The short-term forecasting of air pollutants including ozone, i.e. predictions of expected concentrations over 1–4 d, is relevant for public health and scientific questions . State-of-the-art air quality forecasts, typically on the timescales of hours to days and up to a few days ahead, are based on the output of numerical chemical transport models (CTMs) . These models may be run at higher spatial resolution for the area of interest , in order to better represent processes controlling air pollution at the local level, and may be post-processed to more accurately represent observations . As with other air pollutants, notably PM_2.5 , ML is increasingly being directly applied to the task of short-term, ground-level ozone forecasting, and to bias-correct existing air quality forecasting systems with considerable success. The availability of large and growing observational datasets has facilitated these advances . However, forecasting ozone concentrations as time series with ML comes with significant challenges: forecasting ozone is a spatiotemporal problem, and ozone is controlled by processes of varying spatial and temporal scales as shown in Fig. .

Many short-term forecasting studies using ML have focused on forecasting only at selected observational stations, using observed ozone and additional chemical species, and meteorological variables where they are available from individual stations or external datasets . Furthermore, since it is difficult to downscale a relatively coarse CTM at specific locations, using time series data from a particular station is an attractive way to make predictions at particular locations. However, approaches of this kind do not necessarily provide ozone forecasts across all locations that may be of interest, as a gridded model product model might.

As with other fields, the advances in ML-based ozone forecasting have been pushed by developments on two axes – first, increasing quantities of data and second, larger models with more appropriate inductive biases. The field has a long history (see Fig. ), with studies being published even during the most recent artificial intelligence (AI) “winter”, beginning with a feed-forward neural network (NN) in 1996 . illustrated that a NN could be used to forecast ozone at eight stations in the USA. This was followed by further feed-forward NN approaches, often with datasets drawn from a single location or city . Neural methods were typically evaluated in comparison with (autoregressive) regression models, often finding that NNs were better able to forecast ozone concentrations and extrema on test data , although the improvement was often only marginal. Alongside the successes of feed-forward NN architectures, other work drew attention to methods seen to be more interpretable, such as fuzzy logic systems and regression trees . Further work leveraged methodological advances in ML architectures designed for temporal data, including the use of recurrent neural networks (RNNs) and convolutional neural networks (CNNs) to account for lagged relationships in the time series data . Recent work has combined architectures to model the relationships that control ozone, including combining components such as transformers and CNNs to account for the temporal and spatial information relevant to forecasting . However, datasets have typically been limited to single countries or cities, due to the lack of a combined database of station measurements. The introduction of the TOAR surface database and the TOAR-II database have facilitated recent studies on data drawn from multiple countries . The importance of the curation of large datasets for scientific progress in ML is highlighted in the Outlook section.

Increasingly, more complex architectures are being used to enhance the accuracy of ozone forecasts, and more data are being included as input to the models. The inputs that are relevant to the physical drivers of ozone concentrations, such as past observations of ozone and covariates, and nearby covariates, reflect processes that control ozone observations, and feasibly contribute to improved ozone forecasting and infilling. Recently, methods on the scale of the ML architectures and data used for weather forecasting have been transferred to ozone forecasting by leveraging very large datasets and models . In weather studies there is work on forecasting at observation stations using these methods, and transferring these methods to forecast ozone at ground-level stations is feasible .

Global modeling of atmospheric chemistry is a grand computational challenge due to the high dimensionality of coupled chemical species, the nonlinearity and numerical stiffness of solving chemical mechanisms, and interactions with transport on all scales. The inclusion of comprehensive atmospheric chemistry in Earth system models (ESMs), which simulate the interactions between the atmosphere, oceans, land surface, and biosphere, is a priority science frontier . Atmospheric chemical mechanisms are typically implemented in CTMs, which focus on the distribution and chemical evolution of species in the atmosphere. For some applications, chemistry-climate models (CCMs) may also couple chemical processes with climate dynamics, allowing feedback between chemistry and climate. Current atmospheric chemistry models integrate the coupled chemical kinetic equations for mechanism species over model time steps using high-order implicit numerical solvers, but these solvers are computationally expensive and often dominate the cost of an atmospheric simulation . Such costs put the inclusion of atmospheric chemistry in tension with other computationally intensive ESM/CCM priorities such as increased spatial resolution and ensemble simulations. The current slowdown in the rate of increase in the speed of computer CPUs – the “end of Moore's law” – underscores the need for computationally efficient approaches .

