Weather Forecasting: an unexplored field for Quantum Computing

“Quantum computing is very promising, but we don’t know how to use it yet” says Dr. Peter Dueben, coordinator of machine learning and AI activities at the European Centre for Medium Range Weather Forecasts (ECMWF). Peter is one of ECMWF’s 390 staff members and works at the headquarter in Reading, UK. Established in 1975 as an independent intergovernmental organization with sites in Bologna, Italy, and Bonn, Germany, ECMWF is supported by 34 states in producing different types of numerical weather predictions ranging from global multi-day to seasonal predictions.

”To make good predictions of the weather, you need to represent the atmosphere but also many other components like ocean and land surfaces, cloud-physics, atmospheric chemistry, sea ice, land ice, etcetera” says Peter. This all is combined in a very complex multi-physics numerical forecast model that is run on ECMWF’s in-house supercomputer, one of the biggest machines in Europe and among the top 100 worldwide, multiple times per day. But that’s not the end of the story. The numerical forecast models need to be fed with initial and boundary conditions and that is where observations from weather stations and balloons but also planes and ships crossing the ocean and satellites come into play. For weather predictions of a single day, tens of millions of observations are assimilated into the numerical forecast model which itself is producing data that easily fills tens of terabytes.

For the efficient pre- and post-processing of this vast amount of data, ECMWF is using different types of statistical data analysis tools, among them machine learning, which is one of Peter’s areas of expertise. When asked about the opportunities for quantum computers to speed-up numerical weather forecasting in the future, Peter is not so sure. “The deep learning wave hit us some two to three years ago” he says. Today, AI-based tools that have been developed in collaboration with international research consortia partly funded by the European Commission, are investigated across the computational workflows at ECMWF. But the situation with quantum computing is different. When ECMWF got interested in machine learning, the technology already was at a rather high technology readiness level with ready-to-use hardware and software libraries that allowed researchers to quickly explore application opportunities. This is clearly not the case in quantum computing yet.

According to Peter, it’s not necessary to have the quantum computing hardware ready to make decision-makers interested in this topic. What would be more important is to have proof-of-concept emulations of quantum algorithms for practical applications such as solving global minimization problems, finding the solution to differential equations, or data assimilation. ECMWF is pursuing a lot of in-house research on AI, simulation technologies, and high-performance computing but building up all required expertise in quantum computing and algorithm development from scratch would be a long-term investment. Peter envisions collaborations between quantum computing experts from outside ECMWF and domain experts inside the organization as the most realistic way to go forward.

He sees collaboration as a win-win situation for both parties. “At ECMWF we have a lot of experience in dealing with uncertainties” he says. Numerical weather prediction is based on so-called ensemble models in which multiple weather model simulations are run with slightly different initial conditions. This helps the experts at ECMWF to forecast trends in temperature, winds, and rainfall with good accuracy. But there is more to it. When a tornado is approaching the U.S. west coast it is very difficult to predict long in advance if, where, and how strongly it will hit. In this case, safety measures need to be taken based on a probability analysis of the different scenarios to prepare the region for the most likely event. Whether the probabilistic nature of quantum computing can be of any help is not clear yet, says Peter. However, he sees the other way around, opportunities for ECMWF helping the quantum computing community in dealing with inaccuracies that might occur during computations. ECMWF has recently switched from double precision to single precision arithmetic in forecast simulations and is also exploring the use of half precision for certain parts of the computations. Dealing with inaccuracies in computation, and uncertainties in predictions is therefore not a new topic for ECMWF.

When asked about NP-hard problems in numerical weather forecasting, Peter surprised me by saying “Yes!”. A huge challenge in developing numerical weather prediction codes is the optimal placement and movement of data, which is the major bottleneck today. With high-performance computers becoming more and more heterogeneous this NP-hard problem will become more and more important and difficult to solve at the same time. Using quantum computers for it seems to be plausible for Peter with a big impact on all big-data applications.