EnsAI: An Emulator for Atmospheric Ammonia Ensembles

summary

The gist

EnsAI introduces an AI-based ensemble generation system for atmospheric chemical constituents that can run many orders of magnitude faster than traditional physics-based models When trained on an…

In short

EnsAI is an AI system using a U-Net neural network to generate atmospheric ammonia ensembles much faster than traditional physics models. Trained on existing data, EnsAI accurately reproduces meteorology-dependent features of the original ensemble, significantly reducing computational costs for chemical data assimilation and emissions inversion.

Key concepts

Ensemble Forecasting
This involves running an atmospheric model multiple times with slightly different starting conditions to represent uncertainty. It helps scientists understand the range of possible outcomes and how uncertain different variables are within a weather system.
EnsAI Architecture (U-Net)
EnsAI uses a specific type of deep learning network called a U-Net, which is excellent for mapping input data (like wind and temperature) to output perturbations. It learns the complex relationship between these meteorological inputs and how ammonia concentrations change.
Flow-Dependent Error Covariances
Traditional models use static error estimates, but EnsAI captures 'flow-dependent features,' meaning it understands how errors change based on the movement of air (wind and temperature). This allows for much more accurate error estimation during data assimilation.
Emissions Inversion
This is a process used to estimate unknown sources (like ammonia emissions) by comparing observed concentrations with model forecasts. EnsAI speeds up this process by providing highly accurate ensemble information, leading to faster and more reliable estimates of ammonia sources.

Terminology used across episodes

This episode discusses

The paper

EnsAI: An Emulator for Atmospheric Ammonia Ensembles · Read on arXiv

Environment and Climate Change Canada

Ensemble-based methods for data assimilation and emissions inversions are a popular way to encode flow-dependency within the model error covariance. While most ensemble methods do not require the use of an adjoint model, the need to repeatedly run a geophysical model to generate the ensemble can be a significant computational burden. In this paper, we introduce EnsAI, a new AI-based ensemble generation system for surface ammonia concentrations. When trained on an ensemble for ammonia generated by the GEM-MACH air quality model, it was shown that the ensembles produced by EnsAI can accurately reproduce the meteorology-dependent features of the original ensemble, while generating the ensemble 3,300 times faster than the original GEM-MACH ensemble. While EnsAI requires an upfront cost for generating the ensemble used for training, as well as the training itself, the long term computational savings can greatly exceed these initial computational costs. When used in an emissions inversion system, EnsAI produced similar inversion results to those in which the original GEM-MACH ensemble was used while using significantly less computational resources.

DOI: 10.5194/gmd-19-10283-2026

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "EnsAI: An Emulator for Atmospheric Ammonia Ensembles".

Mira: Here's the gist of the paper. EnsAI introduces an AI-based ensemble generation system for atmospheric chemical constituents that can run many orders of magnitude faster than traditional physics-based models When trained on an existing ensemble for ammonia generated by the GEM-MACH air quality model, it was shown that the ensembles produced by EnsAI can accurately reproduce the meteorology-dependent features of the original ensemble, while generating the ensemble 3,300 times faster than the original GEM-MACH ensemble Ensemble forecasting of atmospheric models has become an invaluable tool in atmospheric modeling over the past few decades An ensemble of atmospheric states has the ability to represent the uncertainty associated with an atmospheric model and can encode information about both the level of uncertainty of particular state variables as well as the correlation between the uncertainties of different variables The goal of an assimilation process is to produce what is know as the analysis, which is a combination of forecast (or background) and observational information in a way that is optimal (or close to optimal) according to some metric While having an accurate estimate of a model’s error covariance is of high importance in many applications, actually obtaining such an estimate is often difficult Challenges with Traditional Ensemble Generation A significant downside of using these methods is the computational cost of creating these ensembles, which often requires running an atmospheric model tens to hundreds of times for each time step The computational challenges associated with ensemble generation are compounded when applied to atmospheric chemistry models, which typically require more computational resources than NWP models When ensemble-based chemical data assimilation is feasible to run, it typically has a much smaller ensemble size than in NWP systems, often with less than a hundred members EnsAI Architecture and Training EnsAI uses a U-Net convolutional neural network to map ammonia emissions perturbations to ammonia surface concentration perturbations, using surface wind and temperature fields as auxiliary inputs to the neural network The input fields to EnsAI are displayed in the upper-left corner of Fig.

