Search NASA⌕ Search

SEARCH · Search NASA

Results for “hypothesis learning”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 37 records · Page 2

Using Big Data Analytics to Advance Precision Radiation Oncology

Big clinical data analytics as a primary component of precision medicine is discussed, identifying where these emerging tools fit in the spectrum of genomics and radiomics research. A learning health system (LHS) is conceptualized that uses clinically acquired data with machine learning to advance the initiatives of precision medicine. The LHS is comprehensive and can be used for clinical decision support, discovery, and hypothesis derivation. These developing uses can positively impact the ultimate management and therapeutic course for patients. The conceptual model for each use of clinical data, however, is different, and an overview of the implications is discussed. With advancements in technologies and culture to improve the efficiency, accuracy, and breadth of measurements of the patient condition, the concept of an LHS may be realized in precision radiation therapy.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

CDRL: Certification-Driven Reinforcement Learning for Neutrino Flavor Model Discovery

Many scientific discovery problems require searching combinatorial hypothesis spaces under complex domain constraints. Reinforcement learning (RL) offers a promising approach, but existing methods rely on scalar rewards that provide limited information about why candidate solutions fail, leading agents to repeatedly explore invalid regions. We introduce Certification-Driven Reinforcement Learning (CDRL), a framework that leverages structured feedback from symbolic reasoning tools. When a candidate violates domain constraints, these tools produce certificates identifying the actions responsible for failure. CDRL converts these certificates into reusable constraints that eliminate classes of invalid solutions and guide exploration toward valid regions. We evaluate CDRL on neutrino flavor model discovery in theoretical particle physics, where the hypothesis space exceeds $10^{26}$ possible models, and compare it with the state-of-the-art RL approach previously used for this task. Across three theory spaces, CDRL achieves up to 1.95$\times$ higher valid model rates and up to 6.33$\times$ higher neutrino model rates while evaluating up to 4$\times$ fewer candidates. We further extract 40 interpretable rules from search trajectories using a post-hoc decision-tree framework and show that reusing them as soft constraints yields gains of up to 2$\times$ in valid model rates and 3$\times$ in neutrino model discovery across all three theory spaces. These results suggest that CDRL uncovers reusable structure in combinatorial search spaces and provides a general framework for scientific model discovery.

Jha, Piyush [Georgia Tech., Atlanta; Georgia Tech]↗

Hypothesis testing via AI: Generating physically interpretable models of scientific data with machine learning (Full Technical Report)

Deep learning has demonstrated an exceptional ability to solve complex tasks (an engineering success); however, it has done so at the expense of the ability to generate new knowledge (a scientific failure). We propose an alternative framework—entitled Deep Symbolic Regression (DSR)—in which artificial neural networks (NNs) rapidly generate hypotheses about physical relationships among inputs. This framework bypasses the need to interpret an NN altogether, while still leveraging the representational power of deep learning. The resulting models are tractable mathematical expressions, which are inherently and readily human interpretable and can provide insights into underlying physical phenomena. Further, we fold this methodology into the scientific process by allowing the scientist to directly integrate a priori knowledge and beliefs to accelerate learning. We demonstrate this methodology on symbolic regression—the problem of rediscovering underlying expressions describing a dataset—and achieve state-of-the-art performance across a wide variety of symbolic regression problems. Further, we generalize our DSR framework to apply to the more general class of symbolic optimization problems, in which one seeks to optimize a sequence of symbols or “tokens” under a black-box reward function. Examples of other symbolic optimization problems include neural architecture search and computational antibody design. Our generalized tool, Deep Symbolic Optimization (DSO), has been demonstrated on the task of learning symbolic control policies for reinforcement learning environments, and has been adopted as an enabling capability for computational antibody design.

