CLiMB: A domain-informed novelty detection clustering framework for galactic archaeology and scientific discovery

Large language models (LLMs) and agent systems are increasingly transforming scientific discovery, driving progress across chemistry, biology, materials science, and physics. Yet most existing work and surveys remain fragmented, focusing on isolated tasks such as idea generation or experiment design without addressing how these components fit within the broader discovery process. To bridge this gap, we introduce the EXHYTE cycle, an iterative framework that formalizes scientific discovery as a sequence of Exploration, Hypothesis generation, and Testing. We assembled a corpus of recent studies, distilled recurring strategies that characterize how AI methods contribute to each EXHYTE substage, and organized the literature accordingly to representative strategies and domain-specific advances. This process-centric perspective unifies diverse methodologies under a single structured workflow, identifies substages that are mature versus underexplored, and reveals complementarities that enable closed-loop discovery systems. It also clarifies the evolving division of labor between human researchers and AI systems, offering a roadmap for developing adaptive, autonomous frameworks for AI-driven scientific discovery. An accompanying website with paper summaries and an LLM-powered interactive survey based on EXHYTE is available at https://webapps.crc.pitt.edu/exhyte/

Large language model agents hold immense promise for automating scientific discovery, yet their real-world application is hindered by an inability to reliably orchestrate tools and execute complex, multi-step plans without encountering hallucinations or logical inconsistencies. Here, we present TeLLAgent, a novel supervisor-executor dual-agent framework that explicitly separates strategic reasoning from precise tool operation to overcome these limitations. The global planning agent, powered by DeepSeek-R1, performs iterative chain-of-thought reasoning to decompose problems and formulate dynamic plans. The local execution agent, leveraging DeepSeek-V3.1, then accurately invokes a curated suite of 30 specialized tools. A critical self-correction loop, mediated by the Model Context Protocol, allows the system to “rethink” and “recover” from failures, significantly enhancing robustness. When rigorously benchmarked on a suite of complex tool-calling tasks, TeLLAgent significantly outperformed GPT-5 and existing agent frameworks, achieving higher success rates in multi-step planning and demonstrating superior scaling with task complexity. Furthermore, TeLLAgent drastically reduced factual hallucinations in knowledge retrieval, as validated by both human experts and LLM judges, underscoring its enhanced reliability. We ultimately demonstrate the power of this approach by deploying TeLLAgent for autonomous discovery in the demanding domain of organic solar cell materials. From a single natural language query, it executed an end-to-end workflow, from molecular design and property prediction to the identification of a high-performance quasi-macromolecular acceptor. This AI-designed molecule was subsequently synthesized and validated, achieving a power conversion efficiency of 16.44%. TeLLAgent establishes a new paradigm for building reliable, autonomous AI systems, proving its potential to accelerate scientific discovery in materials science, drug discovery, and beyond.

We describe the development of a large-scale high-throughput application for discovery in materials science. Our point of departure is a computational framework for distributed multi-scale computation. We augment the original framework with a specialized module whose role is to route evaluation requests needed by the high-throughput application to a collection of available computational resources. We evaluate the feasibility and performance of the resulting high-throughput computational framework by carrying out a high-throughput study of battery solvents. Our results indicate that distributed multi-scale computing, by virtue of its adaptive nature, is particularly well-suited for building high-throughput applications.

본 논문은 그리드 컴퓨팅을 이용한 고성능 과학기술지식처리 프레임워크인 SINDI-Grid의 개발에 관련된 연구이다. SINDI-Grid 프레임워크는 대용량의 데이터 저장소 및 고속의 컴퓨팅 파워를 제공하는 그리드 컴퓨팅의 장점을 이용하여 분산 데이터 분석과 과학기술지식처리를 위한 다양한 그리드 서비스들을 제공한다. 그리고 SINDI-Workflow 도구는 이러한 서비스들을 이용하여 다양한 지식처리 알고리즘을 통합하는 복잡한 과학기술지식처리 애플리케이션을 설계하고 실행하는 역할을 수행한다. In this paper, we propose the SINDI-Grid which is a high-performance framework for scientific and technological knowledge discovery using the grid computing. By using the advantages of the grid computing providing data repository of large-volume and high-speed computing power, the SINDI-Grid framework provides a variety of grid services for distributed data analysis and scientific knowledge processing. And the SINDI-Workflow tool exploits these services so that performs the design and execution for scientific and technological knowledge discovery applications which integrate various information processing algorithms.

