Numerical weather prediction (NWP) is a mainstay of supercomputing. Some of the first applications of the first supercomputers dealt with climate modeling, and even to this day, the largest climate models are heavily constrained by the scale of the supercomputers that run them. While some wait for the exascale era – and beyond – to brute force punishingly accurate and complex climate models into existence, others are looking for a deep learning-powered shortcut to the same results. In a paper for Philosophical Transactions of the Royal Society, eight researchers from the Jülich Supercomputing Center explored whether deep learning could ever actually beat numerical weather prediction at its own game – and if so, how and when that might happen.
The weather and climate supercomputing community is no stranger to deep learning, but it has hitherto mostly been used to augment NWP approaches (e.g. in resolving post-processing issues). These modelers, however, are reticent to incorporate deep learning in more meaningful capacities.
“[There] are still reservations about DL in this community,” the authors write. “Two core arguments in this regard are the lack of explainability of deep [neural networks] and the lack of physical constraints. Furthermore, some scepticism prevails due to the fact that researchers have experimented with rather simple [neural networks] which were clearly unsuited to capture the complexity of meteorological data and feedback processes, but then extrapolate these results to discredit any [neural network] application including the much more powerful [deep learning] systems.”
While the paper explores whether deep learning could eventually replace significant elements of a major NWP model, it’s perhaps more interested in whether deep learning could replace the whole thing.
That bar, of course, is extraordinarily high.
“Over the past decades, the ability of NWP models to predict the future atmospheric state has continuously improved,” the paper reads. “Contemporary global NWP models are not only able to predict the synoptic-scale weather pattern for several days, but they have also reached remarkable accuracy in forecasting end-user relevant meteorological quantities such as the 2m temperature and regional-scale precipitation events.”
But deep learning hasn’t been standing still, either – far from it. Steep increases in available computational power also benefit deep learning applications, which are also boosted by increased data availability and a rapidly expanding library of neural network architectures.
In fact, some researchers have already carried out NWP-mimicking deep learning tests – but, the authors note, these studies have been extremely limited in scope, focusing on forecasting by up to a day.
The authors suggest that any eventual deep learning replacement for an NWP would likely consist of several neural networks trained on subsets of forecast products, allowing deep learning techniques to excel by focusing on specific tasks. Key to this approach, they say, is understanding the distributions of meteorological and climatological variables, which can be both complex and crucial: by way of example, they discuss sea ice, which might change very little over the course of a typical forecast, but which produces profound effects in the medium- to long-term.
A number of challenges face deep learning as it climbs toward NWP. For instance, rare extreme weather events are difficult in terms of training and testing, though the authors report some success across various studies in accounting for this gap. Data availability is another problem: NWP typically uses satellite data where missing values are interpolated, but using such filled-in data with deep learning models poses a serious risk of concept drift, where an assumption made early on leads to cascading built-in biases.
Indeed, the authors say that with respect to data preparation generally, “best practices differ between the meteorological and ML communities.” Machine learning development, they explain, typically involves three datasets: a training dataset, a validation dataset and a test dataset, all of which should be independent from one another. But there’s a problem here, at least for weather prediction: the data is auto-correlated, meaning the datasets aren’t truly independent.
Furthermore, neural networks, the authors say, may need to be directly taught the relationships between certain variables, as short- to medium-term datasets are unlikely to teach a model to understand longer-term variations like El Nino or climate change. This need for intervention extends to limiting factors, as well: deep learning models might be inspired to produce physically impossible forecasts or establish scientifically unsound correlation-causation links. The authors say that some studies have introduced such physical restraints to general success.
“It may be useful to reflect on the potential and necessity of physically constraining [deep learning] models from an abstract point of view,” they add. “In spite of their complexity and dimensionality, [deep learning] models still adhere to the fundamental principles of (data-driven) statistical modelling. This implies that there must be some rules in place to constrain the future, because otherwise extrapolation will be unbound.”
Finally, the authors touch on uncertainty estimation. Ensemble models, which use a series of runs to estimate the relative likelihood of various outcomes, have become more or less the norm in top-of-the-line NWP. However, ensemble approaches introduce exorbitant computational costs for deep learning models. The authors discuss Bayesian deep learning as a reasonable alternative, noting that it has already been tested for weather forecasting applications.
So: where is deep learning-powered weather prediction heading?
“We expect that the field of ML in weather and climate science will grow rapidly in the coming years as more and more sophisticated ML architectures are becoming available and can easily be deployed on modern computer systems,” the authors write. “We [also] expect that the success of [deep learning] weather forecast applications will hinge on the consideration of physical constraints in the [neural network] design. Taken to the extreme, portions or variants of current numerical models could eventually end up as regulators in the latent space of deep neural weather forecasting networks.”
“So, to answer the question posed in the title of this article,” they conclude, “we can only say that there might be potential for end-to-end [deep learning] weather forecast applications to produce equal or better quality forecasts for specific end-user demands, especially if these systems can exploit small-scale patterns in the observational data which are not resolved in the traditional NWP model chain.”
“Whether [deep learning] will evolve enough to replace most or all of the current NWP systems cannot be answered at this point.”
About the research
The paper discussed in this article, “Can deep learning beat numerical weather prediction?“, was published in the February 2021 issue of Philosophical Transactions of the Royal Society. The paper was written by M. G. Schultz, C. Betancourt, B. Gong, F. Kleinert, M. Langguth, L. H. Leufen, A. Mozaffari and S. Stadtler, all of the Jülich Supercomputing Center in Germany.