Deep learning to avoid weather disappointments

Atmospheric aerosols can wreak havoc with weather forecasts. They are also major players in climate and public health. Yet the study of these highly complex and diverse suspensions of particles is hampered by the computational challenge of turning rich light scattering data into information on complex aerosol properties with sufficient speed and accuracy. Using competencies in modelling particle accelerators, a cross-disciplinary team of researchers addressed this challenge with a deep learning architecture that uses invertible neural networks. The ability to rapidly and accurately retrieve aerosol properties from light scattering data is crucial for weather predictions and real-time air quality monitoring.

A few weeks ago, as a Saharan dust storm moved across Europe, the first promised sunny days of spring were a disappointment. Instead of the glorious weather forecast, with highs of 19 degrees, the sun was obscured by dust and the mercury pushed a measly 15 degrees. This dust storm could be seen coming by satellite; so why were its effects not included in the weather forecast?

15th March 2022 at the Jungfraujoch High Altitude Research Station as a Saharan duststorm passes over (Copyright Switch,

It turns out that current weather models do not, or rarely, account for the presence of aerosols – which include dust storms as well as diverse pollutants and sea spray. Yet, these aerosols can have a major effect on radiative transfer, not only by the absorption and scattering of both incoming and outgoing radiation, but also by their effects on cloud formation.

One of the reasons that aerosols are omitted from weather predictions boils down to a computational bottleneck. Typically, the properties of atmospheric aerosols are studied with light scattering. These measurements are made using remote sensors, such as on satellites, or in the laboratory using an instrument known as a polar nephelometer. The data gathered contains rich information on the concentration, size, shape and refractive index of aerosol particles, and to retrieve these properties computational algorithms are used.

The recent dust storms are highly complex aerosols, consisting of variable shapes and sizes of particles, all of which affect the optical properties. To include aerosol data in weather models, large volumes of light-scattering data measured by satellites must first be rapidly analysed to yield aerosol properties. Current computational methods simply do not do this quickly or accurately enough.

To address this bottleneck, scientists from PSI’s Laboratory of Atmospheric Chemistry teamed up with the Laboratory for Simulation and Modelling, and exploited their competencies in modelling particle accelerators. Together they came up with a solution: a deep-learning algorithm using the technique of invertible neural networks. A hallmark of these ‘invertible’ algorithms is that a single model, trained in one direction on a particular data set, can be run in both the inverse and forward directions. Thus, as well as retrieving aerosol properties from light scattering data (the inverse direction), the algorithm can predict light scattering data from a given set of aerosol properties (the forward direction), giving great flexibility in applications.

Traditionally, solving the inverse problem is achieved using iterative optimisation methods based on the underlying physics of light scattering by aerosol particles. These methods, though accurate, are computationally heavy and thus slow. The machine learning approach, published in the Journal of Aerosol Science, speeds up calculations by factors of the order of 1000: improvements that are critical for data that feeds into weather models. “If this analysis is too slow, the observations are useless. The types of speed-ups we’ve achieved are therefore really significant in this context,” explains Rob Modini, scientist in the Aerosol Physics group.

This problem of speed has traditionally been solved using so-called “look-up tables”, which contain pre-computed aerosol properties from simulated light scattering data. This fix comes at the cost of accuracy. Romana Boiger, from the Accelerator Modelling and Advanced Simulations Group, who wrote the algorithm explains, “Our solution maintains the accuracy of the iterative physics-based algorithms, whilst winning on speed.”

A particular practical push for this study is a new polarized, laser imaging-type polar nephelometer, under development in the Aerosol Physics Group at PSI.  This instrument, much more portable than traditional nephelometers, enables laboratory light scattering measurements with high angular resolution. These measurements mimic the measurements made by satellites, and give an opportunity to validate computational techniques so that they can be applied accurately in remote sensing settings: and hence included in weather models. The invertible neural networks from Boiger, Modini et al focused on simulated in situ measurements, as made with the PSI polar nephelometer. Yet this machine learning approach is applicable to all good quality aerosol light scattering data, and provides a proof of concept that can be built upon further in remote sensing settings. Ultimately, the PSI researchers also hope to combine these algorithmic developments with their long-term aerosol monitoring activities, which have been running continuously at the Jungfraujoch since 1995. 

It may be a while before dust storms cease to play havoc with weather predictions. The amount of observational data crunched to produce the forecast is already colossal, and relies on supercomputers. Adding aerosols properties to weather models requires a step up in computational power, which is currently an active area of research. Once this has been reached, improvements in speed, as offered by Boiger, Modini et al will be critical. But the need to quickly and accurately analyse light scattering data of aerosols goes way beyond avoiding weather disappointments. Atmospheric aerosols also have a huge impact on climate and public health. Characterising them is key to understanding them and mitigating their effects. As technological improvements enable ever richer data to be gathered and push towards miniaturisation and real-time measurements of air quality in the field, corresponding improvements in computational techniques are required. Invertible neural networks models to retrieve aerosol properties could be one way of achieving this.

Text: Paul Scherrer Institute / Miriam Arrell