Advecting Superspecies: Efficiently Modeling Transport of Organic Aerosol with a Mass-Conserving Dimensionality Reduction Method

Abstract

The chemical transport model LOTOS-EUROS uses a volatility basis set (VBS) approach to represent the formation of secondary organic aerosol (SOA) in the atmosphere. Inclusion of the VBS approximately doubles the dimensionality of LOTOS-EUROS and slows computation of the advection operator by a factor of two. This complexity limits SOA representation in operational forecasts. We develop a mass-conserving machine learning (ML) method based on matrix factorization to find latent patterns in the VBS tracers that correspond to a lower-dimension set of superspecies. Tracers are reversibly compressed to superspecies before transport, and the superspecies are subsequently decompressed to tracers for process-based SOA modeling. This physically interpretable ML method conserves the total concentration and phase of the tracers throughout the process. The superspecies approach is implemented in LOTOS-EUROS and found to accelerate the advection operator by a factor of 1.5 to 1.8. Concentrations remain numerically stable over model simulation times of two weeks, including simulations at higher spatial resolutions than the ML models were trained on. Results from this case study show that this method can be used to enable detailed, process-based secondary organic aerosol representation in air quality operational forecasts in a computationally efficient manner. Beyond this case study, the physically consistent ML approach developed in this work enforces conservation laws that are essential to other Earth system modeling applications, and generalizes to other processes where computational benefit can be gained from a two-way mapping between detailed process variables and their representation in a reduced-dimensional space.

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

“…Despite conserving total concentration of all tracers, the concentration of total organic aerosol (TOA) may not be conserved due to errors in the mass distribution over volatility bins after decompression. A variation of this method to conserve TOA instead of total concentration, as well as an alternative way to conserve total concentration only using W from NMF, is explored in Sturm (2021). However, the compromise of conserving TOA versus total concentration is avoidable by adding another cross section: creating compression and decompression matrices B and W for each phase as well as VBS class, for example, one transformation for all biogenic gaseous VBS tracers and a separate transformation for all biogenic particle tracers.…”

“…The advection operator consumes a significant amount of wall time in LOTOS‐EUROS, from about 20% of total wall time in sequential runs (only chemistry and sometimes deposition calculations take longer) to over 50% of total wall time in parallel runs using domain decomposition. Wall time of advection can double with the inclusion of organic aerosol tracers (Sturm, 2021). Therefore, the current default in LOTOS‐EUROS is to include only one passive tracer for organic aerosol, which is one of the reasons for an underestimation of total particulate matter (Timmermans et al., 2022).…”

“…When using the VBS module, the number of advected tracers increases from 46 to 104. Model timing experiments in Sturm (2021) found that wall time for the advection operator can double when using the VBS module. Advection is a bulk process and does not perform OA‐specific calculations.…”

“…One strategy for physically consistent data-driven models reposes the learning targets: rather than estimate important properties or their tendencies, instead estimate fluxes between the properties. The fluxes can then be related to tendencies in a way that balances mass, energy, or atoms (Sturm & Wexler, 2020, 2022Yuval, O'Gorman, et al, 2021). Custom neural network architectures can also obey conservation laws by incorporating hard constraints in their hidden layers (Beucler et al, 2021), such as flux balances (Sturm & Wexler, 2022): this can also improve the physical interpretability of the inner working of neural networks.…”