RESEARCH

FDL Heliolab 2025

Active Region Characterization and Analysis of Dynamics and Evolution

Can we capture and codify the dynamic state of active regions on the Sun to inform downstream chained models? 

The changing features of active regions on the Sun potentially offer clues to inform models of internal drivers of solar activity - some of which may have negative downstream consequences for technology and humans on Earth, Moon or Mars. This challenge will leverage data from solar instruments such as SDO/HMI and 4K imagery of the solar disk to examine active regions in detail.  Employing AI-driven feature extraction, physics-informed helioseismology, and continual learning to derive key parameters, like magnetic field strength and configuration, which will serve as essential boundary conditions for subsequent modeling and can be seen as the “first step in the chain” of understanding.

  • Active Region Characterization and Analysis of Dynamics and Evolution Live Showcase

Results

Space weather events, such as powerful solar flares and coronal mass ejections (CMEs), originate in the Sun's complex magnetic field and pose a significant threat to modern infrastructure on Earth. They can disrupt electrical grids, damage satellites, and expose astronauts to dangerous radiation. The earliest warning signs of these events are found in the Sun's active regions areas where the magnetic field becomes highly concentrated and unstable. Existing physics-based models, such as Surface Flux Transport (SFT) models, predict the evolution of the solar magnetic field but struggle to handle the vast, multi-modal datasets now available from observatories like the Solar Dynamics Observatory (SDO). The scientific community has a wealth of data, but a historical gap exists between traditional physics-based models and the advanced AI techniques that could extract the full predictive power of this information.

Team ARCADE has developed a novel hybrid system that marries traditional heliophysics principles with modern AI methodologies to forecast the emergence and evolution of active regions for the first time. The system, known as ARCADE, addresses the long-standing “physics and AI divide” by incorporating a wide range of solar data modalities and using a neural network (specifically, a ResNet architecture) to solve the governing physics equations for SFT. This approach allows the model to learn the underlying parameters that describe the physical processes, creating a more interpretable and reliable forecast than a pure black-box AI model. A key innovation of ARCADE is its integration of Uncertainty Quantification (UQ) techniques, which provide a probabilistic perspective on the forecasts, making them more trustworthy and useful for scientists and stakeholders.

ARCADE's approach represents a significant step forward in space weather modeling. By blending physical realism with the adaptability and efficiency of AI, the system provides accurate short-term forecasts of active region emergence up to six hours in advance. This capability is crucial for creating an early warning system that provides Earth with more time to prepare for potential space weather hazards. ARCADE is the first model of its kind to successfully integrate physics-informed machine learning with uncertainty quantification to create an interpretable, cloud-native, and scalable framework. This will not only advance scientific research by allowing for rapid hypothesis testing but also provide critical tools for practical forecasting, helping to safeguard our technology and society from the impacts of solar activity.

See this work as a poster presentation at the NeurIPS 2025 AI4Science Workshop: Data-Driven Solar Surface Flux Transport Modeling with Uncertainty Quantification.

Decoding Solar Wind Structure

Can we identify and understand the precursors and drivers of solar wind structures, with an aim to forecast these in advance? 

This challenge seeks to identify and understand the solar precursors and drivers of different solar wind structures. The challenge itself will consist of two stages. In the first stage, work from FDL 2021 on clustering and classifying solar wind structures such as switchbacks and high speed streams emanating from coronal holes will be extended and include the latest in-situ Parker Solar Probe and Solar Orbiter data. Both unsupervised and interpretable machine learning approaches will be used to explore the complex nature of solar wind structures. The second stage takes this a step further, using the distinct solar wind structures identified, solar disk imagery, and foundation model embeddings from SDO data to link precursors and drivers on the solar disc to target structures. An aim here is to drive physics informed modelling, including the physics-based HUXt model and magnetic field data. A direct application of being able to link dynamics on the Sun with solar wind structure is being able to inform solar wind prediction and answer key scientific questions, such as the extent of magnetic reconnection in the corona driving observed switchbacks. From a solar wind prediction aspect the modelling of coronal holes directly informs the preconditioning of the solar wind and downstream impacts to the near-Earth geospace environment.