Chemical solvers in atmospheric models compute the local evolution of species concentrations over a chemical time step that may range from minutes to hours depending on the model . The chemical mechanisms used in regional to global atmospheric models and ESMs typically include O(100) coupled species with chemical lifetimes ranging from less than a second to much larger than the model time step. High-order implicit solvers can integrate this system of stiff coupled differential equations with high accuracy and fast implementations of these schemes are available, but they are still extremely costly for atmospheric models. Atmospheric models may combat that cost by decreasing the size of the chemical mechanism, breaking down the stiffness of the problem, or using lower-order approximations. However, these methods rarely achieve a speedup of more than a factor of two and sometimes lead to loss of accuracy in the model results. As a consequence, these computational barriers limit the ability for high-resolution simulations, prevent detailed uncertainty analyses, and complicate the coupling of atmospheric chemistry into CCMs/ESMs for long-term climate simulations without significant compute resources. ML methods could be transformative in this area for both reducing the cost of an atmospheric chemistry simulation and facilitating their incorporation into ESMs. ML methods seem well-suited to replace chemical solvers in atmospheric models because the chemical computation is very repetitive, involving the integration of similar conditions in neighboring grid cells and successive time steps. However, the large number of coupled species brings a “curse of dimensionality” to the problem, and ML methods have no check on error growth, unlike in standard chemical solvers where errors are dampened by the negative response to perturbations (Le Chatelier’s principle).

Largely, ML methods in atmospheric chemistry modeling currently involve emulating model components to improve model parameterizations, reduce computational bottlenecks, and create simplified, reduced-order models. Here, emulation refers to an ML model reproducing the same calculations as a component of a complex physical or simulated system for a set of inputs. There exists a growing number of studies forecasting ozone on short- (hourly) and longer-term timescales, spanning from city- to regional-level spatial scales. However, few of these studies have been implemented in operational settings (i.e., within CTMs, CCMs) to offer insight beyond that of traditional model-to-observation comparison methods.

Satellite measurements provide detailed information on the spatiotemporal distribution of atmospheric composition and related parameters, such as those associated with surface air quality. Satellite measurements have greater spatial and temporal coverage compared to in-situ observations and they can fill the gaps in those sparse distributions, particularly in remote areas where in-situ observations are not available.

Over the past few decades, multiple satellites have been launched to measure total ozone columns. However, total column measurements cannot be used to provide insight into near-surface ozone because the amount of stratospheric ozone is much larger than the amount of tropospheric ozone. Tropospheric ozone information has been directly retrieved using measurements from nadir-viewing thermal infrared (TIR) sounders, such as the Tropospheric Emission Spectrometer (TES) and the Infrared Atmospheric Sounding Interferometer (IASI) , and by combining measurements from both ultraviolet (UV) and visible (VIS) wavelengths by the Tropospheric Emissions: Monitoring of Pollution instrument . In addition, the limb-nadir matching method employs stratospheric ozone data from limb-viewing measurements, such as those from the Microwave Limb Sounder (MLS), to derive tropospheric columns from observed total columns . Recently, multispectral satellite approaches, such as IASI and the Global Ozone Monitoring Experiment (GOME) 2 and TES and the Ozone Monitoring Instrument (OMI) , have been implemented to derive tropospheric ozone profiles with increased sensitivity to the lower troposphere.

Nevertheless, satellite observations of ozone are still limited in spatial, temporal, and vertical resolution and are not sufficiently sensitive to ground surface levels. On the other hand, measurements of precursors, such as NO₂ and CH₂O from OMI, GOME-2, the Tropospheric Monitoring Instrument (TROPOMI), and the Ozone Mapping and Profiler Suite, have provided unprecedented information to assess the formation processes and surface concentrations of pollutants such as ozone and aerosols. Despite these advancements, technical challenges remain in accurately assessing near-surface air pollutant concentrations from satellite observations of precursors. ML techniques can be used to fill the gaps in the information available from satellite observations and to improve the estimation of surface air pollutants.

ML has been widely used in satellite applications, especially in remote sensing imagery in the past and is becoming more widely applied to atmospheric composition data. ML has been applied to satellite observations in two main categories: (1) to generate atmospheric concentration retrievals and blend multi-satellite products, and (2) to fill gaps in observational information, including surface concentrations and emissions estimates.