In the early 1910s, Italy was world leader in cinema with its spectacular films of Roman and Carthaginian antiquity. Despite their innovations in storytelling and mise en scène, filmmakers like Guazzoni and Pastrone also looked backward to the 19th century, by appropriating not only literature and theatre, but also painting, which has been hitherto little researched. ‘Archaeologist’ painters like Gérôme, Alma-Tadema and Rochegrosse, who combined painstaking historical research with their own imagination of antiquity, thus experienced a second life in the 20th century medium of film. Thanks to the use of mechanical reproduction, their works became part of public memory and were reused by filmmakers, most evident so in two key films of the early years of Italian cinema: Quo vadis? (1913) and Cabiria (1914). Yet, particularly for Cabiria, this book also creates a new archaeological framework from which to approach early Italian epics.

Convective storms vertically transport water vapor and condensate from Earth’s surface to the upper troposphere. Life on Earth is fundamentally linked to this transport which determines the hydrological cycle, and the intensity of severe weather responsible for the destruction of life and property. Despite advances in high-resolution modeling and better observational capabilities, the scientific community continues to be confronted with knowledge gaps about convective storms that limit our predictive capabilities. The ongoing developments in the high-resolution Energy Exascale Earth System Model (E3SM), large eddy simulations, and AI-based analytics to evaluate uncertainties are expected to provide a comprehensive framework for new scientific discovery. The model-experiment (MODEX) approach suggests that the aforementioned advancements in model development and AI-based inference techniques should be complemented by similar advancements in the experimental (observational) side so that the former does not outstrip the ability of the latter to provide meaningful constraints. What are the recent advancements in observations that will provide the necessary leap forward in improving our predictive capabilities? To address this question, we propose a new experimental paradigm called Multisensor Agile Adaptive Sampling (MAAS) that capitalizes on advancements in communications (5G), computational resources (edge/fog computing), sensor capabilities, and machine learning (ML) and AI techniques (Kollias et al., 2020). The MAAS framework allows for the collection of higher spatiotemporal resolution and quality observations of convective storms than is traditionally possible. The MAAS framework is scalable and applicable to atmospheric observatories such as those operated by the Department of Energy (DoE) Atmospheric Radiation Measurement (ARM) facility.

Knowledge graphs have emerged as a powerful paradigm for structuring, organizing and reasoning over complex scientific knowledge, and are increasingly recognized as catalysts for accelerating AI for science. This study provides a comprehensive survey of scientific knowledge graphs (SciKGs), covering their construction methodologies and diverse applications across biology, chemistry and materials science. We examine how SciKGs support tasks such as drug development, omics analysis, reaction prediction and materials design, and highlight how the synergistic integration of SciKGs and large language models (LLMs) forms a knowledge- and language-driven framework for scientific discovery, in which SciKGs serve as the foundational knowledge infrastructure and LLMs act as dynamic semantic engines. We further identify key challenges and outline emerging opportunities for building auditable, interoperable and self-evolving SciKGs. Looking forward, we envision a new generation of SciKG-centered ecosystems where self-updating graphs, co-evolving with LLMs and embodied within AI scientists, become core infrastructures that autonomously drive, verify and accelerate scientific discovery.

Data collected by large-scale instruments, observatories, and sensor networks (i.e., science facilities) are key enablers of scientific discoveries in many disciplines. However, ensuring that these data can be accessed, integrated, and analyzed in a democratized and timely manner remains a challenge. In this article, we explore how state-of-the-art techniques for data discovery and access can be adapted to facilitate data and develop a conceptual framework for intelligent data access and discovery.