Results

The solar wind, a continuous flow of particles from the Sun, is a dynamic and complex system that directly influences space weather. While it is the backdrop for more extreme events like CMEs, its own structures, such as fast and slow wind streams, stream interaction regions, and switchbacks, can cause geomagnetic storms that impact satellite orbits, communications, and navigation systems. Understanding and predicting these structures is crucial for mitigating their effects. The recent Parker Solar Probe (PSP) mission provides unprecedented data from closer to the Sun than ever before, but the sheer volume and complexity of this multi-instrument data, which changes on a second-by-second basis, pose a significant challenge for traditional analysis methods. A systematic way to classify and understand these solar wind structures is needed to push the boundaries of heliophysics research.

To address this, the Decoding Solar Wind team developed a novel framework called “CIPHER” which combines a human-in-the-loop, semi-automated approach to systematically classify solar wind structures from PSP time series data. CIPHER operates in four stages: it compresses the time series data, clusters segments based on their properties, allows a human expert to review and label the results, and then classifies the entire dataset. This process allowed us to create a cleaned and labeled PSP dataset, including new catalogs of stream interaction regions and switchbacks. We then used this labeled data to train a sophisticated machine learning model for prediction. Our model is built upon the SDOFM foundation model, which is pre-trained on Solar Dynamics Observatory (SDO) images of the Sun. By combining PSP's in-situ data with SDO's images and the labels generated by CIPHER, our model learns to predict solar wind classes at PSP's location based solely on images of the solar disk.

The model is not limited to predicting conditions at PSP's specific location. By leveraging neural fields, it can predict solar wind types anywhere in the inner heliosphere, providing a comprehensive, global view of solar wind generation and propagation. This capability allows prediction of the solar wind impacting Earth or any other spacecraft, offering a powerful new tool for space weather forecasting. The project has produced several key outcomes for the heliophysics community: a cleaned and labeled ML-ready dataset from the PSP mission, the flexible and extensible CIPHER pipeline for analyzing any in-situ solar wind data, and a new catalog of classified solar wind events. Most importantly, the team has created a robust solar wind prediction model based on the SDOFM foundation model, ready to be used for both scientific research and operational prediction tasks, and providing new insights into the massive volume of PSP data.

See this work as 3 poster presentations at the NeurIPS 2025 Machine Learning and the Physical Sciences Workshop: (1) CIPHER: Scalable Time Series Analysis for Physical Sciences with Application to Solar Wind Phenomena, (2) Uncovering Solar Wind Phenomena with iSAX, HDBScan, Human-in-the-loop and PSP Observations, and (3) Scalable Machine Learning Analysis of Parker Solar Probe Solar Wind Data

Ionosphere-Thermosphere Digital Twin

Can we resolve local ionospheric disturbances at a global scale?

This challenge aims to develop a global model of Total Electron Content (TEC) by leveraging data from dual-band GNSS receivers and expanding to large, low-level mobile GNSS datasets. The model will support ionospheric scintillation studies and integrate with the FDL-developed Karman-CL pipeline, enabling advanced system integration for forecasting.

The primary goal is to establish a framework that unifies individual models to study magnetospheric-ionospheric coupling. Building on prior success in forecasting neutral density, this project extends its application to model ionospheric scintillation—critical for understanding space weather impacts on communication and navigation systems.

The scientific value lies in developing an AI-driven model connecting processes from the Sun to the solar wind, magnetosphere, upper atmosphere, and ground. Key innovations include enabling inter-model communication, where outputs from one AI model inform another, and the sheer scale of the challenge in getting mobile GNSS data integrated. The main risks involve ensuring effective integration and managing the complexity of these interconnected systems.

