Earth Systems Lab 2024 Case Study

Project Background

Coronal Mass Ejections (CMEs) are large-scale solar events that release vast quantities of charged particles into space. When these particles reach Earth, they can cause significant damage to critical infrastructure, including satellite systems, power grids, and communication networks. A single, major CME event could lead to widespread disruptions, with economic damages estimated at up to $2.6tn, and recovery efforts extending over two years. As global dependence on technology grows, particularly in electronics and satellite-based communications, the need for an advanced early-warning system is increasingly urgent.

Solar flares, often precursors to CMEs, are sudden bursts of energy resulting from magnetic activity on the Sun’s surface. Accurately forecasting solar flares is a first key step in predicting potential CMEs, providing the advanced lead time for space agencies, governments, and infrastructure operators to take protective measures. However, forecasting solar flares and CMEs presents significant scientific and technical challenges, as the Sun’s magnetic field is highly dynamic and complex as it goes through the solar cycle, and our understanding of solar activity remains incomplete as the amount of activity varies in different cycles.

A further major limitation is the current observational infrastructure. Most real-time solar observations come from spacecraft positioned at the first Lagrange point (L1) - Lagrange points are positions in space where the gravitational forces of a two-body system like the Sun and Earth produce enhanced regions of attraction and repulsion - L1 lies between the Earth and the Sun. Observations from L1 are useful for solar flare predictions because this position provides a direct line of sight to the Sun, observing solar regions as they face Earth. However, large solar structures such as CMEs extend above or around active regions, so detection of these structures becomes a challenge as in many cases, they are difficult to observe from the front-facing view provided by L1.

The above diagram shows how observations from an angle, such as from L5, would provide a better perspective on whether a structure is forming above an active region, and to comprehensively understand the structure. The L5 position can provide a more complete view on a certain region, compared to one from the L1 position, as illustrated below where the view of the Sun from L5 (left) and Sun seen from L1, geometrically projected onto the view from L5 (right). It demonstrates that the regions at the border of the Sun in L1 are smeared due to the geometrical projection. This implies a loss of spatial resolution due to the curvature of the Sun at its borders, so being able to observe the Sun from different angles would mitigate this problem.

Furthermore, large solar flares and CMEs are relatively rare events, making it difficult to collect enough observational data to train and validate predictive models. Therefore, extending the current state-of-the-art forecast lead times, and improving accuracy will require innovative approaches, including leveraging new observational technologies and advanced computational methods.

Project Results

The team’s primary objective was to demonstrate that the HelioWatch framework can achieve state-of-the-art forecasting performance. The first stage involved looking at how labelling of datasets could be improved as this is essential aspect of building effective machine learning models for flare forecasting. In the heliophysics community, the standard practice is to assign binary labels ("Flare" or "No Flare") to images taken within a specified lead time - typically 24 hours - before a flare. However, the team identified limitations in this approach, so they developed a more dynamic option using a regressive labelling scheme, predicting the number of hours remaining until the next flare. This provides more granular information and could improve the models’ ability to forecast not just if, but when a flare is likely to occur. A third approach they explored combines the previous two and involved predicting the occurrence of a flare across multiple future time steps simultaneously, rather than focusing on a single lead time or a regression to the next event. In this method, the model forecasted a vector where each element corresponds to a specific time step in the future e.g., one hour, three hours, six hours, etc. For each of these time steps, the model outputs a binary label ("Flare" or "No Flare") depending on whether a flare is predicted to occur at that particular point in time. This approach enabled the model to not only predict the occurrence of a flare within a given time but also provided a temporal distribution of flare activity over a defined future period. The three methods are illustrated in the diagrams below.

Standard labels

Regressive Labels

Multi-timestep Labels

Each framework labelling option was then incorporated with models such as ResNet50, ResNet18, and cutting-edge, pre-trained transformer-based architectures, such as Meta’s DINOv2 and Google’s Vision Transformer (ViT). The team also tested two different lead times for forecasting, specifically 12 and 24-hour windows, and sampled data at two different cadences: 12 minutes and 1 hour, allowing the team to evaluate the effectiveness of different temporal resolutions for flare prediction.

The highest score was achieved by ResNet50, trained on hourly samples of Helioseismic and Magnetic Imager (HMI) data from the SDO satellite. This predicted whether a flare would occur within the next 24 hours. An example is shown below, where the red stars on the magnetogram images show where a flare occurred, and the vertical dashed line indicates the time at which the images were taken.

The model’s prediction probability (in blue) rose as the time of the flare approached October 20th. The model then accurately predicted subsequent flares between October 21st-22nd and October 22nd-23rd, with high probabilities.

The Scan Partnership

Scan is a major supporter of ESL 2024 and FDL Europe, building on its participation in the previous four years events. As an NVIDIA Elite Solution Provider Scan contributes multiple DGX supercomputers in order to facilitate much of the machine learning and deep learning development and training required during the research sprint period.