Deep Neural Networks Accelerate Weather Model Simulation Speed.

2019-11-17

When you check the weather forecast in the morning, the results you see are likely derived from the Weather Research and Forecasting (WRF) model, which is a comprehensive model that simulates many aspects of the natural world around us.


From clouds to solar radiation, from snow to vegetation, and even wind disturbed by skyscrapers.


Countless features of weather and climate are coupled and interact with each other. Scientists have not used simple unified equations to describe these complex relationships. Instead, they use an approach called parameterization to approximate these relationships, in which they model the relationships at a larger scale than the actual phenomena.Although parameterization simplifies the physical processes in some way, allowing the model to produce relatively accurate results in a reasonable amount of time, the computational cost is still high. Environmental scientists and computer scientists from Argonne are collaborating to use a type of machine learning - deep neural networks - to replace parameterization of certain physical schemes in the WRF model, thereby reducing processing time.


In a recent study, scientists focused on the Planetary Boundary Layer (PBL), the lowest part of the atmosphere. PBL is the atmospheric layer most affected by human activities, extending only a few hundred meters above the surface. The dynamics in this layer, such as wind speed, temperature, and humidity profiles, are crucial for determining many physical processes in the atmosphere and on Earth.


PBL is a key component in the WRF model, but it is also one of the least computationally expensive components. This makes it an excellent testbed for studying how deep learning neural networks can enhance more complex components in a similar way.


"We used 20 years of computationally generated data from the WRF model to train the neural networks and two years of data to evaluate whether they can provide accurate alternatives to physics-based parameterizations," said Prasanna Balaprakash, a computer scientist at the Department of Energy.


Any machine learning project requires a large amount of high-quality data, and this study is not lacking in data. The ALCF and the National Energy Research Scientific Computing Center's supercomputing resources generated over 300 years (700 TB) of data, describing past, present, and future weather and climate in North America.Balaprakash developed the neural network and trained it to learn the abstract relationship between inputs and outputs by providing over 10,000 data points (8 per day) from two locations (one in Kansas and one in Alaska). The result is an algorithm that scientists believe can replace the PBL parameterization in the WRF model.


The algorithm, called domain-aware neural network, takes into account known relationships and can not only predict environmental data more accurately but also requires much less training data compared to algorithms that do not consider domain expertise.


Scientists have demonstrated that deep neural networks with underlying structures that consider the relationships between input and output variables can successfully simulate time-varying wind speed, temperature, and water vapor data. The results also indicate that a trained neural network can predict the state of nearby locations from one location with a correlation of over 90% compared to the test data.


The ultimate goal of scientists is to replace all expensive parameterization settings in the WRF model with deep learning neural networks to achieve faster and higher-resolution simulations. Currently, the team is working on parameterizing longwave and shortwave solar radiation to minimize the time required for the two parts of the WRF model to come together.

RECENT AI NEWS

RECENT AI TOOLS