Results

The ionosphere critically influences Global Navigation Satellite Systems (GNSS), satellite communications, and Low Earth Orbit (LEO) operations, yet accurate prediction of its variability remains challenging due to nonlinear couplings between solar, geomagnetic, and thermospheric drivers. Total Electron Content (TEC), a key ionospheric parameter, is derived from GNSS observations, but its reliable forecasting is limited by the sparse nature of global measurements and the limited accuracy of empirical models, especially during strong space weather conditions. In this work, the Ionosphere-Thermosphere Twin team developed a machine learning framework for ionospheric TEC forecasting that leverages Temporal Fusion Transformers (TFT) to predict sparse ionosphere data. The approach accommodates heterogeneous input sources, including solar irradiance, geomagnetic indices, and GNSS-derived vertical TEC, and applies preprocessing and temporal alignment strategies. Experiments spanning 2010–2025 demonstrate that the model achieves robust predictions up to 24 hours ahead, with root mean square errors as low as 3.33 TEC units. Results highlight that solar EUV irradiance provides the strongest predictive signals. Beyond forecasting accuracy, the framework offers interpretability through attention-based analysis, supporting both operational applications and scientific discovery. To encourage reproducibility and community-driven development, we release the full implementation as the open-source toolkit ionopy.

See this work as a poster presentation at the NeurIPS 2025 Machine Learning and the Physical Sciences Workshop: Forecasting the Ionosphere from Sparse GNSS Data with Temporal-Fusion Transformers

Multimodal Flare Prediction

How can we synthesize data from multiple NASA missions to develop a 4π virtual flare monitor?

This challenge aims to develop a virtual 4π GOES soft X-ray (SXR) flare monitor using front- and far-side EUV images to enable continuous solar flare prediction across the entire solar disc, enhancing space weather forecasting. By merging this virtual instrument output with ensemble forecasting and integrating observational data, the system will continuously update the probability distribution of imminent solar eruptions across the Sun.

The scientific value lies in overcoming current limitations in flare monitoring. For over 30 years, GOES SXR data has been essential for classifying flare strength and informing forecasts. However, existing methods focus only on front-side observations, leaving significant limb and far-side data underutilized.

Key innovations include using EUV data to assess whether an event is eruptive, improving forecasting capabilities. A potential risk is inadequate uncertainty quantification of the EUV flare index for reliable research and operational use but overcoming this, the result would be transformative for human space exploration.  Once developed, this virtual instrument will be able to integrate and be driven by foundation models, such as HelioFM.

Results

Solar flares, sudden bursts of energy from the Sun, pose a significant threat to our technological infrastructure, including satellites, power grids, and GPS. Real-time monitoring of these events is crucial for mitigating their impact. The GOES satellite series provides essential data on a flare's soft X-ray flux, a key indicator of its strength. However, GOES data is an integrated measurement, meaning it tells us how strong a flare is but not its specific location on the Sun's surface. To pinpoint a flare's origin, we need complementary data, such as the extreme ultraviolet (EUV) images from NASA's Solar Dynamics Observatory (SDO). The current reliance on two separate systems (GOES for strength and SDO for location) is a major limitation, preventing a unified approach to flare forecasting and real-time response.

To solve this problem, the Multimodal Flare Prediction team developed “FOXES”, a novel machine learning model that can extract both the strength and location of a solar flare from SDO EUV images alone. The solution is based on a Vision Transformer (ViT) architecture, a class of models that treats images similarly to how large language models process sentences. By segmenting solar images into smaller “patches”, like words in a sentence, the model learns complex patterns between these patches. This allows it to identify which regions of the sun are most active and accurately translate the visual information from SDO images into a predicted soft X-ray flux value, essentially creating a “virtual GOES” instrument. Training the model on a combination of SDO and GOES data, the model consistently shows high accuracy, particularly for the most powerful and dangerous flares.

