Minisymposium

Aerosol mixing state refers to the way that different chemical components are distributed amongst the particles in an aerosol population. It is an important emergent property of the atmospheric aerosol that affects aerosol radiative forcing and aerosol–cloud interactions. However, current aerosol models used in global Earth system models are limited in their capability to represent aerosol mixing state, thereby suffering from largely unquantified structural uncertainty. In contrast, particle-resolved aerosol simulations are able to capture aerosol mixing state faithfully, however, they are too computationally expensive to be directly implemented into Earth system models.

This study integrates machine learning and particle-resolved aerosol simulations to develop emulators that predict submicron aerosol mixing state indices from the Earth system model simulations. The emulators predict aerosol mixing state using only quantities that are predicted by the Earth system model, including bulk aerosol species concentrations, which do not by themselves carry mixing state information. This work is a prototypical example of using machine learning emulators to add information to Earth system model simulations and to quantify structural uncertainty in Earth system models.

Predicting secondary organic aerosol (SOA) mass is of crucial importance as it is a significant contributor to the atmospheric particulate load. SOA therefore has an impact on aerosol optical, hygroscopic and toxic properties. There is a still a large gap between the complexity of processes involved in SOA formation and the simplicity of their representation in air quality and climate models. GECKO-A is a tool used for generating explicit organic chemistry chemical mechanisms, aiming at reproducing the complexity of SOA formation processes. These mechanisms cannot be applied in 3D models due to their unpractical sizes. Machine learning has been used to accelerate chemistry solvers in 3D models by emulating their behavior at a fraction of their computational cost. Here we present a similar approach to emulate the behavior of complex GECKO-A mechanisms to predict SOA formation. Different methods, including neural networks and random forests, are used and we quantify their performances and weaknesses. This specific problem requires the construction of ad-hoc datasets, and we illustrate the issues associated with this.

The last decade has been a huge leap forward in atmospheric new particle formation (NPF) research. Novel instrument development has allowed us to measure more and more details of the dynamics of even the smallest clusters. At the same time, however, for example NPF and particle growth rates have been analyzed typically by very simple regression or balance equation approaches, permitting no proper estimation of the uncertainties. Here we combine a Bayesian approach with finite element method (FEM) approximation of the size distribution to estimate unknown rate parameters in the aerosol general dynamic equation. The method is based on Kalman Filter and Kalman Smoother methods which allow for the estimation of the parameters and their error covariance matrices. The unknowns are modeled as random variables, and their prior probability distributions are incorporated in the solution of the inverse problem. As a first step, we test our methodology with synthetic data, generated by a detailed aerosol dynamics model. The advantage of this approach is that the ‘answers’ are known, i.e. we know in detail, for example, the time evolution of the nucleation and condensational growth rates as well as the size dependence of the deposition and condensational growth rates.

Following the presentations, we will have a discussion on the challenges and the potentials of scaling data-driven methods for aerosols and clouds to global climate predictions.