Event Abstract Back to Event NIQuery: Neuroimaging Informatics Query Framework for Data Sharing, Discovery, and Analysis B. Nolan Nichols1*, Robert F. Dougherty2, Landon T. Detwiler1, Gunnar Schaefer2, Randall J. Frank1, James F. Brinkley1, Brian A. Wandell2 and Thomas J. Grabowski1 1 Integrated Brain Imaging Center, University of Washington, Seattle, WA, United States 2 Center for Cognitive and Neurobiological Imaging, Stanford University, Stanford, CA, United States Scientific discovery about the human brain will be accelerated by neuroinformatics services on the programmable web. Just as bioinformatics databases provide services for molecular data, a scalable service-oriented framework is needed to take advantage of the large number of human neuroimaging data sets now available online. We have developed a specification, NIQuery, for remote access to observation-level data in distributed and heterogeneous image-specialized databases. NIQuery integrates emerging open source standards for metadata description, access & query, and provides investigators with computational access to voxel-level data. The protocol supporting this functionality consists of: 1) a persistent Session object that wraps databases (e.g. XNAT[1], NIMS[2], Allen Institute[3]), exposes the NIQuery application programming interface, and serves objects and requests; 2) a Query object that provides a mechanism to interrogate databases with user defined and/or predefined web-accessible queries; 3) a Data object conforming to a supported 'image' data model (e.g., DICOM, NIfTI, etc.) that provides a mechanism to return pixel data to an application; and 4) a Workflow object through which a server provides a computational service on a Data object. A registry service (www.niquery.org) provides an index of available NIQuery servers, as well as the query, data, and workflow objects available on each server. NIQuery enables client applications to discover shared neuroimaging data using metadata-level distributed queries and then execute image processing workflows on discovered data at their source, on a cached copy in the cloud, or locally. A sample implementation of this framework involves exporting a snapshot of XNAT and NIMS databases into XCEDE XML files, indexing the snapshots with the NIQuery registry service, and remotely calculating quality control metrics on resting-state fMRI data. These informatics tools will support agile exploration and reuse of open access neuroimaging data.

We suggest that the next era of Heliophysics should focus on the Sun–Heliosphere and Geospace as each a system-of-systems, and recommend a coordinated, deliberate, worldwide scientific effort to answer long-standing questions that will remain unanswered without a unified program. Many of the biggest unanswered science questions that remain across Heliophysics center around the interconnectivity of the different physical systems and the role of mesoscale dynamics in modulating, regulating, and controlling that interconnected behavior. Heliophysics has made key progress understanding both the large-scale dynamics and the microphysical processes that occur in these dynamic systems. Such understanding grew out of a systematic approach to study both limits of the system, from global, with the coordinated missions of the International Solar Terrestrial Physics (ISTP) program, to micro, with largely uncoordinated (albeit coincident) missions such as Cluster, Time History of Events and Macroscale Interactions during Substorms (THEMIS), Van Allen Probes, Magnetospheric Multiscale (MMS), Parker Solar Probe, and Solar Orbiter. We suggest that the international Heliophysics community should embark on a grand program to study these system-of-systems holistically, with coordinated, multipoint measurements. We particularly recommend an emphasis on resolving the mesoscale dynamics that links micro to global, and a whole-of-science approach that includes ground-based measurements and advanced numerical modeling. In effect, we propose a mesoscale ISTP type program that would consist of a system of Great Observatories capable of revealing the connections among systems from the solar interior to the top of Earth’s atmosphere. The paradigm and specific approaches outlined in this paper could serve as a strategic imperative and overarching theme that binds our Solar and Space Physics communities together under a common scientific objective. By its very nature, the type of program we argue for would be large, with several coordinated elements, and international in scope. It would include space-borne missions and coordinated ground-based observatories, artificial intelligence/machine learning (AI/ML) methods of analyzing large and complex datasets, and next-generation numerical modeling. The need to coordinate and integrate these different elements is independent of any specific mission implementation. Hence, we suggest the Heliophysics community organize around an ISTP-type program, ISTPNext, with associated Heliophysics “Great Observatories”.

The design and discovery of new materials is fundamental to advancing scientific and technological innovation. The recent emergence of the materials genome concept holds great promise in revolutionising materials science by enabling the systematic utilisation of data for efficient prediction and optimisation of ‘superior’ materials. However, the materials genome approach can be stymied by the vast complexity of design spaces, which often demand substantial computational resources and sophisticated data processing capabilities. To address these challenges, this work introduces a generative design framework called the non-dominant sorting optimisation-based generative adversarial networks (NSGAN). Capitalising on the synergies of genetic algorithms (GA) and generative adversarial networks (GANs), NSGAN provides a robust and efficient approach for tackling high-dimensional multi-objective optimisation design problems. To validate the efficacy of the proposed framework, we applied the model to a comprehensive dataset of aluminium alloys. Additionally, an online tool was created as a supplementary resource, offering a brief introduction to this innovative method for the wider scientific community. This study explores the potential of a predictive and data-driven approach in material design, indicating a promising pathway for widespread applications in the field of materials science.

