Atmospheric chemistry models—used as components in models that simulate air pollution and climate change—are computationally expensive. Previous studies have shown that machine-learned atmospheric chemical solvers can be orders of magnitude faster than traditional integration methods but tend to suffer from numerical instability. Here, we present a modeling framework that reduces error accumulation compared to previous work while maintaining computational efficiency. Our approach is novel in that it: 1) uses a recurrent training regime that results in extended (>1 week) simulations without runaway error accumulation, and 2) can reversibly compress the number of modeled chemical species by >80% without further decreasing accuracy. We observe a ~260× reduction in computation time (~1900× when run on specialized hardware) compared to the traditional solver. We use random initial conditions in training to promote general applicability across a wide range of atmospheric conditions. For ozone (with an initial concentration range of 0–70 ppb), our model predictions over a 24-hour simulation period match those of the traditional solvers with median error of 2.7 ppb and less than 19 ppb error across 99% of simulations initialized with random noise. Error can be significantly higher in the remaining 1% of simulations, which include among the most extreme concentration fluctuations simulated by the reference model. Results are similar for total particulate matter (median error of 16 ug/m3 and <32 ug/m3 across 99% of simulations with concentrations ranging from 0-150 ug/m3). Finally, we discuss practical implications of our choice of modeling framework and next steps for improving performance. The machine learning models described here are not yet suitable replacements for traditional chemistry solvers but represent a step toward that goal.

IntroductionPhotonic packet switching fabrics are emerging as an attractive solution for ultra-high bandwidth, low switching latency interconnection within next -generation high-performance computing architectures. This is because lightwave networks are not limited by the conventional bottlenecks found in electronic systems, and have demonstrated the transmission of gargantuan volumes of data [1]. Furthermore, by utilizing intelligent routing procedures, the latency of such networks can be reduced to speed-of-light limitations. A topology for a switching fabric with minimal routing logic, designed specifically for photonic implementation, has been recently demonstrated and evaluated [2][3][4]. Termed the Data Vortex, this topology employs implicit synchronous signaling and routing, thereby eliminating the need for optical buffering.The predominant switch elements considered for optical packet networks have been semiconductor optical amplifiers (SOAs) due to their fast switching times, low switching power, and high extinction ratio [5]. Furthermore, in addition to switching, SOAs may be configured to re-amplify the signal and compensate for routing and transmission losses. The primary limitation of SOAs is the in -band signal noise introduced due to amplified spontaneous emissions (ASE). Crosstalk and other effects also contribute to an increase in the signal's bit error rate (BER) as it propagates through the switching nodes [6].The present work characterizes and simulates signal propagation through the physical layer of the Data Vortex switching fabric. Because traditional time-domain physical modeling techniques such as those based on rate equations, traveling wave analyses, and transfer matrix methods (viz. [6-8]) require precise knowledge of a myriad of physical parameter values, they are often quite difficult to implement. Instead of relying on physical analyses, we capture the essence of the signal distortions and perform the modeling from a signal-level phenomenological perspective. In order to construct the model, experimental data was collected from a number of actual Data Vortex switching nodes using optical packets with 10Gbit/sec payloads. It should also be noted that the methodology described herein can be applied to systems with differing architectures in a straightforward manner.