97 MATHEMATICS AND COMPUTING↗

Using Deep Machine Learning to Conduct Object-Based Identification and Motion Detection on Safeguards Video Surveillance

Video surveillance is one of the core monitoring technologies used by the International Atomic Energy Agency (IAEA) Department of Safeguards at safeguarded nuclear facilities worldwide. Current IAEA image-review software has functions for scene-change detection, black image detection and missing scene analysis, but their capabilities are not optimum. The current workflow for the detection of safeguards relevant events heavily depends on inspectors’ laborious visual examination of surveillance videos, which is a time-consuming process and prone to errors. To improve the accuracy of the process and reduce inspectors’ burden, the paper proposes using deep machine learning to detect objects of interest in video streams and to conduct object-based motion detection. The hypothesis of this work is that deep machine learning will reduce the burden on inspectors and reduce errors by automatically locating and identifying objects of interest in video streams. Objects of interest include casks and fuel assemblies that are typically monitored by inspectors. The algorithm being developed in this work is based on a computationally efficient deep machine learning algorithm – You Only Look Once (YOLO) – but is further devised to address specific challenges related to the operation of nuclear facilities. The developed model (which is called YOLO-NSG – YOLO for Nuclear Safeguards) is evaluated with data sets collected at the test facilities at Brookhaven National Laboratory (BNL) and Sandia National Laboratories (SNL). The initial focus of the research is for application at safeguarded nuclear reactors, such as pressurized heavy-water reactors, where video surveillance is broadly deployed, but can be extended to other use cases of nuclear safeguards. The detailed structure of YOLO-NSG model is introduced and the test results are reported in the paper.

36 MATERIALS SCIENCE↗

Toward a machine-guided approach to energetic material discovery

In this article, we trained a machine learning (ML) model to connect microstructural details of an energetic material formulation to its performance for the purpose of guiding the discovery of new explosive formulations. Our hypothesis was that the algorithm would robustly learn the training data and produce an accurate surrogate model. Specifically, the algorithm learned the relationship between details of the void size distribution (VSD), initiating shock pressure, and the energetic material performance. We used realistic constraints on the VSD and a range of cases were ingested by a physically informed reactive flow model working within a hydrodynamic solver running on high-performance computing resources. The ML algorithm produced a surrogate model that accurately predicted known test points around the parameter space. In addition to the utility of the model and the process used for its development, we noted interesting comparisons between what we, the authors—subject matter experts, would heuristically conclude from the training data and the surrogate model predictions. We detected nuanced details that were missed by the surrogate model; however, these details are not important to an energetic material formulator. We concluded that the algorithm did indeed robustly learn the training data and produce an accurate surrogate model. We further concluded that the surrogate model is a powerful tool to guide the formulator in the absence of subject matter experts and limited-access computing resources.

42 ENGINEERING↗

Off-the-shelf deep learning is not enough, and requires parsimony, Bayesianity, and causality

Abstract Deep neural networks (‘deep learning’) have emerged as a technology of choice to tackle problems in speech recognition, computer vision, finance, etc. However, adoption of deep learning in physical domains brings substantial challenges stemming from the correlative nature of deep learning methods compared to the causal, hypothesis driven nature of modern science. We argue that the broad adoption of Bayesian methods incorporating prior knowledge, development of solutions with incorporated physical constraints and parsimonious structural descriptors and generative models, and ultimately adoption of causal models, offers a path forward for fundamental and applied research.

97 MATHEMATICS AND COMPUTING↗

Anomaly Detection in Power System State Estimation: Review and New Directions

Foundational and state-of-the-art anomaly-detection methods through power system state estimation are reviewed. Traditional components for bad data detection, such as chi-square testing, residual-based methods, and hypothesis testing, are discussed to explain the motivations for recent anomaly-detection methods given the increasing complexity of power grids, energy management systems, and cyber-threats. In particular, state estimation anomaly detection based on data-driven quickest-change detection and artificial intelligence are discussed, and directions for research are suggested with particular emphasis on considerations of the future smart grid.

42 ENGINEERING↗

A Siamese CNN + KNN-Based Classification Framework for Non-intrusive Load Monitoring

Through the development of smart grids, programs such as demand side response, have been presented as auxiliary services to the real-time operation of distributed networks. In order to provide consumers information on their energy consumption, so that a modulation in consumption is possible, non-intrusive load monitoring has been introduced as an solution to this pattern recognition problem. Non-intrusive load monitoring enables the modeling of electrical loads connected to the low-voltage system, considering only a single measurement point. Presented state-of-the-art solutions though, consider availability of data as well as representation of all possible classes of the environment. This is of course a most conservative hypothesis, since in real-life applications availability of such data is much difficult, as well as the dynamic behavior of models is implicitly evolving in time. Here, a framework that uses neural Siamese networks with k-nearest neighbor clustering is presented toward non-intrusive load monitoring. Online learning feature is implemented, which relaxes the hypothesis of data requirements as well addresses the evolving nature of load profile. k-nearest clustering allows nonlinear characteristic space modelling. Test results using synthetics and real-life data show that the solution, besides obtaining a good generalizability in the classification, also obtained results with an accuracy of 95.77%.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Advanced Analytics Studies Applied to US Department of Veterans Affairs' Corporate Data Warehouse (Initial Draft)