We present a data-derived, ecosystem mapping approach for the global ocean as commissioned by the Group on Earth Observations (GEO) and as a contribution to the Marine Biodiversity Observation Network (MBON). These ecological marine units (EMUs) are comprised of a global point mesh framework, created from over 52 million points from NOAA’s World Ocean Atlas with a spatial resolution of 1 by 1 degree (∼27 x 27 km at the equator) at 44 varying depths and a temporal resolution that is currently decadal. Each point carries attributes of chemical and physical oceanographic structure (temperature, salinity, dissolved oxygen, nitrate, silicate, phosphate) as likely drivers of many marine ecosystem responses. We used a k-means statistical clustering algorithm to identify physically distinct, relatively homogenous, volumetric regions within the water column (the EMUs). Backwards stepwise discriminant analysis determined if all of six variables contributed significantly to the clustering, and a pseudo F-statistic gave us an optimum number of clusters worldwide at 37. A major intent of the EMUs is to support marine biodiversity conservation assessments, economic valuation studies of marine ecosystem goods and services, and studies of ocean acidification and other impacts. As such, they represent a rich geospatial accounting framework for these types of studies, as well as for scientific research on species distributions. To further benefit the community and facilitate collaborate knowledge building, data products are shared openly and interoperably via www.esri.com/ecological-marine-units. This includes provision of 3D point mesh and EMU clusters at the surface, bottom, and within the water column in varying formats via download, web services or web apps, as well as generic algorithms and GIS workflows that scale from global to regional and local. Work is in progress to delineate EMUs at finer spatial and temporal resolutions and to include ocean currents and various biodiversity observations. A major aim is for the ocean science community members to move the research forward with higher-resolution data from their own field studies or areas of interest, with the original EMU project team assisting with GIS implementation (especially via a new online discussion forum), and hosting of additional data products as needed.

As the Cosmology and Fundamental Physics (CFP) panel is fully aware, the next decade will see major advances in our understanding of these areas of research. To quote from their charge, these advances will occur in studies of the early universe, the microwave background, the reionization and galaxy formation up to virialization of protogalaxies, large scale structure, the intergalactic medium, the determination of cosmological parameters, dark matter, dark energy, tests of gravity, astronomically determined physical constants, and high energy physics using astronomical messengers. Central to the progress in these areas are the corresponding advances in laboratory astrophysics which are required for fully realizing the CFP scientific opportunities within the decade 2010-2020. Laboratory astrophysics comprises both theoretical and experimental studies of the underlying physics which produce the observed astrophysical processes. The 5 areas of laboratory astrophysics which we have identified as relevant to the CFP panel are atomic, molecular, plasma, nuclear, and particle physics. Here, Section 2 describes some of the new scientific opportunities and compelling scientific themes which will be enabled by advances in laboratory astrophysics. In Section 3, we provide the scientific context for these opportunities. Section 4 briefly discusses some of the experimental and theoretical advances in laboratory astrophysics required to realize the CFP scientific opportunities of the next decade. As requested in the Call for White Papers, Section 5 presents four central questions and one area with unusual discovery potential. Lastly, we give a short postlude in Section 6.

New intent discovery is of great value to natural language processing, allowing for a better understanding of user needs and providing friendly services. However, most existing methods struggle to capture the complicated semantics of discrete text representations when limited or no prior knowledge of labeled data is available. To tackle this problem, we propose a novel clustering framework, USNID, for <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">u</b> nsupervised and <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">s</b> emi-supervised <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">n</b> ew <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i</b> ntent <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</b> iscovery, which has three key technologies. First, it fully utilizes of unsupervised or semi-supervised data to mine shallow semantic similarity relations and provide well-initialized representations for clustering. Second, it designs a centroid-guided clustering mechanism to address the issue of cluster allocation inconsistency and provide high-quality self-supervised targets for representation learning. Third, it captures high-level semantics in unsupervised or semi-supervised data to discover fine-grained intent-wise clusters by optimizing both cluster-level and instance-level objectives. We also propose an effective method for estimating the cluster number in open-world scenarios without knowing the number of new intents beforehand. USNID performs exceptionally well on several benchmark intent datasets, achieving new state-of-the-art results in unsupervised and semi-supervised new intent discovery and demonstrating robust performance with different cluster numbers.