Central to the success of ML modeling efforts and their utility are the choice of datasets and workflows, i.e., ML model choice and training methods. As noted above, in the field of air pollution and atmospheric composition research, the use of ML is hampered by the absence of benchmark datasets suitable for training different model types with varying sizes and complexity. Such well-defined benchmarks including datasets, training objectives, evaluation scores, and baseline models have been instrumental for the rapid development of ML models in other fields . In particular, WeatherBench and WeatherBench2 have been key factors driving the transformation of ML weather forecasting between 2022 and 2024. A similar dataset to perform the same function for ozone forecasting would allow the robust comparison of different methods, and may guide the field towards more accurate models. Careful curation and data fusion of the TOAR surface ozone database with other relevant datasets might provide a robust and representative benchmark dataset, building on existing work . Ultimately, a lack of sufficient surface observations will impact the study of air quality and downstream impacts on health or vegetation. Therefore improving data coverage over poorly monitored areas remains a priority .

The breadth of information available in datasets like TOAR, GHOST and reanalysis products like Copernicus Atmosphere Monitoring Service (CAMS) is vast, but these products provide significant challenges to the development of ML models for ozone due to heterogeneous data formats and lack of succinct documentation that focuses on the use of such data for ML applications. It may be that ML methods can also be used for infilling missing data (e.g. cloud-filtered satellite data, gaps in in-situ observations) for meteorological variables and ozone . Overall, there is a clear need for a harmonized benchmark dataset(s) for ozone to further enable ML models to be developed. These should (as much as possible) follow the vision outlined by and the principles defined by .

With regards to model choice and development, it is worth noting that, in contrast to CTMs and other methods of simulation, ML models do not a priori require simulation, outputting and aggregating of high-resolution time series data to generate predictions for relevant ozone metrics. Instead, ML models can be trained to directly generate forecasts of these metrics (see Sect. ). In this regard, it is necessary that ozone ML benchmarks should, where appropriate, include target objectives both for forecasting concentrations and for forecasting (a set of) aggregate ozone metrics. See and for a more detailed discussion on relevant ozone metrics.

With regard to ML model training, there is a wide array of data-splitting approaches that can answer subtly related scientific questions. For forecasting it is common practice to divide the data temporally such that the training data completely precedes the testing data, such as using the last few years of a longitudinal dataset for testing. This is commonly recommended in benchmarking studies . We emphasize that while these procedures likely make intuitive sense, they do not match the default setting in ML packages , such as scikit-learn , where the default cross-validation procedure will randomly split over individual data instances rather than over spatial blocks or temporal blocks. Without using these correct procedures, performance will be overestimated and may not reflect real performance when deployed. In practice, it remains a challenge to carefully define and document the data selection and splitting procedures and adapt them to the scientific problem at hand.

These challenges of data selection and splitting become particularly relevant when looking into climate timescales. Not only can this cause out-of-distribution samples of model input data (for example higher temperatures), but climate change may also affect atmospheric chemical and physical processes so that the mapping between inputs and outputs may drift. This problem is known as “concept drift”, where ML predictions become less accurate over time, which can arise from a non-representative training dataset or an ML model that lacks expressiveness, for example by being unable to extrapolate effectively beyond the bounds of the training data. For the latter, tree-based ML models especially are poor with respect to extremes and outliers. Here, model architecture may play a role. Exploring generative AI models, such as Generative Adversarial Networks (GANs) and transformer models, holds promise for the next generation of ML-based atmospheric models. These newer ML architectures can generate more internally consistent dynamics and require less training data than classical CNNs, while also demonstrating improved accuracy and stability over time.

In addition to appropriate handling of training data, ensuring the trained model is as generally useful as possible, both in and out of sample, remains an enduring challenge. As is common in ML tasks, models trained on data from one geographical region may not necessarily transfer to another region, even when the underlying task and physics remain the same. This limitation often arises from variations in spurious features or unobserved variables specific to each domain, or differences in emissions and climate in different regions. Many approaches in the ML literature seek to improve the performance of ML models across domains, or under domain shifts, which are yet to be used for ozone forecasting , while recent studies suggest that large-scale weather forecasting models may generalize to unseen conditions and perturbations . generalization is particularly important for the use of ML models trained in high-data domains and then deployed in low-data domains. In this context, it may be useful to exploit the benefits of probabilistic forecasting (see following section), using models that report uncertainty in unfamiliar domains.