In this report, we describe our experience of applying several advanced analytics algorithms to the US Department of Veterans Affairs’ (VA’s) Corporate Data Warehouse (CDW) electronic health records datasets during FY2017-18. While various algorithms were applied to the CDW data, the main goal of this effort was to provide useful insights into the operational aspects (as opposed to the purely clinical aspect) of this specific implementation. Since there are not many reports in the public literature on advanced analytics applied to this data, this report provides unique insight in this regard. We focused on two machine learning (ML) applications: (1) clinical pathway inference (patterns dominating clinical pathways applied to a male dataset of Stable Ischemic Heart Disease patients as use case) and (2) medical-concept representation learning’s capability to inform and refine cohort membership based on probable patient outcomes. During our study on clinical pathway inference, we used two main modeling techniques: (a) topic probabilistic modeling (latent Dirichlet allocation [LDA]) and (b) feature imbedding. We observed that the applicability of LDA to the pathway inference remains in question. LDA provides results that are intuitive and easy for humans to comprehend; however, when applied to the pathway inference, it also generates pathway patterns that consist of a few dominant pathway components. In addition, LDA provides no information regarding the relationships among components in the same pathway. The outputs of LDA are also found to categorize patients based on their trace data of clinical procedures. However, the results suggest that LDA is biased toward statistically dominant components. This makes it particularly hard to discriminate pathway subbranches. To address the issue, we designed a new pathway inference methodology that integrates temporal ordering of pathway components into LDA results, as existing feature embedding tools such as word2vec are not readily applicable to our task. In addition, we developed our own feature embedding tool customized for patient traces. We conclude this study by suggesting two approaches to use embedding representation of a component into LDA and briefly list our lessons learned and recommendations for future work. For our study on medical-concept representation learning ability to inform and refine cohort membership based on probable patient outcomes, during our empirical evaluation we compared methods from medical-concept learning to standard one of a kind (one-of-K) encoding to evaluate the change in effectiveness as done in Choi, Schuetz, Stewart, and Sun (2016). We performed two primary empirical evaluations. The first evaluation was on a curated collection of 60,000 patients with no more than 1 year of medical history included. The second was a collection of patients with no restriction to the amount of medical history included. The hypothesis across both evaluations is that representation learning is broadly useful independent of downstream processing models; thus, for each experiment, we trained three models—a logistic regression model, a two-layer neural network model, and a nearest-neighbor model—and then averaged the results. Each model was trained using fivefold cross validation. We provided evidence that medical-representation learning improves predictions of the primary diagnosis category of a short patient history through 12 experiments. In addition, we provided evidence that medical-representation learning fails to improve prediction of the primary diagnosis category of an arbitrarily long patient history through 12 experiments. We conclude this study with a brief list of our lessons learned and recommendations for future work. We conclude the report with a discussion of recent technology trends as they relate to our artificial intelligence and, more specifically, to our ML research and approaches as described previously.

97 MATHEMATICS AND COMPUTING↗

Model-independent measurement of the Higgs boson associated production with two jets and decaying to a pair of W bosons in proton-proton collisions at $\sqrt{s}=13$ TeV

A model-independent measurement of the differential production cross section of the Higgs boson decaying into a pair of W bosons, with a final state including two jets produced in association, is presented. In the analysis, events are selected in which the decay products of the two W bosons consist of an electron, a muon, and missing transverse momentum. The model independence of the measurement is maximized by employing a discriminating variable, developed through machine learning, that is agnostic to the signal hypothesis. The analysis is based on proton-proton collision data at $\sqrt{s}=13$ TeV collected with the CMS detector from 2016–2018, corresponding to an integrated luminosity of 138 fb −1 . The production cross section is measured as a function of the difference in azimuthal angle between the two jets. The differential cross section measurements are used to constrain Higgs boson couplings within the standard model effective field theory framework.

