Weather forecasting foundation models (FMs) are increasingly fine-tuned to predict air quality, offering fast global pollution forecasts at lower computational cost than conventional chemical transport models. These FMs are typically trained on reanalysis data and generate forecasts through autoregressive rollout. They do not explicitly represent governing physical or chemical processes. Therefore, high forecast skill does not reveal whether a model has learned physical mechanisms or exploits statistical regularities in its training data. Here, we present the first study of what a FM fine-tuned for atmospheric chemistry has learned by examining Microsoft's Aurora model. We impose controlled chemical perturbations on its forecasts and test them against known photochemical relationships. We then examine the internal representations that generate these forecasts. We find that Aurora captures a first-order ozone response to reactive nitrogen but does not enforce the chemical constraints that a process-based model encodes. It generates chemically inconsistent combinations of related species and relaxes localized emission features such as wildfire plumes toward background. Internally, its representations remain largely organized around the meteorology inherited during pretraining, with little structure specific to chemistry. Using sparse autoencoders, we identify internal components that causally control the chemical forecast but do not map cleanly onto individual atmospheric processes. This work provides a framework for testing whether AI forecasting systems learn atmospheric chemistry from reanalysis data. As these models are increasingly positioned to inform environmental policy decisions, we argue that composition forecasts should also be judged by their internal mechanisms rather than by benchmark skill alone.

Conservation laws are time-invariant properties that constrain many physical systems. For systems of chemical reactions, the law of mass conservation constrains how atoms flow between chemical species. Chemical reaction networks can display emergent conservation not explained by mass conservation: these hidden symmetries arise instead from coupled kinetics. Kinetic invariants emerge when branching reactions with proportional rates cause species concentrations to evolve in lockstep. We detect emergent conservation in a simplified atmospheric chemical mechanism of ozone formation through a data-driven analysis of simulated concentrations, a result matching the theoretical kinetic explanation. Surveying 35 widely used atmospheric chemical mechanisms spanning five orders of magnitude in complexity, we discover emergent conservation in 15 mechanisms. Kinetic invariants constrain the intrinsic dimensionality of chemical systems: mechanisms with emergent conservation evolve in lower-dimensional spaces than their size suggests. Identifying emergent conservation can provide theoretical bounds for exact mechanism reduction and uncover kinetic symmetries in atmospheric chemistry.

Detailed knowledge of chemical processes in the atmosphere is key to our understanding of regional air pollution and global climate change. However, a complete description of all atmospheric chemical reactions is still out of reach. This necessitates the discovery of new reactions for improved predictability and process understanding. Here, we propose a data-driven, chemical kinetics-oriented approach for atmospheric chemical reaction discovery. Our approach leverages time series of species abundances and an incomplete chemical mechanism to predict the existence of new chemistry by “completing” the mechanism. Species abundances and the incomplete mechanism serve as inputs to a variant of graph neural networks known as graph autoencoders (GAEs). The GAE learns a low-dimensional representation of the chemical system to predict the existence of pairwise chemical interactions occurring between species. We assess our model using GEOS-Chem, a widely used atmospheric chemical mechanism that represents the complex set of chemical interactions in the atmosphere. Our reaction discovery model achieves high predictive performance (0.9085 mean AUC; 90.06% average precision) in recovering unseen reactions and outperforms other competitive baselines. The success of this method solidifies its promise in discovering unknown chemical reactions and warrants further application to additional atmospheric chemistry contexts.

Abstract Chemical cycling drives the production and loss of many important atmospheric constituents. The speed of atmospheric chemical cycling is a particularly valuable indicator for characterizing and measuring the effects of such cycles on oxidant chemistry, air quality, and climate. Here, we apply graph theoretical methods to explicitly quantify and analyze the characteristic timescales of gas‐phase chemical cycles in the troposphere and stratosphere, as simulated by the GEOS‐Chem chemical mechanism. We identify all two‐, three‐, and four‐reaction cycles in the mechanism and calculate a characteristic timescale for each individual cycle. We find that the speed of chemical cycling varies by orders of magnitude at any given location but tends to be faster in urban‐ and biogenically‐dominated chemical regions, and slower during the night. We further quantify the fraction of cycling that contains a rate‐determining step, and explicitly demonstrate the large potential for mechanisms to recycle oxidants like OH.

Computational models of atmospheric composition are not always physically consistent. For example, not all models respect fundamental conservation laws such as conservation of atoms in an interconnected chemical system. In well performing models, these nonphysical deviations are often ignored because they are frequently minor, and thus only need a small nudge to perfectly conserve mass. Here we introduce a method that anchors a prediction from any numerical model to physically consistent hard constraints, nudging concentrations to the nearest solution that respects the conservation laws. This closed-form model-agnostic correction uses a single matrix operation to minimally perturb the predicted concentrations to ensure that atoms are conserved to machine precision. To demonstrate this approach, we train a gradient boosting decision tree ensemble to emulate a small reference model of ozone photochemistry and test the effect of the correction on accurate but non-conservative predictions. The nudging approach minimally perturbs the already well-predicted results for most species, but decreases the accuracy of important oxidants, including radicals. We develop a weighted extension of this nudging approach that considers the uncertainty and magnitude of each species in the correction. This species-level weighting approach is essential to accurately predict important low concentration species such as radicals. We find that applying the uncertainty-weighted correction to the nonphysical predictions slightly improves overall accuracy, by nudging the predictions to a more likely mass-conserving solution.

