Machine Learning
The TOMCAT chemical transport model (CTM) is being used to create novel long-term datasets of various atmospheric trace gases, including ozone, methane, nitrous oxide and the hydroxyl radical, by combining model simulations with satellite observations. These datasets are crucial for understanding the long-term changes and variability of these gases in the stratosphere. The primary goal of these applications of machine learning (ML) is to produce consistent, gap-free datasets that overcome the limitations of individual satellite records, which often have shorter durations and may contain biases due to differences in measurement methods, sampling patterns, and retrieval algorithms. The datasets can be used to evaluate models and as reference fields in satellite retrieval algorithms.
We have combined TOMCAT with satellite observations in two ML approaches: (1) XGBoost to bias-correct TOMCAT output using (sparse) satellite observations and (2) Graph Neural Network with Transformers to create improved, denoised and gap-filled dataset with model-informed constraints. These two approaches are described below.
Our aim is to provide valuable resources for the scientific community to improve understanding of stratospheric processes and climate change. Suggestions for further species of interest are welcome.
XGBoost and ML-TOMCAT / TCOM Datasets
To address these limitations, a machine-learning (ML) approach is employed, specifically using the Random Forest/XGBoost algorithms to correct biases in the TOMCAT model outputs using available satellite measurements. For example, in the creation of the ML-TOMCAT ozone dataset, the random forest (RF) method is used to merge satellite data with the TOMCAT model output. The ML algorithm is trained using satellite data during periods of good temporal sampling. The trained model is then used to simulate differences for the entire time period. The ML approach allows for the reconstruction of long-term vertical profile data for key stratospheric species.
In effect, the ML methods are used as a computationally cheap and easy-to-apply form of data assimilation. The following publications describe the ML-TOMCAT methodology and its use in constructing long-term datasets:
- Dhomse et al. (2021) (ML-TOMCAT): This paper describes the creation of a long-term ozone profile dataset by combining the TOMCAT CTM with a random forest ensemble learning method, using the Stratospheric Water and OzOne Satellite Homogenized (SWOOSH) dataset for training.
- Dhomse and Chipperfield (2023) (TCOM): This paper discusses the methodology for creating long-term methane and nitrous oxide profile data sets by applying corrections to TOMCAT output based on profile measurements from HALOE and ACE-FTS.
- Dhomse and Chipperfield (2026) (TCOM, v2): Update TCOM method applied to CFC-11 and CFC-12.
The datasets created using the TOMCAT model and ML techniques are publicly available on Zenodo. A summary of updated links is provided on the TOMCAT model data pages on this site (https://tomcat.leeds.ac.uk/tomcat/tcom/), or links are provided in the publications.

TCOM: Merging satellite observations and TOMCAT output
DrCAT - Enhanced OH Dataset from Microwave Limb Sounder
The hydroxyl (OH) radical is a key species in stratospheric chemistry, driving ozone loss and the coupling of other chemical families. However, global OH observations are very sparse, with a critical gap in satellite-based observations since 2009. Although alternative chemical methods attempt to mitigate such gaps, they have limitations. DRCAT (Yin et al., Science Advances, 2026) is a chemistry-informed deep learning network to predict stratospheric OH profiles using coincident satellite observations of chemically related species. DRCAT provides reconstruction of a continuous and reliable global stratospheric OH dataset from late 2004 to the present, outperforming both TOMCAT itself and an observation-constrained steady-state approximation. Moreover, DRCAT successfully predicts anomalous OH enhancements after the 2022 Hunga volcanic eruption, showing generalizability to extreme conditions outside of the training regime. DRCAT thus provides a robust pathway to fill OH observational voids. Beyond OH, DRCAT offers a scalable framework for similar reconstruction of other key short-lived species.

Zonal mean OH from MLS observations, machine learning, simple steady-state and TOMCAT.
The code and tools for the DRCAT OH study are available on the TOMCAT GitHub pages.
Publications
Dhomse, S.S., Arosio, C., Feng, W., Rozanov, A., Weber, M., and Chipperfield, M.P.: ML-TOMCAT: machine-learning-based satellite-corrected global stratospheric ozone profile data set from a chemical transport model, Earth Syst. Sci. Data, 13, 5711–5729, https://doi.org/10.5194/essd-13-5711-2021, 2021.
Dhomse, S.S. and Chipperfield, M.P.: Using machine learning to construct TOMCAT model and occultation measurement-based stratospheric methane (TCOM-CH4) and nitrous oxide (TCOM-N2O) profile data sets, Earth Syst. Sci. Data, 15, 5105–5120, https://doi.org/10.5194/essd-15-5105-2023, 2023.
Dhomse, S.S. and Chipperfield, M.P.: TCOM-CFC11 and TCOM-CFC12: a gap-free, observationally constrained global dataset of stratospheric CFC-11 and CFC-12 profiles (v2.0), Earth Syst. Sci. Data, 18, 6171–6189, https://doi.org/10.5194/essd-18-6171-2026, 2026.
Yin, W., C. Zhou, W. Feng, S.S. Dhomse, R.J. Pope, L. Millán, and M.P. Chipperfield, A chemistry-informed deep learning network for mitigating the stratospheric OH data gap,
Science Advances, 12, doi:10.1126/sciadv.aee2319, 2026.
