A DECADE OF FIRSTS: FDL-X AND GOOGLE CLOUD CONTEMPLATE WHAT’S NEXT IN AI FOR SCIENCE.

Artificial IntelligenceEarthSpaceGoogle Cloud
Written By Ana Freitas

FDL-X and FDL Europe have quietly been using AI to tackle humanity’s most pressing problems on Earth and in space for nearly a decade in close partnership with Google Cloud.

AI has been dominating the headlines recently. We’ve seen some astonishing progress in generative AI. However, out of the limelight another revolution has been quietly gaining momentum - the application of AI to applied problems in science: climate, disaster response, energy efficiency, space situational awareness, health and deep space exploration. 

The application of AI in these areas has been so rapid and profound, it’s difficult to take it all in. Buried in science journals, data and code repos on GCP and jargon-filled technical memos is a treasure trove of technical and scientific progress that has blasted through pernicious problems and, in many cases, unlocked bold new capabilities for space agencies and scientific institutions. Perhaps due to the innate suspicion of the science community to hyperbole, the rapid and profound progress in the application of AI for science has been celebrated cautiously, buried in careful language of science, where sometimes implications require a second or third read before the “oh wow” really hits.   

“Perhaps due to the innate suspicion of the science community to hyperbole, the rapid and profound progress in the application of AI for science has been celebrated cautiously’”

This is of course open science’s inherent promise; that a tricky problem once solved and accepted by the community, catalyzes a wave of previously unconsidered application. This ability to combine and derive new opportunities, surfing a wave of rapid discovery is an exciting aspect of the AI for science revolution. Techniques such as physics-informed learning, causal discovery, continuous learning and foundation models are transforming how we think about science and discovery. New scientific insight can be extracted from the petabytes of data being generated. Data and algorithms are now key tools in the science toolbox. But the revolution is more than that; in the same way that container shipping changed global trade, the ability to package ‘insight’ within a data product and co-lab allows science to be glued together like Legos.

This concept of integrated AI, sometimes called a ‘Mixture of Experts’ is a key focus of FDL-X. We believe that complex, interconnected science problems like space weather, climate, health and space exploration lend themselves to orchestrated approaches working in concert - which is why a unified development and AI serving platform like GCP is key.

AI for science and the cloud are intrinsically linked. 

Cloud technology and AI are often bundled together on deep tech lists (you’ve seen them, often with IOT, blockchain, quantum, the metaverse et al) but the actual mechanisms of how AI and cloud tech work hand in glove to drive applied science aren’t often talked about. Put simply, AI research efforts like FDL-X and FDL Europe wouldn’t be possible without cloud infrastructure. The data handling at scale, availability of CPUs, GPUs and TPUs as needed, as well as numerous tools to support AI pipeline development and deployment reduce the barriers to teams doing great work. 

It’s perhaps worth taking a breath to reflect on how cloud technology has super-charged the research of organizations like FDL. 

  1. The Cloud enables science to be ‘living’ and seamlessly shared. 

  2. The Cloud enables enormous datasets to be ingested into AI pipelines from anywhere

  3. AI projects can spin-up as little or as much compute as they need on demand

  4. Interdisciplinary teams can collaborate effortlessly from anywhere from any laptop

  5. Multiple AI pipelines can be stitched together to make an orchestrated outcome  
    FDL.ai brings together some of the brightest interdisciplinary minds in space science, Earth science, health, AI and ML, academia, and the commercial sector to tackle here-and-now challenges for NASA in the United States and ESA in Europe. 

 Over eight weeks, researchers cracked the following questions: 

  1. Can we develop a live stream of the Sun and provide analysis-ready data of ultraviolet irradiance for the scientific community?

  2. Can we improve our ability to model thermospheric drag (which affects satellite orbits) by integrating ultraviolet irradiance data? 

  3. Can we provide reliable forecasts of the effect of geomagnetic storms on the Earth?  

  4. Can we use machine learning to build a near-real-time warning system for potentially dangerous solar weather events?

  5. Can we harness large language models (LLMs) to provide accurate and detailed information to natural disaster response and recovery teams? 

  6. Can we develop foundation models for analyzing synthetic aperture radar (SAR) Earth observation data that can generalize to a range of downstream tasks? 

FDL.AI is a model of how global teams of experts can come together to blast through difficult  problems, and Google Cloud enables distributed teams to work seamlessly together on enormous datasets; its significance to FDL.ai’s work cannot be overstated. Google Cloud’s robust infrastructure, versatile services (including TensorFlow and Cloud-AI pre-trained ML models), and extensive ecosystem provide a powerful foundation for AI development and deployment. Google Cloud’s storage capabilities and scalable computing resources allow teams to handle large datasets, and BigQuery allows for efficient data analysis. Moreover, Google Cloud’s emphasis on security and compliance ensures that AI projects meet rigorous industry standards.

 

FDL-X 2023 Challenges

Featured
FDL-X 2023: Improving Thermospheric Drag Measurements
FDL-X 2023: Improving Thermospheric Drag Measurements

This project built upon the EUV virtual instrument for observing in this wavelength from FDL.ai 2018; the 4pi view of the Sun from FDL.ai 2022; and the image-to-image translation approach and auto-calibration techniques of FDL.ai 2019. The goal was to ensure safer space traffic management, reduce the errors in thermospheric density modeling, and assess the impact of EUV irradiance in thermospheric density modeling.