Kai: 3, which consists of the four input fields

δe, u, v, T: on the 256 × 256 horizontal grid Training was done on a NVIDIA A100 GPU and took approximately 18 days to train the network through 50 epochs Performance Comparison with Static Covariances The analysis showed that including flow-dependence in the error covariances has a considerable effect on the covariances and that EnsAI can encode much of the meteorological-dependent features of the original ensemble EnsAI produced very similar results to inversions performed using the GEM-MACH ensemble, being able to generate ensembles in a fraction of the time required for physics-based models The RMSE(∆e) values for the increments produced using the static covariances are between 33% and 70% larger than the RMSE values for EnsAI, with the exception of the January increment where RMSE(∆e) for the static covariances is only slightly higher than for EnsAI It was demonstrated that including only surface information in EnsAI was sufficient to reproduce most of the time-dependent features of the ensemble, thereby allowing EnsAI to produce accurate ensemble information for ammonia with much smaller memory requirements as compare to systems that require multiple vertical levels An ensemble-based emissions inversion system for ammonia was used as a test application for EnsAI, and it was demonstrated that EnsAI-based emissions inversions produced very similar results to inversions performed using the GEM-MACH ensemble The longer term goal is to expand EnsAI to more atmospheric chemical constituents to use in chemical data assimilation Code and Data Availability The source code for EnsAI (Sitwell, 2025) is available from https://doi.org/10.5281/zenodo.15342768 under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license GEM-MACH version 3.1.0 (ECCC, 2023), which was used to generate the training data sets, is available from https://doi.org/10.5281/zenodo.8346780 under the GNU Lesser General Public License The CrIS ammonia retrieval data use for evaluation is available at https://doi.org/10.5281/zenodo.18865369 under the Creative Commons Attribution 4.0 International license References The paper cites numerous works supporting its methodology and context, including those on ensemble Kalman filter, variational data assimilation, and machine learning weather forecasting The references include work by Sitwell et al., 2025, Ross N Bannister, Kaifeng Bi et al., 2023a, and others Appendix Details The appendix details the Ensemble-Variational Emissions Inversion method, where the cost function is minimized to find the optimal combination between a priori (background) and observational information The emissions inversion then consists of minimizing this cost function with respect to the control vector χ, followed by the transformation of the optimal control vector to the emissions (or concentration) increment given by Equation (13) The components that comprise the U-Net used for EnsAI are described in this appendix The decomposition of the correlation length defined in Equation (8) provides definitions for L

0: , L

1: , and L

2: The integral in Equation (16a) shows that at each angle θ, the integral over r is equal to the (one-sided) variance of ϱ in that particular direction The RMSE values for Lec

m: are smaller than those for the static covariances for all months, as seen in the three lower panels of Fig.

Mira: 11 The ACC(Lec

m: ) values with m ≤ 2 for the static covariances are close to zero for all cases, while the EnsAI values are above 0.5 for all cases with the exception of the m = 1 mode in January and February The RMSE(Lec

0: ) values for EnsAI are about 7 km less than the RMSE value for the static covariances, which corresponds to about a 10% larger RMSE value for the static covariances The ACC(σc) values throughout the test period show dramatic differences between EnsAI and the static covariances, with EnsAI having larger ACC values for all months The RMSE(σc) values for EnsAI are lower than the static covariance values for all months The RMSE(Lec

1: /Lec

0: ) taken over the whole domain are 0.07 to 0.14 less for EnsAI than for the static covariance, and between 0.04 and 0.