Galaxy evolution is complicated. Throughout their lifetimes, galaxies are subject to an amalgamation of astrophysical and cosmological processes that direct the growth of their stellar masses, the transformation of their morphologies, and the cessation of their star formation. The variable action of these processes begets a diverse population of galaxies, which exhibit a variety of brightnesses, colours, shapes, and sizes, among myriad other features. Many of these features are bimodally distributed, which has led to the general acceptance of a simple empirical paradigm of galaxy evolution. However, connecting this diversity among galaxies with the array of processes that are involved in their evolution, and constraining the relative influences of each of these processes, requires that several features are analysed simultaneously. This has been enabled by the recent advent of machine learning techniques, which are capable of extracting scientifically useful information from complicated, multi-dimensional datasets, to astronomy and astrophysics. Unsupervised machine learning techniques, free from the requirement for pre-labelled training data, are especially well suited to the exploration of the data structures of galaxy samples in multi-dimensional feature spaces. This thesis assesses the use of clustering, an unsupervised machine learning technique, for the research of galaxy evolution. Clustering is first tested on a well-characterised sample of galaxies from the GAMA survey. Galaxies are represented in five dimensions by a set of intrinsic astrophysical features. Use of a unique cluster evaluation framework enables the robust identification of reproducible and astrophysically meaningful clustering structures via the k-means method. Outcomes consisting of two, three, five, and six clusters are deemed stable, and form a hierarchical structure that agrees well with established notions of the galaxy bimodality. The two- and three-cluster outcomes are dominated in their structures by the stellar masses, colours, and star formation activity of galaxies, with Sersic indices and half-light radii becoming important for the five- and six-cluster outcomes. Clusters also exhibit broad correspondence with detailed morphological classifications, and it is suggested that the inclusion of additional morphological features might improve this correspondence further. The five- and six-cluster outcomes indicate the differential role of environment in the evolution of galaxies with intermediate colours. This cluster evaluation framework is then applied for the validation of the cosmological, hydrodynamical EAGLE simulations against the GAMA survey. Outcomes consisting of seven and five clusters respectively, determined using the same five features for both samples, are selected for analysis. These outcomes produce an agreement score of Vₐ = 0.76, indicating broad, overall agreement, but differences in their substructures. These differences include discrepancies in the growth of the central bulges of galaxies along the star-forming main sequence, an over-abundance of low-mass, bulge-dominated, star-forming galaxies in the EAGLE sample, and a subpopulation of high-mass, disc-dominated, star-forming galaxies in the EAGLE sample that is not present in the GAMA sample. These differences are attributed to the resolution of EAGLE, and to an active galactic nucleus feedback prescription that is not sufficiently effective in EAGLE. Finally, clustering is used to compare samples of galaxies at low (z ~ 0.06; GSWLC-2) and intermediate (z ~ 0.67; VIPERS) redshifts, in order to examine the evolution of subpopulations of galaxies. Galaxies are clustered in a nine-dimensional feature space defined by a series of ultraviolet-through-near-infrared colours using the Subspace Expectation-Maximisation algorithm, which includes iterative dimensionality reduction. The algorithm models both samples using seven clusters: four containing mostly star-forming galaxies, and three containing mostly passive galaxies. Both sets of star-forming clusters form clear morphological sequences, capturing the gradual internally-driven growth of galaxy bulges at both epochs. At high stellar masses, this growth is linked with quenching. However, it is only at low redshifts that additional, environmental processes appear to be involved in the evolution of low-mass passive galaxies. The results of this thesis demonstrate the utility of clustering as a method with which to analyse the large galaxy samples that are anticipated from next-generation surveys, and with which to facilitate the multi-dimensional comparison of cosmological galaxy simulations with observations. Clustering is robustly able to identify astrophysically meaningful substructures in complex, multi-dimensional feature spaces, and these substructures may readily be interpreted with respect to the evolutionary contexts of the galaxies that they encompass.