The team was able to produce SDOML embeddings to encode the sun; use SDOML embeddings to replace solar proxies; assess the impact of SDOML embeddings as inputs for thermospheric density models; and look at differentiable physics-based models and losses for thermospheric density modeling.

FDL-X 2023: Curating Analysis Ready Data (ARD) for Solar Spectral Irradiance
FDL-X 2023: Curating Analysis Ready Data (ARD) for Solar Spectral Irradiance

This project built upon the Solar Dynamic Observator’s Machine Learning Dataset (SDOML) and the work of FDL.ai 2022 in terms of a 4pi view of the Sun across four EUV channels in order to provide 4pi datasets for EUV solar spectral irradiance. The ARD team was able to build an SDO computational platform; a cloud-based large-scale data ingestion, processing, and ML platform. They launched SDO ML V2 which includes analysis ready data for SDO AIA, HMI, and EVE from 2010 until 2023. They also created Virtual EVE, a deep learning architecture for offline and online solar irradiance forecasting based on AIA and HMI input data.

FDL-X 2023: Creating a Sun-To-Mud GIC Forecasting Scheme
FDL-X 2023: Creating a Sun-To-Mud GIC Forecasting Scheme

To mitigate the risk associated with geomagnetically induced currents (GICs), there is a ned to provide end-users with forecasts that are fast, localized, and actionable. This project built upon FDL.ai’s DAGGER pipeline and pursued a hybrid approach of combining ML and AI with human-in-the-loop methodologies to track the geoffective potential of solar transients.

DAGGER is now the first and only ML forecasting model able to predict localized geomagnetic fields globally; it learns the physics in multiple regimes and provides actionable forecasts through multiple lead teams and with associated uncertainty. Integration provides a basis for extended lead times, real-time data ingestion, coupling to downstream models, and operational deployment into the future.

FDL Europe 2023 Challenges

Featured
FDL Europe 2023: Developing a Foundation Model for Analyzing SAR EO Data
FDL Europe 2023: Developing a Foundation Model for Analyzing SAR EO Data

SAR is independent of weather and day or nighttime conditions; its signal can penetrate solid objects and measure millimeter-scale land deformations. But unlike optical data, SAR is difficult to use. It needs a lot of preprocessing under various different parameters, and deep learning SAR models don’t work reliably in novel regions.

With this challenge, the SAR team developed a flexible machine learning pipeline to leverage the wealth of information contained in different SAR input imagery. They also fine tuned these pre-trained models on a variety of labeled downstream tasks. This included dealing with complex and missing data, generalizing across time and space domains, and developing tooling to facilitate easy analysis and compelling visualizations.

You can read more about our FDL-X 2023 teams and research challenges here and about our FDL Europe 2023 teams and research challenges here.

FDL Europe 2023: Creating An SSA Live Twin for Space Weather
FDL Europe 2023: Creating An SSA Live Twin for Space Weather

Extracting the direction and density of coronal mass ejections (CMEs) is crucial for predicting their impact. The upcoming ESA VIGIL mission will provide a much needed additional viewpoint to allow triangulation and 3D construction of CMEs — and the ability to do so onboard can save valuable time. The SSA Live Twin for Space Weather team was up for the challenge and developed a machine learning pipeline that could analyze data both onboard the spacecraft and from Earth observatories, significantly reducing downlink bandwidth requirements and alert latency.

FDL Europe 2023: Building a Foundation Model Adaptor for Disasters
FDL Europe 2023: Building a Foundation Model Adaptor for Disasters

Climate change is already causing more frequent and extreme weather events; floods are some of the most prevalent and dangerous. Disaster reports are needed to make better flood relief decisions and resilience plans, but disaster reporting is not an easy task.

The FM Adaptors for Disasters team combined Earth Observation (EO), flood prediction models, flood-mapping tools, human-in-the-loop, and weather monitoring systems with LLMs to create an NLP agent that can create accurate and tailored flood reports that can be used to assess disaster impact across the globe. Check out their work at floodbrain.com (password: floodbrain).

Looking Forward

Thank you to our partners at Google Cloud for supporting the next frontier in research excellence. Read more about our 2023 research challenges and get involved at fdleurope.org (FDL Europe) and fdl.ai (FDL-X). 

“Google Cloud has enabled us to keep up with the big AI labs, and they’ve been steadfast in their support and ambition for a better future, for all humankind. Cloud-based AI technology allows us to think big, operate globally, and deliver reliably. Without Google Cloud, we wouldn’t have the continued ability to over-deliver, which we always aspire to,” said James Parr, Founder of FDL.ai and CEO of Trillium Technologies, which operates FDL.ai in partnership with NASA and ESA. 

FDL.AI’s mission is grounded in open science in order to enable scientists, scholars, and science enthusiasts alike by providing access to machine learning projects, code, MLOps, server-side compute, analysis ready data, and maintained benchmark datasets. To this end, all of FDL.ai’s research will be available at SpaceML.org in the coming months.

Learn more about past and present FDL.AI programs at fdleurope.org and fdl.ai.

Ana Freitas
Previous

Space Science & Astrobiology Division Seminar Series
Next

Living with our Star, Space Weather: GNSS