Title and authors: Mira: So this paper introduces something called EnsAI, which is this new AI-based system designed to generate ensembles for atmospheric chemical constituents. It’s interesting because it aims to run many orders of magnitude faster than traditional physics-based models when you need those ensembles.

Kai: Right, so instead of running a full model repeatedly for every ensemble member, this EnsAI system is trained on an existing ammonia ensemble—specifically one generated by the GEM-MACH air quality model—and then it can reproduce the meteorological features of that original set while being three thousand three hundred times faster at generating new ensembles.

Lev: three thousand three hundred times faster is a huge number for computational chemistry, Kai. That suggests we might be able to do emissions inversions much more quickly than current methods allow us to run them on real hardware, which is where the bottleneck usually lands.

Mira: And the core idea here is that EnsAI doesn't need to run the heavy physics model every time you want an ensemble for concentrations. It uses a U-Net convolutional neural network to map ammonia emissions perturbations directly to surface concentration perturbations, using surface wind and temperature fields as its auxiliary inputs.

Kai: So it’s essentially learning the relationship between what’s being emitted at the source and how that translates into the actual concentrations we measure on the ground, bypassing a lot of the traditional model steps.

Lev: That bypass is key because running a full atmospheric model for every ensemble member is what makes those ensembles so expensive in terms of memory and time, Lev. If you can get close to that fidelity with this neural network approach, it really changes how we think about data assimilation speed.

Mira: The paper highlights that when they compared EnsAI's results against static covariances—which are simpler error estimates—the RMSE values for the increments produced by those static covariances were between thirty three percent and seventy percent larger than the RMSE values produced by EnsAI.

Kai: That comparison shows that just using simple, non-flow-dependent error estimates doesn't capture much of what’s happening in the atmosphere, Mira. EnsAI manages to encode those meteorology dependent features that static models miss, which is a big deal for accuracy.

Title and authors: Lev: Capturing flow dependence in the error covariance is where things get tricky; it means the uncertainty isn't just a fixed number everywhere, but it changes based on where you are and what the wind conditions are.

Mira: Exactly. The paper shows that including this flow dependence in the error covariances has a considerable effect on those covariances, and EnsAI can encode much of those features when it generates ensembles.

Kai: They were able to produce results very similar to inversions performed using the GEM-MACH ensemble, but they did it in a fraction of the time required by physics-based models. That’s a massive speedup for an emissions inversion system.

Lev: If we look at this from an engineering standpoint, that speedup means we can run these complex chemical data assimilation systems much more often on the hardware we have available, which directly impacts how many inversions we can practically perform in a given timeframe.

Mira: The study also points out that including only surface information in EnsAI was enough to reproduce most of the time dependent features of the ensemble. This suggests that you don't need to model every vertical level or every atmospheric process just to get good results for ammonia concentrations.

Kai: That reduces the memory requirements significantly, which is another major win compared to systems that require modeling multiple vertical levels, Mira. It’s much more data efficient for the same accuracy.

Lev: So, it's not just about speed; it’s about efficiency in terms of what we need to store and process while still maintaining that necessary accuracy for the chemistry.

Mira: The test application they used was an ensemble-based emissions inversion system specifically for ammonia, and they demonstrated that EnsAI-based inversions produced results very similar to those from the GEM-MACH ensemble. This validates its capability in practice.

Kai: But I wonder what happens when we try to apply this to other chemical constituents, like methane or other pollutants? The paper mentions that another goal is developing an AI system flexible enough for other chemical constituents relevant for air quality.

Lev: That flexibility is the long term goal, right? We’re hoping this architecture can be adapted without starting from scratch every time we want to model a new constituent.

Title and authors: Mira: And the training data setup itself is important; they trained EnsAI on an existing ensemble for ammonia generated by GEM-MACH, which means you have a known, established baseline to start with.

