Machine Learning-Based Parameterisation of Photolysis in GEOS-Chem - Version 14.2.2 - Masters of Science by Research

Aug 2, 2026

Upload by Oliver

Photolysis schemes are an essential but computationally expensive component of chemical transport models, where GEOS-Chem uses the Fast-JX formulation. This thesis develops a proof-of-concept machine-learning parameterisation for Fast-JX using an ensemble of XGBoost models. The trained models are integrated into the GEOS-Chem Fortran code and evaluated quantitatively against Fast-JX. It further assesses how photolysis prediction changes impact GEOSChem concentrations of key species such as O3 and NO2 during simulation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages88

Reading Time3 h 42 min

AccessFree to download

Introduction

Atmospheric Chemistry, Environmental Problems, and Chemical Transport Models

Photolysis

Different Photolysis Mechanisms

Research Aims

Theoretical Background

Machine Learning Foundations

Key Machine Learning Algorithms

XGBoost Algorithm over Random Forest Regression

Machine Learning in Atmospheric Chemistry

Original Photolysis Scheme in GEOS-Chem

Model Description (GEOS-Chem)

Data and Methodology

Training and Evaluation Data

Different Photolysis Rate Machine Learning Models

Model Training and Parameters

Objective Function and Regularisation

Implementation

Model Performance on Validation Data

Evaluation Metrics

The Model Statistics for Validation data

Feature Importance

Model performance when integrated into GEOS-Chem

Prediction of Photolysis Rates

Predictive Ability in the Vertical Columns

Concentration Calculations using the Predicted Rates

Discussion

Hardware and Software Optimisations

Limitations

Conclusions

Appendix

Machine Learning-Based Parameterisation of Photolysis in GEOS-Chem: Version 14.2.2 Marcus Eliot Brady Masters of Science by Research University of York Chemistry September, 2024 Abstract Photolysis schemes are an integral part of chemical transport models. However, they are computationally intensive. The chemical transport model GEOS-Chem uses the photolysis scheme Fast-JX, which takes a significant amount of time to run. Through the developments of machine learning, parameterised approaches for physical processes have become a common way of speeding up calculations. This study looks to develop a proof-of-concept approach for a machine learning-based parameterisation of the Fast-JX photolysis calculations using a collection of XGBoost models. The machine learning models were integrated into the GEOS-Chem Fortran code base and were quantitatively evaluated against the standard Fast-JX scheme. This work additionally determines the wider impact the photolysis predictions had on the GEOSChem simulation in regards to the calculated concentration of key components such as O 3 and NO2. Results show high accuracy for most species, with 103 out of 105 unique photolysis rates maintaining an R2 greater than 0 .95 throughout a six month simulation period. While the current implementation is minimally optimised, and hence computationally slower than Fast-JX, it successfully demonstrates that a machine learning parameterisation of photolysis rates is feasible in Fortran based chemical transport models and provides a foundation for future optimisations. Acknowledgments First and foremost, I would like to thank to thank my supervisor Prof. Mat Evans. I am additionally grateful to have made acquaintance with everyone in the Evans group: Ryan Pound, Amy Lees, Hansen Cao, Matthew Rowlinson, and Killian Murphy. All the support, ranging from learning Fortran and plotting guidance to the photos of various trips at the end of the Thursday slides, has been greatly appreciated over this year. You have all greatly helped me develop my skills as a researcher. It would be inappropriate to not thank my family for their continued support, if I mentioned everyone, it would be a very long section of text so I will just say I have an appreciation for all. However, a very special thank you to my Grandpa and my Grandma, whose generosity led me to this masters and hence paved the way for all future research that I will take. Aside from my family there are a couple names I must mention. I must thank Lily, for your discussions, patience, and support throughout the past year not just academically but personally-I’m incredibly grateful. Also Immy and Willow, for keeping my company during my writing period. For compute power, the Viking cluster was used during this project, which is a high performance compute facility provided by the University of York. I am grateful for computational support from the University of York, IT Services, and the Research IT team. Through everyone’s support and following the completion of this thesis I can now go on to the next steps of my research career of completing my PhD over the next four years at UCL using the skills I have built throughout this year. Declaration I, Marcus Brady, declare that this thesis is a presentation of original work and I am the sole author. This work has not previously been presented for a degree or other qualification at this University or elsewhere. All sources are acknowledged as references in the Bibliography. Contents Abstract i Acknowledgments ii Declaration iii Table of Contents v List of Tables vi List of Figures ix 1 Introduction 1 1.1 Atmospheric Chemistry, Environmental Problems, and Chemical Transport Models ......................................... 1 1.2 Photolysis ..................................... 2 1.3 Different Photolysis Mechanisms ......................... 3 1.4 Research Aims .................................. 4 2 Theoretical Backgroun

FAQs

Why is photolysis parameterisation important in GEOS-Chem?

Photolysis schemes are integral to chemical transport models but are computationally intensive. GEOS-Chem’s Fast-JX photolysis component significantly increases runtime, motivating faster parameterised alternatives.

How does the thesis implement the machine learning approach?

XGBoost models are trained to parameterise Fast-JX photolysis calculations and then integrated into the GEOS-Chem Fortran codebase. Model outputs are evaluated against the standard Fast-JX scheme using quantitative metrics.

How accurate are the predicted photolysis rates compared with Fast-JX?

Results show high accuracy for most species, with 103 out of 105 unique photolysis rates keeping an R2 above 0.95 throughout a six-month simulation period.

What impact does the new photolysis prediction have on GEOS-Chem concentrations?

The thesis evaluates how updated photolysis predictions affect simulated concentrations of key components, including O3 and NO2, using the GEOS-Chem results driven by the ML-predicted rates.

Aug 2, 2026

Upload by Oliver