Data-Driven Modelling of Earth’s Magnetic Field Prediction
This project is currently being undertaken by Alya Anesya, a Bachelor of Engineering Technology (Digitalization and Automation) student at Universiti Teknologi MARA, under my supervision. The project investigates the use of data-driven techniques to predict rapid variations in the Earth’s magnetic field, represented by dB/dt, an important indicator in studies of geomagnetically induced currents. Such variations may affect power networks, pipelines and communication infrastructure.
Using one-minute SuperMAG data from five geomagnetic stations; ABG, GAN, GUA, HYB and PHU, the study compares Ridge Regression, XGBoost, Moving Average and Exponential Smoothing for one-step-ahead prediction. Model performance is evaluated using MAE, RMSE, MAPE and R² while preserving the chronological order of the time-series data. A Streamlit dashboard is also being developed to automate data preparation, model comparison, visualisation and result export. Overall, the project provides an accessible framework for applying machine learning and time-series analysis to space-weather monitoring and geomagnetic research.
2. Automated Dashboard for Geomagnetic Time-Series Data Imputation
This project is currently being undertaken by Nur Syazwani, a Bachelor of Engineering Technology (Digitalization and Automation) student at Universiti Teknologi MARA, under my supervision. The project focuses on reconstructing missing geomagnetic observations, which may occur because of sensor faults, communication failures, equipment maintenance or data-transmission interruptions. Incomplete data can reduce the reliability of geomagnetic analysis and affect derived parameters such as the horizontal magnetic field and its rate of change, dB/dt.
Using SuperMAG data from the ABK, CYG and HYB stations, the study evaluates four imputation methods: Linear Interpolation, K-Nearest Neighbours, PCHIP, and Forward Fill with Backward Fill. Their performance is tested under MCAR, MAR and MNAR missing-data conditions at different missing-data percentages and assessed using MAE and RMSE. The project also develops two interactive dashboards: a Developer Dashboard for comparing imputation methods and a User Dashboard for automatically reconstructing incomplete geomagnetic datasets. The resulting system aims to provide a practical and user-friendly tool for supporting geomagnetic analysis and future space-weather research.
Suitable for Bachelor’s or Master’s degree candidates
This project aims to develop a machine-learning system capable of classifying geomagnetic conditions as quiet, moderate or disturbed using ground-magnetometer data. Relevant features may include magnetic-field variations, dB/dt and selected geomagnetic indices. The project may also include an interactive dashboard for visualising detected events and comparing the performance of different classification models.
This project investigates the use of deep-learning models such as LSTM, GRU or Transformer architectures to predict short-term variations in the Earth’s magnetic field. Their performance will be compared with conventional regression and statistical methods to determine whether deep learning offers meaningful improvement, particularly during periods of rapid geomagnetic change.
This project focuses on developing an automated platform to detect missing observations, duplicate timestamps, abnormal spikes, inconsistent sampling intervals and other data-quality issues in geomagnetic datasets. The system may also recommend or apply suitable data-recovery methods and generate a clear quality report for each monitoring station.
This project aims to predict rapid magnetic-field variations using external space-weather parameters such as solar-wind speed, IMF Bz, Kp, Dst and SYM-H. By incorporating both solar-wind and geomagnetic information, the study may provide better insight into the factors influencing ground-level magnetic disturbances.
This project explores how and why machine-learning models produce particular space-weather predictions. Explainable-AI techniques such as SHAP analysis, feature importance and residual analysis will be used to identify the variables that most strongly influence model outputs. The goal is to improve model transparency, scientific interpretation and confidence in data-driven space-weather applications.