Kai: But of course, there’s that upfront cost for generating the training ensemble and then the time it takes to train the neural network itself on that NVIDIA A100 GPU. It's not a plug-and-play solution right out of the box.

Lev: That initial computational cost is what we need to weigh against those long term savings, since you mentioned in page one that the long term computational savings can greatly exceed these initial costs.

Mira: And it’s worth noting that they were able to show improvements when they looked at flow-dependent error covariances versus static ones, showing EnsAI had smaller RMSE values for Lec zero than the static covariances for all months.

Kai: That reduction in error metrics, especially for the Lec zero mode over the whole domain, shows that capturing this flow dependence is genuinely helpful and not just theoretical.

Lev: And those RMSE differences you mentioned—about ten percent larger when using static covariances for Lec zero—that’s a concrete metric showing why we need to model those dependencies if we want to get tight error bounds.

Mira: So, the overall conclusion of EnsAI: it proves that this AI approach can generate ensemble information for ammonia with much smaller memory requirements than traditional systems, and its emissions inversions produce results comparable to using the GEM-MACH ensemble but with significantly less computational resource usage.

Kai: It’s an emulator for atmospheric ammonia ensembles, and it shows that we can get high-fidelity outputs without needing those massive model runs every single time.

Lev: Yeah, the implication is that we can accelerate the chemical data assimilation pipeline considerably if we can integrate these kinds of AI emulators into our standard workflows for air quality modeling.

Mira: For now, it seems like a really promising test case for emissions inversions, and I’m looking forward to seeing how they extend this technique to other chemical species in the future.

Kai: We’ll keep an eye on those extensions, because if this architecture works well for ammonia, it opens up the door for applying it to a much broader range of atmospheric constituents.

The paper's summary: Kai: So to recap, EnsAI is essentially an AI system that takes an existing ammonia ensemble from another model and can generate new ones three thousand three hundred times faster than the original physics model does, while still capturing the important meteorological details of that original set.

Mira: Right, so it’s not just about speed; it’s about getting this high level of fidelity for chemical data assimilation without having to run those incredibly expensive physics models repeatedly.

Lev: The paper shows that they use a U-Net convolutional neural network, and they train it on surface wind and temperature fields to map ammonia emissions perturbations directly to surface concentration perturbations.

Kai: That’s the setup, right? They’re using surface data as inputs for this neural network to predict how emissions affect the concentration on the ground.

Mira: And what I find interesting is how they handled the error covariance; they showed that including flow dependence in those error estimates makes a big difference in what EnsAI can encode.

Lev: That’s where it gets tricky because static covariances, which are simpler and easier to calculate, just don't capture all the necessary uncertainty information about how things move around.

Kai: The results were pretty impressive because EnsAI produced very similar results to inversions done using the full GEM-MACH ensemble, but it took a fraction of the time required by those physics-based models.

Mira: And they quantified that difference by showing that the root mean square error for the increments produced by static covariances was between thirty three percent and seventy percent larger than what EnsAI got.

Lev: That comparison is really telling; it means that when you use these simpler static error estimates, you end up with much bigger errors in your final concentration estimates.

Kai: It also highlights that just using surface information was enough to capture most of the time dependent features, which is a big win because it means you don't need to model every vertical level for this application.

Mira: So it’s proving that you can get accurate ensemble information with much smaller memory requirements compared to traditional systems that try to track every single atmospheric layer.

Lev: And the test application they used, which was an emissions inversion system for ammonia, showed that EnsAI produced results very similar to those from the GEM-MACH ensemble.

Kai: So it’s a functional emulator; it’s not just a theoretical exercise, it can actually produce usable outputs for chemical data assimilation.

Mira: The long term goal they mentioned is expanding this EnsAI architecture to cover more atmospheric chemical constituents, which would be the next big step for this kind of research.