Hadron-Hadron Scattering↗

Using Auto-Segmentation to Reduce Contouring and Dose Inconsistency in Clinical Trials: The Simulated Impact on RTOG 0617

Contouring inconsistencies are known but understudied in clinical radiation therapy trials. We applied auto-contouring to the Radiation Therapy Oncology Group (RTOG) 0617 dose escalation trial data. We hypothesized that the trial heart doses were higher than reported due to inconsistent and insufficient heart segmentation. We tested our hypothesis by comparing doses between deep-learning (DL) segmented hearts and trial hearts.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

BETO 2021 Peer Review - Inverse Bioproduct Design Through Machine Learning and Molecular Simulation

This work aims to identify performance advantaged bioproducts (PABPs) through property prediction, which will guide experimental synthesis. The impact of this work will be faster market adoption of bioproducts with greater performance relative to incumbent products. We have identified >106 bioproduct candidates, but only some will have superior performance to create a market pull. High-throughput property prediction, enabled by machine learning, and elucidation of structure-function relationships, enabled by molecular simulation, provide a hypothesis driven approach for down selection of candidate biomolecules to pursue experimentally. To enable machine learning and molecular simulation for bioproduct discovery, automated structure generation and embedding must capture relevant features for prediction, databases must cover domains applicable to biobased products, and best practices for simulation of polymer systems must be developed. To address these challenges, we have established bioproduct relevant datasets, developed high-throughput polymer structure generation, and built end-to-end neural networks that have predicted 8 properties for >1.4 x 106 biopolymers. A molecular simulation pipeline for building, running, and analyzing polymers and polymer additives is being used to predict performance and develop design principles of biobased products. In collaboration with the PABP synthesis project, these computational tools are guiding synthesis and informing design of PABPs.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗

LENS: Learning Enabled Network Synthesis

RTRC and UMD have developed novel machine learning based methods under the ARPA-E DIFFERENTIATE program for rapid acceleration of hypothesis generation in complex architecture design spaces involving both discrete choices of component inclusion and interconnection and continuous parametric decisions. The project named Learning Enabled Network Synthesis (LENS) further demonstrated the developed methods on challenging electrical power converter design problems by identifying the most suitable circuit topologies and simultaneously selecting the most appropriate components to achieve optimized design of power converter with improved performances. We demonstrated that LENS could enable exploration of very large design space of circuit topologies and components by addressing the limitations of conventional design process in non-linear, high switching speed, multi-dimensional power converter design and optimization. The key innovation developed in LENS is the seamless integration of statistical learning and logical reasoning techniques and building on the individual strengths of these techniques for rapid hypothesis discovery. The main component of LENS comprises of: 1) Graph Reasoning Engine (GRE) to enforce composition rules that rapidly reject all discrete architectures that are composed incorrectly and generates an adaptive database of feasible designs which can be used by ML modules, 2) Graph Generative Learning module which is a deep neural network based generative model for graph architectures which can enable design space exploration beyond the dataset generated by the GRE, 3) Graph Reduced Order Model (ROM) for graph domains for accelerating computation of output metrics, and 4) Active learning and Rule Discovery module for sample efficient learning and extracting logical rules from the learned ML models which will be integrated in the GRE to enhance the filtering effectiveness. LENS approach can be applied to any design domains where designs can be represented as multi-attribute graphs. The LENS team integrated the various technical innovations listed above into an optimization pipeline and exercised the optimization pipeline on the converter design problem. The LENS project demonstrated that the developed AI/ML technologies can be used to generate novel converter circuits >45x faster than experts on chosen use-cases. This can enable faster design space exploration and identification of new designs which are not considered by experts due to the increasing design space complexity. This has significant potential impact on the public and energy needs of the country. It is currently estimated that 30% of all electrical powers generated passes through power converters. The future estimate is that 80% of all power generated would be passing through converters. LENS fills a critical gap in this space since by accelerating the design process the designers would be able to generate more efficient converters which can lead to significant energy savings for the country.

