Physics-informed learning of aerosol microphysics

Abstract

Aerosol particles play an important role in the climate system by absorbing and scattering radiation and influencing cloud properties. They are also one of the biggest sources of uncertainty for climate modeling. Many climate models do not include aerosols in sufficient detail due to computational constraints. To represent key processes, aerosol microphysical properties and processes have to be accounted for. This is done in the ECHAM-HAM (European Center for Medium-Range Weather Forecast-Hamburg-Hamburg) global climate aerosol model using the M7 microphysics, but high computational costs make it very expensive to run with finer resolution or for a longer time. We aim to use machine learning to emulate the microphysics model at sufficient accuracy and reduce the computational cost by being fast at inference time. The original M7 model is used to generate data of input–output pairs to train a neural network (NN) on it. We are able to learn the variables’ tendencies achieving an average $ {R}^2 $ score of 77.1%. We further explore methods to inform and constrain the NN with physical knowledge to reduce mass violation and enforce mass positivity. On a Graphics processing unit (GPU), we achieve a speed-up of up to over 64 times faster when compared to the original model.

Addressing Complexity in Global Aerosol Climate Model Cloud Microphysics

“…(2022); Harder et al. (2021); Fletcher et al. (2022)), but rather to understand how the model can be simplified.…”

“…Note that we use the emulation merely as a tool for model analysis. Our aim is by no means to replace model components with machine learned substitutes or to replace the full model (in contrast to e.g., Arcomano et al (2022); Harder et al (2021); Fletcher et al (2022)), but rather to understand how the model can be simplified.…”

“…On a smaller scale, machine-learning is also used to replace or improve single parameterizations or schemes (e.g., A. Seifert and Rasp (2020); Gettelman et al (2021); Harder et al (2021); Meyer et al (2022)). However, while these approaches may help to reduce the computational load of climate simulations, they do little to improve interpretability of the model (Rudin, 2019).…”

Machine Learning for numerical weather and climate modelling: a review

“…As has been briefly touched on in previous sections, a promising and increasingly popular method for improving the performance of ML applications in weather and climate modelling is to include physics-based constraints in the ML model design (e.g. Karpatne et al, 2017;de Bézenac et al, 2017;Beucler et al, 2019;Yuval et al, 2021;Harder et al, 2022). This can be done through the overall design and formulation of the model, and through the use of custom loss functions which impose physically-motivated conservations and constraints.…”

Advecting Superspecies: Efficiently Modeling Transport of Organic Aerosol With a Mass‐Conserving Dimensionality Reduction Method

“…Though physical consistency is an important result by itself, imposed constraints do not necessarily improve accuracy of such tools beyond adherence to whichever physical law(s) the constraints enforce. For example, Harder et al (2022) found the accuracy of a neural network surrogate model of aerosol microphysics was not improved when adding a completion constraint during training, where a chosen variable was reassigned to the sum of all other variables' tendencies to conserve mass. However, constraints can be implemented in ways that add domain knowledge to the data-driven algorithm: Sturm and Wexler (2022) found that by adjusting a feed-forward neural network archi tecture to include a flux-tendency constraint during training, the overall prediction accuracy of chemical species concentrations improved.…”