Lev: That implies a path forward where we might be able to apply this fast emulation technique across a wider range of atmospheric science problems.

Kai: It really shows that we can trade some complexity in the modeling approach for massive gains in computational efficiency, which is always something I’m interested in when I’m working with these heavy simulations.

Mira: And it suggests that we don't necessarily need to run a full atmospheric model every time we want to do an emissions inversion for ammonia.

Lev: It definitely changes the cost equation; the overhead lands much more favorably when you move from running a full simulation repeatedly to using this kind of learned mapping.

Kai: So, it’s about building tools that let us access high-fidelity ensemble information without having to be bottlenecked by the computational power required for traditional modeling.

Mira: And we need to keep looking at how they handle those flow-dependent error covariances, because that’s where a lot of the real physical nuance is hiding.

Lev: I agree; if we can get that part working robustly on real hardware, it unlocks a whole new level of accuracy for these types of analyses.

The paper's improvements: Tom: So we’re talking about how EnsAI can be improved beyond just being fast, Mira, what are the authors suggesting to make it even better?

Kai: Well, they spent time looking at those error metrics again and found that focusing on flow dependence in the error covariances really helps things.

Mira: Right, so they’re saying that instead of just using a static picture of uncertainty everywhere, you need to let the model account for how uncertainty changes based on the wind and temperature fields.

Lev: That means incorporating this flow-dependent feature into the emissions inversion process is crucial for getting tighter error bounds.

Kai: And they showed that when they used this approach, their root mean square errors for Lec zero were about seven kilometers smaller than what they got from static covariances, which is a big reduction.

Mira: That reduction in error across all months is what the paper really needs to emphasize; it shows that modeling the flow dependence isn't just theoretical refinement.

Lev: And when you look at the other correlation modes, like Lec one and Lec two, EnsAI actually produced smaller root mean square errors for those too, which is good because it means the spatial correlation capture is better.

Kai: So they’re suggesting that this method is more robust because it handles these complex spatial dependencies better than simpler models.

Mira: And they also touched on how they trained the network; using only surface information turned out to be sufficient to get most of the time dependent features right, which simplifies things for deployment.

Lev: That’s a practical point; it means you don't need to feed the AI every single vertical layer if you only care about what’s happening at the surface.

Kai: The main thing they are pushing for is expanding this concept to other chemical constituents, which opens up a whole new avenue for using this fast AI emulators.

Mira: Exactly; it shows that the architecture itself is flexible enough to handle different types of chemical data assimilation problems if you adjust the training setup correctly.

Lev: So it’s not just about ammonia anymore; it’s about building a versatile emulator framework for various atmospheric chemistry challenges.

Conclusion: Kai: So to wrap up, EnsAI is an emulator for atmospheric ammonia ensembles that uses AI to generate those ensembles three thousand three hundred times faster than traditional models while still capturing the important meteorological details of the original set.

Mira: It really shows that we can get high fidelity output for chemical data assimilation without needing those massive physics model runs repeatedly.

Lev: That speedup is significant because it means the computational cost of running these chemical data assimilation systems drops a lot when you use this kind of AI approach.

Kai: The implication is that we can run these complex simulations much more often on the hardware we have available, which directly impacts how many inversions we can practically perform in a given timeframe.

Mira: And the accuracy improvements, especially with those flow-dependent error covariances, mean that the results are actually better than what you’d get from simpler methods.

Lev: From an engineering standpoint, if this method can be scaled up to handle more chemical constituents, it could significantly lower the barrier for applying these advanced inversion techniques in real atmospheric science applications.

Kai: So that's what we saw with the paper "EnsAI: An Emulator for Atmospheric Ammonia Ensembles".

Mira: It’s a solid step forward because it proves that learning how to emulate an ensemble can work really well for atmospheric chemistry.

Lev: I think the real challenge now is getting this system to run on real hardware reliably at that speed without introducing too much noise or error.

More episodes

← Home