SuperdropNet: a Stable and Accurate Machine Learning Proxy for Droplet-based Cloud Microphysics

TL;DR

The results suggest that ML models can effectively emulate cloud microphysics, in a manner consistent with droplet‐based simulations, and how multistep autoregressive training improves performance is revealed.

Abstract

Cloud microphysics has important consequences for climate and weather phenomena, and inaccurate representations can limit forecast accuracy. While atmospheric models increasingly resolve storms and clouds, the accuracy of the underlying microphysics remains limited by computationally expedient bulk moment schemes based on simplifying assumptions. Droplet-based Lagrangian schemes are more accurate but are underutilized due to their large computational overhead. Machine learning (ML) based schemes can bridge this gap by learning from vast droplet-based simulation datasets, but have so far struggled to match the accuracy and stability of bulk moment schemes. To address this challenge, we developed SuperdropNet, an ML-based emulator of the Lagrangian superdroplet simulations. To improve accuracy and stability, we employ multi-step autoregressive prediction during training, impose physical constraints, and carefully control stochasticity in the training data. Superdropnet predicted hydrometeor states and cloud-to-rain transition times more accurately than previous ML emulators, and matched or outperformed bulk moment schemes in many cases. We further carried out detailed analyses to reveal how multistep autoregressive training improves performance, and how the performance of SuperdropNet and other microphysical schemes hydrometeors' mass, number and size distribution. Together our results suggest that ML models can effectively emulate cloud microphysics, in a manner consistent with droplet-based simulations.

Figures (11)

• Figure 1: Stochasticity of superdroplet simulations with box size $25m^{3}$. (a) and (b) correspond to 100 simulations at multiplicity of 1024. (c) and (d) correspond to 100 simulations at multiplicity of 25600. For both sets of simulations, rain water mass ($L_r$) and rain number concentration($N_r$) are compared. Other initial conditions for both simulations are the same with ${L_0}$=0.3 g/m$^3$, $r_0$=13 $\mu$m, $\nu$=0.5.
• Figure 2: Stochasticity in simulations at a box size of 2500 $m^{3}$ and multiplicity is 25600. (a) and (b) correspond to 100 simulations at $\nu$=0.5. (c) and (d) correspond to 100 simulations at $\nu$=3. For both sets of simulations, Rain water mass($L_r$) and Rain number concentration($N_r$) are compared. Other initial conditions for both simulations are the same with ${L_0}$=0.3 g/m$^3$, $r_0$=13 $\mu$m.
• Figure 3: Top panels, red -- Superdroplet-derived changes in bulk moments over single time steps ($\Delta t = 20$s) vs. changes over single time steps predicted by a neural network trained to predict one time steps into the future. Bottom panels, blue -- Superdroplet-derived bulk moments vs. predictions on time step ahead by the same network. Results are shown only for initial conditions in the held-out testing data. Mean absolute percentage errors (MAPE) are shown for each comparison.
• Figure 4: Superdroplet-derived bulk moments (black lines) compared to rollouts from a neural network trained to predict 1 step into the future (red-dashed lines), SuperdropNet (blue solid line), PRNet (yellow-dashed lines) and from a classical bulk moment scheme (blue-dotted lines). Results are shown for a simulation with $L_0=2$ g/$m^3$, $r_0=9\mu$m, $\nu=0$. Shaded region indicates +/- 1 standard deviation over 100 superdroplet simulations. A single time step corresponds to 20 s of simulation time.
• Figure 5: Superdroplet-derived bulk moments (black lines) compared to rollouts from a neural network trained to predict 1 step into the future (red-dashed lines), SuperdropNet (blue solid line), PRNet (yellow-dashed lines) and from a classical bulk moment scheme (blue-dotted lines). Results are shown for a simulation with $L_0=0.2$ g/$m^3$, $r_0=9\mu$m, $\nu=2$. Shaded region indicates +/- 1 standard deviation over 100 superdroplet simulations. A single time step corresponds to 20 s of simulation time.