hochreither/State of Penn

Geoscientific models allow researchers to test potential scenarios with digital representations of the Earth and relevant systems, from predicting the effects of large-scale climate change to aiding in the development of land management practices. However, estimating the parameters of traditional models is computationally expensive and calculates results for specific locations and scenarios that are difficult to extrapolate to other scenarios, according to Chaopeng Shen, associate professor of civil and environmental engineering at Penn State. .

To address these issues, Shen and other researchers have developed a new model known as differentiable parameter learning that combines elements of traditional process-based models and machine learning for a method that can be applied at scale and lead to more aggregated solutions. Their model, published in Nature Communicationis publicly available to researchers.

“A problem with traditional process-based models is that they all need some kind of parameters – the variables in the equation that describe some attribute of the geophysical system, such as the conductivity of an aquifer or rainwater runoff — which they don’t have direct observations for, Shen said. “Normally you would go through this process called parameter inversion or parameter estimation where you have a few observations of the variables that the models are going to predict and then you go back and ask, ‘What should my parameter be?’

A common process-based model is an evolutionary algorithm, which evolves through many iterations of operation so that it can better adjust parameters. However, these algorithms are not able to handle large scales or to be generalized to other contexts.

“It’s like I’m trying to fix my house, and my neighbor has a similar problem and is trying to fix his house, and there’s no communication between us,” Shen said. “Everyone is trying to do their own thing. Similarly, when you apply evolutionary algorithms to an area – say in the United States – you’re solving a separate problem for each little piece of land, and there’s no communication between them, so there’s a lot of wasted effort, plus everyone can solve their problem inconsistently, and that introduces a lot of physical unrealism.

To solve larger region problems, Shen’s model considers data from all locations to arrive at a solution. Instead of entering data from location A and getting the solution from location A, then entering data from location B for the solution from location B, Shen enters the data from the locations A and B for a more complete solution.

“Our algorithm is much more holistic because we use an overall loss function,” he said. “This means that during the process of parameter estimation, the loss function of each location – the deviation between your model output and observations – is aggregated. Problems are solved together at the same time. I am looking for a solution to the whole continent. And when you bring more data points into this workflow, everyone gets better results. While there were also other methods that used an overall loss function, humans derived the formula, so the results were not optimal.”

Shen also noted that his method is much more computationally cost-effective than traditional methods. What would normally take a 100-CPU super cluster two to three days now only requires a single hour-long GPU.

“The cost per grid cell has dropped tremendously,” he said. “It’s like economies of scale. If you have a factory that builds one car, but now you have the same factory that builds 10,000 cars, your cost per unit drops dramatically. And the same thing happens when you bring more points in this workflow. At the same time, each site now has a better service thanks to the participation of other sites.”

Pure machine learning methods can make good predictions for widely observed variables, but they can produce results that are difficult to interpret because they do not include assessment of the causal relationship.

“A deep learning model can make a good prediction, but we don’t know how it did it,” Shen said, explaining that while a model can do a good job of making predictions, researchers can misinterpret the apparent causal relationship. “With our approach, we are able to organically link process-based models and machine learning at a fundamental level to leverage all the benefits of machine learning as well as insights that come from the physical side. ”

Machine learning refines flash graphene

More information:
Wen-Ping Tsai et al, From Calibration to Parameter Learning: Harnessing Big Data Scale Effects in Geoscience Modeling, Nature Communication (2021). DOI: 10.1038/s41467-021-26107-z

Provided by Pennsylvania State University

Quote: New geoscientific modeling tool yields more holistic results in forecasts (2022, March 4) retrieved March 4, 2022 from

This document is subject to copyright. Except for fair use for purposes of private study or research, no part may be reproduced without written permission. The content is provided for information only.

Comments are closed.