FOXES represents a significant step forward in space weather forecasting. By consolidating two separate data streams into a single, unified system, it provides a powerful tool for real-time flare detection. The model can not only predict flare strength with high fidelity, even when traditional GOES data is missing, but also provide interpretable heatmaps that pinpoint a flare's precise location. This capability is a game-changer for mission planners and space weather teams. Furthermore, because FOXES operates on image data, it can be applied to solar images from viewpoints beyond Earth's line of sight, such as those from the STEREO mission. This will enable us to create the first-ever comprehensive multimodal flare catalog and estimate the strength of events on the far side of the Sun. This revolutionary leap forward will be critical for protecting astronauts on future missions to deep space, such as to Mars, and will enhance our ability to prepare for solar events before they impact Earth.
See this work as a poster presentation at the NeurIPS 2025 Machine Learning and the Physical Sciences Workshop: FOXES: A Framework For Operational X-ray Emission Synthesis

Orchestrator: Ionosphere-Thermosphere

Can we enable confidence in model predictions by including context from multiple models?

Enabling a decision intelligence system: this challenge aims to establish a framework that enables seamless integration, interaction, and fine-tuning of models developed across multiple Heliolab research challenges. This framework will allow AI agents to interact with diverse models—adapting inputs, running complex workflows, and sourcing outputs—thereby enhancing flexibility, interoperability, and scientific discovery. A key innovation is enabling the use of embeddings from foundation models, such as HelioFM, to enrich model interactions and support advanced reasoning capabilities.

This challenge is critical for realizing the full potential of RADIANT, serving as a departure point for future model integration across the community. Providing a common platform and specified framework will enable RADIANT to link models. The goal is to create an adaptable environment where models can be fine-tuned for specific objectives, and AI agents can autonomously manage complex workflows as a stepping stone to a full sensing-reasoning system for the solar-terrestrial environment.

  • Orchestrator Live Showcase

Results

As part of Heliolab 2025, the Orchestrator team created an agentic orchestration system for heliophysics tasks. Heliophysics research faces significant challenges in synthesizing vast, heterogeneous datasets from multiple ground-based observatories and space missions, with traditional methodologies remaining largely manual and siloed. The agentic orchestration system developed addresses these limitations by enabling integration and interaction between computational models across heliophysics research. The system employs Large Language Model-based (LLM-based) agents structured according to established design patterns. Our implementation leverages state-of-the-art orchestration primitives, specifically Anthropic's Model Context Protocol (MCP) for tool description and Google's Agent Development Kit (ADK) for agent-to-agent communication. The system incorporates domain-specific tools ranging from ionospheric models to solar surface simulations, augmented by Retrieval Augmented Generation (RAG) containing heliophysics literature and worked examples. Valuation was conducted through demonstrated capabilities in ionospheric modeling, solar surface analysis, automated pipeline generation, and tool discovery. The orchestrator system autonomously generates data pipelines, creating and managing computational infrastructure, all whilst requiring human oversight for critical decisions. The system reduces prototyping time from months to minutes, providing natural language access to sophisticated heliophysics simulations and machine learning models. This work establishes a first-attempt for accelerated scientific discovery in heliophysics by improving access to computational tools and enabling rapid hypothesis testing through automated workflow orchestration.

Additionally, the team created Reasoning With a Star (RWS), a new contribution for a heliophysics dataset applicable to LLM-reasoning, and provide an initial benchmarking approach. These data are constructed from NASA/UCAR Living With a Star (LWS) problem sets and manually compiled into a readily consumable question-and-answer structure with context, unit-aware numeric ground truth, metadata (expected type, required units, format hints), and acceptance checks. A programmatic grader grades answers via symbolic equivalence, unit-converted numeric tolerance, and schema rules. This work provides a single-shot baseline and initial agent-based benchmarks, and an improvement was found when tasks are broken down via patterns of agents.

See this work as a poster presentation at the NeurIPS 2025 AI4Science Workshop: An Agentic Orchestration System for Heliophysics Tasks and Machine Learning and the Physical Sciences Workshop: Reasoning With a Star: A Heliophysics Dataset and Benchmark for Agentic Scientific Reasoning.