In the context of observational data, ML is increasingly used to derive or enhance geophysical variables from satellite measurements (see Sect. 4). However, a conceptual and practical challenge lies in the appropriate application of these ML-derived products across spatio-temporal scales or atmospheric regimes . There is a need to evaluate and document the scale-dependence of such products, and to guide their use in downstream modeling and analysis applications accordingly.

A relevant application for ozone forecasting and modeling is the study of extremes, including both accurate forecasting and attribution. Extreme ozone concentrations or fluxes can have a large impact on health and vegetation, and are also referred to as low-likelihood high-impact events. By definition, extreme events occur rarely and are hence challenging to accurately represent. There has been some work on approaches to weight extremes more during model training . The ability of models to represent extremes is also an important metric that can be used to evaluate the quality of these models. Extremes can thus play a role for uncertainty quantification of the predictive performance of the models (e.g., important for forecast emulators and assessing ML performance on the extremes), where one can distinguish between epistemic (systematic) and aleatoric (statistical) performance. This connects with an increasingly recognized need to evaluate performance in more rigorous and consistent ways: including the development of new benchmark datasets, diagnostics, and metrics (see above). Progress on ML evaluation include causal evaluation (process-oriented approach) and eXplainable AI (xAI), for understanding (in)consistencies of the ML algorithms with physical processes (in other words: whether accurate answers are found for the right reasons). However, such methods are generally only applicable to relatively small-scale ML models. Dynamical tests and counterfactual experiments provide a means to test the credibility of large ML models .

Forecasting of potential extreme events is particularly challenging because these events are beyond the typical ozone variability, and naturally, extreme events are rarely and infrequently represented in data. For data-driven models this challenge is further exacerbated by (1) the need to forecast not only the presence of threshold exceedances, but also the intensity and duration of extreme events; and (2) ozone extreme events are often related to other anomalous mechanisms, such as heatwaves and wildfires, which are difficult to take into account based on limited extreme information, also due to the fact that ozone responses to these mechanisms are heterogeneous. Although the extreme value theory is widely adopted, its limitations are frequently acknowledged, including the IID (independent and identically distributed) assumption and independence between extreme and non-extreme events. On the other hand, approaches based on probabilistic forecasting may better characterize the uncertainty and likelihood of extreme events. One solution may be to use metrics and data scenarios to evaluate performance under different types of evaluation scenarios, taking advantage of evaluation metrics in weather forecasting which have been studied extensively. For example, if a key consideration is the ability of a forecasting model to capture extreme events, then metrics that capture relevant performance explicitly on those events should be used. This allows for robust comparison of both the existing and novel models on both traditional metrics and metrics focused on extreme event prediction to more comprehensively evaluate model performance. Evaluation of extreme events is limited in the literature, with recent studies highlighting the lower accuracy of ML models when forecasting spring and summertime ozone concentrations . In addition to helping with forecasting extreme events, probabilistic forecasting more generally provides a number of advantages compared to the deterministic forecasting methods that are currently more common, as outlined in . Furthermore, ML weather forecasting models are increasingly adopting probabilistic and diffusion-based architectures that are able to produce sharp forecasts and uncertainty estimates. This is a promising line of work, however, these ML architectures may be challenging to implement for ozone forecasts due to the uncertainty driven by the meteorological fields themselves.

Interpreting and explaining ML models used to study ozone remains difficult. While these two terms are often used interchangeably, for this article we follow the distinction that interpretability focuses on designing and exploring models that are transparent and have comprehensible internal data transformations, whereas explainability methods focus on post-hoc explanations of how black-box models are working . Models that are directly and trivially interpretable, such as multiple linear regression, are typically not the most performant, and in the high-data regime, the most performant ML models are typically variants on deep NNs that are difficult to interpret or explain. There is some literature that explores whether PINNs provide more interpretability. For example, efforts are underway to enhance the interpretability of ML models in atmospheric sciences by incorporating or diagnosing conservation priorities such as mass and stoichiometry . Additionally, neural operators are being employed to learn the solution operators of ODEs/PDEs from the chemical training data . However, while incorporating chemistry and physics constraints has been shown to increase interpretability, there is no guarantee that these methods will improve the stability of the ML model over time . Often, there is a trade-off between interpretability and ML model accuracy, especially with more complex models . While methods to interpret and explain NNs more generally have been studied widely, mechanistic interpretability of NNs is a challenging task , and only a limited range of XAI methods have been tested with ML methods developed for ozone forecasting, often focused on sensitivity approaches which look at the post-hoc explanations where the inputs to models are perturbed to see how predictions change . Recent studies have investigated the importance of model input parameters through bootstrapping, i.e. random perturbations of individual inputs . Input data perturbation experiments are also possible and informative for very large models as, for example, demonstrated by for the Pangu-Weather forecast model. Furthermore, ML approaches are increasingly employed to develop end-to-end models that process raw input data (e.g., emissions, meteorological fields) and directly predict outputs such as ozone concentrations. While end-to-end models bypass the challenges of emulating individual components, which are less prone to short-term instabilities and operator splitting issues, they also limit the ability to track uncertainty metrics tied to physical parameters and processes.