42 ENGINEERING↗

Explainable Artificial Intelligence Technology for Predictive Maintenance

The domestic nuclear power plant fleet has relied on labor-intensive and time-consuming preventive maintenance programs, thus driving up operation and maintenance costs to achieve high-capacity factors. Artificial intelligence and machine learning can help simplify complex problems, such as diagnosing equipment degradation, to enable more effective decision-making. Benefits will be felt not only within existing analog and digital instrumentation and control, but also work processes, the integration of people with technology, and most importantly, the business case. Together, these hold promise to make nuclear power more efficient and reduce costs associated with operation and maintenance. While the artificial intelligence and machine learning technologies hold significant promise in the nuclear industry, there are challenges or barriers to their adoption. This report outlines the those different machine learning adoption barriers (categorized as historical, technical, economic, regulatory, and user) that the industry must overcome to realize the full benefits of artificial intelligence and machine learning capabilities for long-term economic sustainability. This report also provides solutions for some of these barriers by focusing on improving the explainability of machine learning to encourage trust from the end-user. Trust and explainability are essential to machine learning adoption. This report focuses on research-developed solutions to some of these barriers while analyzing a non-safety-related system, namely the circulating water system. This system frequently experiences waterbox fouling which our models preemptively diagnoses then explains to the operator how those conclusions were reached. This report presents and discusses the inherent trade-off between machine learning performance (in terms of accuracy) and explainability, where highly accurate machine learning methods (such as deep-learning) are the least explainable, and the most explainable methods (such as decision trees) are the least accurate. In addition, explainability of artificial intelligence techniques in terms of transparency and post-hoc metrics are discussed. This report outlines the importance of data novelty and value of new information in evaluating both the explainability and trustworthiness. Novelty detection helps to establish consistency or inconsistency of the new data with respect to the training data. On the other hand, value of information could be a part of the user-centric visualization recommendation system that request additional information to be collected, thereby strengthening the machine learning outcomes. During this project, a copyrighted user-centric visualization that aligns with a human-in-the-loop approach was developed. The user-centric visualization presents different levels of information and can be tailored as per user credentials to gain user confidence. One of the salient features of the user-centric visualization is it presents machine learning methods with explainability metrics. A simplified version of the user-centric visualization was presented to 32 users with varying levels of machine learning expertise. Feedback was solicited to test the hypothesis that the app contained sufficient explainability and that the users would trust the algorithm. Overall, the app was positively received, and the hypothesis was supported. This report discusses the trust-but-verify framework – a potential approach to build user trust artificial intelligence. The framework discusses trust from the human level to artificial intelligence level. The fundamental premise of the trust but verify framework is derived from an observation of nuclear safety culture (i.e., nuclear power plant personnel do not rely on a singular source of data to make a decision). This also ties back to the user-centric visualization that presents different levels of information to achieve both explainability and trustworthiness of artificial intelligence. Even so, the adoption of artificial intelligence and machine learning in the nuclear industry faces additional barriers, namely regulatory and stakeholder readiness. To overcome these challenges, new solutions must gain regulatory approval and cater to stakeholder needs. The Nuclear Regulatory Committee has a 5-year strategic plan which prepares them for reviewing artificial intelligence technologies in licensee submissions. Early and frequent engagement with the regulator is encouraged. Additionally, artificial intelligence solutions should incorporate human-in-the-loop considerations and offer explainability. Stakeholders must prepare by hiring or training staff to adapt to advancing technology in everyday plant tasks.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

Automated Identification of Characteristic Droplet Size Distributions in Stratocumulus Clouds Utilizing a Data Clustering Algorithm

Abstract Droplet-level interactions in clouds are often parameterized by a modified gamma fitted to a “global” droplet size distribution. Do “local” droplet size distributions of relevance to microphysical processes look like these average distributions? This paper describes an algorithm to search and classify characteristic size distributions within a cloud. The approach combines hypothesis testing, specifically, the Kolmogorov–Smirnov (KS) test, and a widely used class of machine learning algorithms for identifying clusters of samples with similar properties: density-based spatial clustering of applications with noise (DBSCAN) is used as the specific example for illustration. The two-sample KS test does not presume any specific distribution, is parameter free, and avoids biases from binning. Importantly, the number of clusters is not an input parameter of the DBSCAN-type algorithms but is independently determined in an unsupervised fashion. As implemented, it works on an abstract space from the KS test results, and hence spatial correlation is not required for a cluster. The method is explored using data obtained from the Holographic Detector for Clouds (HOLODEC) deployed during the Aerosol and Cloud Experiments in the Eastern North Atlantic (ACE-ENA) field campaign. The algorithm identifies evidence of the existence of clusters of nearly identical local size distributions. It is found that cloud segments have as few as one and as many as seven characteristic size distributions. To validate the algorithm’s robustness, it is tested on a synthetic dataset and successfully identifies the predefined distributions at plausible noise levels. The algorithm is general and is expected to be useful in other applications, such as remote sensing of cloud and rain properties. Significance Statement A typical cloud can have billions of drops spread over tens or hundreds of kilometers in space. Keeping track of the sizes, positions, and interactions of all of these droplets is impractical, and, as such, information about the relative abundance of large and small drops is typically quantified with a “size distribution.” Droplets in a cloud interact locally, however, so this work is motivated by the question of whether the cloud droplet size distribution is different in different parts of a cloud. A new method, based on hypothesis testing and machine learning, determines how many different size distributions are contained in a given cloud. This is important because the size distribution describes processes such as cloud droplet growth and light transmission through clouds.