Discovering conservation laws for a given dynamical system is important but challenging. In a theorist setup (differential equations and basis functions are both known), we propose the sparse invariant detector (SID), an algorithm that autodiscovers conservation laws from differential equations. Its algorithmic simplicity allows robustness and interpretability of the discovered conserved quantities. We show that SID is able to rediscover known and even discover new conservation laws in a variety of systems. For two examples in fluid mechanics and atmospheric chemistry, SID discovers 14 and 3 conserved quantities, respectively, where only 12 and 2 were previously known to domain experts.

Atmospheric simulationchambers continue to be indispensable toolsfor research in the atmospheric sciences. Insights from chamber studiesare integrated into atmospheric chemical transport models, which areused for science-informed policy decisions. However, a centralizeddata management and access infrastructure for their scientific productshad not been available in the United States and many parts of theworld. ICARUS (Integrated Chamber Atmospheric data Repository forUnified Science) is an open access, searchable, web-based infrastructurefor storing, sharing, discovering, and utilizing atmospheric chamberdata [https://icarus.ucdavis.edu]. ICARUS has two parts: a data intake portal and a search and discoveryportal. Data in ICARUS are curated, uniform, interactive, indexedon popular search engines, mirrored by other repositories, version-tracked,vocabulary-controlled, and citable. ICARUS hosts both legacy dataand new data in compliance with open access data mandates. Targeteddata discovery is available based on key experimental parameters,including organic reactants and mixtures that are managed using thePubChem chemical database, oxidant information, nitrogen oxide (NOx)content, alkylperoxy radical (RO2) fate, seed particleinformation, environmental conditions, and reaction categories. Adiscipline-specific repository such as ICARUS with high amounts ofmetadata works to support the evaluation and revision of atmosphericmodel mechanisms, intercomparison of data and models, and the developmentof new model frameworks that can have more predictive power in thecurrent and future atmosphere. The open accessibility and interactivenature of ICARUS data may also be useful for teaching, data mining,and training machine learning models.

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 dimensionality reduction method based on matrix factorization to find latent patterns in the VBS tracers that correspond to a smaller 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 data-driven 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 1.5–1.8. Concentrations remain numerically stable over model simulation times of 2 weeks, including simulations at higher spatial resolutions than the data-driven models were trained on. The reversible compression of VBS tracers enables detailed, process-based SOA representation in LOTOS-EUROS operational forecasts in a computationally efficient manner. Beyond this case study, the physically consistent data-driven 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.

Discovering conservation laws for a given dynamical system is important but challenging. In a theorist setup (differential equations and basis functions are both known), we propose the Sparse Invariant Detector (SID), an algorithm that auto-discovers conservation laws from differential equations. Its algorithmic simplicity allows robustness and interpretability of the discovered conserved quantities. We show that SID is able to rediscover known and even discover new conservation laws in a variety of systems. For two examples in fluid mechanics and atmospheric chemistry, SID discovers 14 and 3 conserved quantities, respectively, where only 12 and 2 were previously known to domain experts.

In section 2 you introduce two networks. One is a simple two layers dense neural network which you refer to as naïve. The other network, which is the focus of your work, features two architecture modifications with regards to the naïve network. One affects the input with the addition of bilinear terms correponding to the two bodies reactions in your chemical system. Another modification affects the output, resulting in your network effectivelly predicting the rate of each reaction in the second to last layer and subsequently assesing the change in concentration of each species in the last layer, according to stochiometric constraints.

Models of atmospheric phenomena provide insight into climate, air quality, and meteorology and provide a mechanism for understanding the effect of future emissions scenarios. To accurately represent atmospheric phenomena, these models consume vast quantities of computational resources. Machine learning (ML) techniques such as neural networks have the potential to emulate computationally intensive components of these models to reduce their computational burden. However, such ML surrogate models may lead to nonphysical predictions that are difficult to uncover. Here we present a neural network architecture that enforces conservation laws to numerical precision. Instead of simply predicting properties of interest, a physically interpretable hidden layer within the network predicts fluxes between properties which are subsequently related to the properties of interest. This approach is readily generalizable to physical processes where flux continuity is an essential governing equation. As an example application, we demonstrate our approach on a neural network surrogate model of photochemistry, trained to emulate a reference model that simulates formation and reaction of ozone. We design a physics-constrained neural network surrogate model of photochemistry using this approach and find that it conserves atoms as they flow between molecules while outperforming two other neural network architectures in terms of accuracy, physical consistency, and non-negativity of concentrations.

Large air quality models and large climate models simulate the physical and chemical properties of the ocean, land surface, and/or atmosphere to predict atmospheric composition, energy balance and the future of our planet. All of these models employ some form of operator splitting, also called the method of fractional steps, in their structure, which enables each physical or chemical process to be simulated in a separate operator or module within the overall model. In this structure, each of the modules calculates property changes for a fixed period of time; that is, property values are passed into the module, which calculates how they change for a period of time and then returns the new property values, all in round-robin between the various modules of the model. Some of these modules require the vast majority of the computer resources consumed by the entire model, so increasing their computational efficiency can either improve the model's computational performance, enable more realistic physical or chemical representations in the module, or a combination of these two. Recent efforts have attempted to replace these modules with ones that use machine learning tools to memorize the input–output relationships of the most time-consuming modules. One shortcoming of some of the original modules and their machine-learned replacements is lack of adherence to conservation principles that are essential to model performance. In this work, we derive a mathematical framework for machine-learned replacements that conserves properties – say mass, atoms, or energy – to machine precision. This framework can be used to develop machine-learned operator replacements in environmental models.

Taking into consideration all conceivable branching off at traffic nodal points, a formalized process was developed to determine the waiting time for non-steady traffic bound to parts, so that the traffic at nodal points can be described more realistically than by earlier processes, as an aid to balancing the requirements of environmental acceptability, economy and traffic quality.