Modeling ozone using ML proves challenging due to the multitude of sources driving model error (emissions, chemistry, transport, deposition) and the nonlinear response of ozone to these sources. Parameter tuning an appropriate ozone ML model for a complex, high-dimensional parameter space is possible given large computational resources and adaptive learning on pre-defined metrics. However, such an approach is largely inefficient given that atmospheric chemistry data lies on relatively low-dimensional manifolds with respect to the possible input parameter space. That is, many ozone-related relationships are structured with individual signals often being sparse and low-rank. Here, domain knowledge from atmospheric chemistry can help identify the optimal training dataset and define meaningful loss functions and targeted timescales (Fig. 1) for the ML model problem. In particular, domain knowledge of chemical and physical processes can help explain errors in ML models at short time scales versus long time scales. For example, ML models of atmospheric chemistry tend to predict well fast chemical processes (e.g., seconds to days) but diverge over longer time scales (e.g., months to years) . In addition, knowledge of slow chemical processes, such as the role of peroxyacetyl nitrate (PAN) decomposition for ozone formation over polluted and/or remote areas, may help define appropriate training targets for ML models. An emphasis should be placed on emulating chemistry on longer timescales (>1 year) as issues of long-range stability are more challenging than shorter-term accuracy, and are a necessity for inclusion into CCMs and ESMs.

On the other hand, a heightened focus on domain knowledge may unintentionally limit the potential of ML models. Atmospheric chemists typically leverage well-established relationships of the chemical system, such as NO_x-limited vs. VOC-limited regimes, which are easily uncovered by linear regression or principal component analysis. By invoking such a strong prior assumption, it may impose constraints that hinder an ML model's ability to learn more complex, non-obvious interactions within the data. This bias toward known relationships risks overlooking patterns that could be hidden in the chemical state space that may promote greater accuracy and stability over longer time scales. Striking a balance between leveraging domain expertise and allowing ML models the flexibility to explore complex dynamics is essential for advancing the predictive capability of ozone modeling.

Although an open data infrastructure such as the TOAR-II database gives the impression of low barriers to data access, this might in fact not be true for everyone. Poor internet connectivity from developing countries may limit researchers from retrieving data and subsequently running a computationally demanding model . Furthermore, not all possible data providers agree with sending their data to an open access database, which is one important factor that limits global coverage of surface measurement data. The increasing resolution of satellite products and models is often considered to be an improvement, but the larger data size can complicate the processing and analysis of data for some researchers . Some data services require a registration and compliance with data use policies, which could conflict with institutional policies of researchers or exacerbate language barriers that non-native English researchers can experience. Finally, whereas advanced APIs can be ideal for technically skilled researchers and allow for reproducible workflows, they might hinder less technical researchers or policy makers that want to explore data sets.

In particular to developing nations, which may not have the economic ability to acquire high-resolution satellite products outside of those freely-available, it is imperative to develop high-quality, globally generalizable solutions to ozone modeling. Data hosting platforms like Google Earth Engine (GEE) enable users to freely access global data relevant for ozone modeling studies, ranging from land-use information from MODIS to human modification data from VIIRS nighttime lights , Gridded Population of the World , and more. Recent work by investigated the ability to detect spatially resolved ozone pollution trends using time-series Sentinel-5 imagery from GEE, highlighting the quality of spatial distribution and accuracy available an open-source product and platform. This demonstrates the necessity to co-design data services and their hosting platforms to provide efficient and performant access to high-quality, well-documented data.