54 ENVIRONMENTAL SCIENCES↗

High-dimensional and permutation invariant anomaly detection

Methods for anomaly detection of new physics processes are often limited to low-dimensional spaces due to the difficulty of learning high-dimensional probability densities. Particularly at the constituent level, incorporating desirable properties such as permutation invariance and variable-length inputs becomes difficult within popular density estimation methods. In this work, we introduce a permutation-invariant density estimator for particle physics data based on diffusion models, specifically designed to handle variable-length inputs. We demonstrate the efficacy of our methodology by utilizing the learned density as a permutation-invariant anomaly detection score, effectively identifying jets with low likelihood under the background-only hypothesis. To validate our density estimation method, we investigate the ratio of learned densities and compare to those obtained by a supervised classification algorithm.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Toward a Diverse Next-Generation Energy Workforce: Teaching Artificial Photosynthesis and Electrochemistry in Elementary Schools through Active Learning

Artificial photosynthesis is a promising approach to generate commodity chemicals using abundant chemical feedstocks and renewable energy sources. Despite its importance, affordable and effective hands-on classroom activities that demonstrate artificial photosynthesis and teach key concepts, especially for primary school students, are lacking. Educating young students on this topic is a critical step in the development of the next-generation energy workforce, especially one that is diverse in race and gender. Here, we hypothesize that an effective approach to educate a broad range of young students on the topic of artificial photosynthesis is through the use of an active learning-based lesson plan that employs cheap and accessible materials. This hypothesis is confirmed by evaluating the understanding of fifth grade students with a survey before and after a lesson plan on artificial photosynthesis that uses active-learning techniques and uses safe and highly accessible materials (baking soda, tap water, plastic jars, Ni coil, alligator clips, and a solar cell) to perform solar-powered water splitting. The lesson plan and survey questions are designed to align with the educational outcomes for fifth grade classrooms in California and to address four general learning objectives: (1) Motivations of Artificial Photosynthesis, (2) Applications of Artificial Photosynthesis, (3) Inputs and Outputs of Artificial Photosynthesis, and (4) Engineering Design for Artificial Photosynthesis. The survey data demonstrate a statistically significant improvement in overall student understanding from the lesson plan. Importantly, the data show that the lesson plan presented here is effective at narrowing the performance gap between minority students and overly represented groups.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

An Agenda for Multimodal Foundation Models for Earth Observation

Archives of remote sensing (RS) data are increasing swiftly as new sensing modalities with enhanced spatiotemporal resolution become operational. While promising new breakthroughs, the sheer volume of RS archives stretches the limits of human analysts and existing AI tools, as most models are: i) limited to single data modalities; ii) task-specific; iii) heavily reliant on labeled data. The emerging Foundation Models (FMs) have the potential to address these limitations. Trained on vast unlabeled datasets through self-supervised learning, FMs enable generic feature extraction that facilitate specialization to a wide variety of downstream tasks. This paper describes a vision towards an FM for multimodal Earth Observation data (FM4EO), discussing key building blocks and open challenges. We put particular emphasis on multimodal reasoning, a topic underexplored in EO. Our ultimate goal is a practical path toward FM4EO with capacity to unlock breakthroughs in few-shot learning scenarios, multimodal geographic knowledge integration, synthesis, and hypothesis generation.

Ambrozio Dias, Philipe↗