fields/climate-weather/model families
Multilayer perceptrons in weather, climate and oceans
How weather, climate and ocean researchers use the plainest kind of neural network, a stack of fully connected layers, to turn tables of measurements into estimates of quantities they cannot observe directly.
How it is used
Weather, climate and ocean science studies the atmosphere, the sea and the exchanges between them and the land. Much of the work is turning one set of measurements into another: from temperature, wind, humidity and radiation to a flux of carbon, a pollutant concentration, a current, or a forecast. A multilayer perceptron is the simplest neural network for that job. It takes a fixed list of numbers, passes them through several layers of weighted sums and simple nonlinear functions, and returns one or more numbers. It assumes nothing about how the inputs are arranged, so it is used where the data already sits in a table: one row per site, per hour or per month.
In the examples here it appears in two guises. Sometimes it is the whole model, fitted to a table of monthly emissions, fuel use, traffic and weather averages, or to half-hourly tower fluxes and their meteorological drivers. Sometimes it is only the last part of a larger network: convolution layers or a recurrent branch first compress a map or a time series, and fully connected layers at the end combine what they found into the answer. In that role it is the piece that mixes everything together, and it is doing the same arithmetic either way.
Why researchers use it
Researchers reach for it because the relations they want are often known to be nonlinear but not known in form. Radiation, temperature and moisture act together on a carbon flux, and air traffic, fuel use and the weather act together on particle levels, in ways no tidy equation captures. A multilayer perceptron fits such a surface from examples without being told its shape. It stands in, in effect, for a regression equation the researcher would otherwise have to guess, and for the long chain of physical calculations that a process model would need.
It is also cheap and undemanding. It trains on modest tables on ordinary hardware, sits easily inside off-the-shelf software alongside linear and tree-based methods, and can be compared against them in the same run. Because it imposes no structure, it is the natural default when the inputs have no grid and no ordering, and the natural final stage when some other part of the network has already handled the space or the time.
In the record
Limits
Having no structural assumptions is also the weakness. The network cannot see that neighbouring grid cells are neighbours, or that yesterday came before today, so anything spatial or temporal must be fed in as hand-made features or handled by another part of the model. It interpolates within the conditions it saw and can drift badly outside them, which matters when the question is a future climate. With short records it readily fits noise. And it offers no mechanism, so a good fit need not mean the physics was captured.
Written by AIxSci: the background from general knowledge, the examples from checked records.
Related
Every article
aix-00306
Benchmarking the cost of ensemble Kalman methods for calibrating chaotic models
aix-00284
Neural network turns hourly satellite heat images into Gulf Stream current maps
aix-00282
Neural network trained on a decade of records forecasts typhoon strength and position
aix-00292
Neural network forecasts daily sea surface temperature up to 90 days ahead
aix-00293
Machine learning used to predict fine particle emissions at a Greek regional airport
aix-00294
Neural network trained on simple equations spots tipping points in real data
aix-00297
Weather and climate model built from physics equations plus a trained neural network
aix-00299
Neural network forecasts weather straight from raw satellite and station observations
aix-00280