Search NASA⌕ Search

SEARCH · Search NASA

Results for “Foundation models”

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 127 records · Page 7
All sourcesNASA NTRSDOE OSTI
DOE OSTI2026

Uncertainty quantification of fireball features extracted from nuclear test films using computer vision

Films from the US’s historic nuclear testing era comprise the only extensive collection of imagery depicting high-yield detonations. These films offer unique insights into the characteristics of flows occurring on scales that are difficult to replicate experimentally, and they are a valuable source of data for the validation of models used to describe nuclear detonations. In recent work, we implemented modern computer vision and machine learning techniques to extract features of the fireball following nuclear detonation. With a training dataset of fireball films, we fine-tuned a You Only Look Once 11 (YOLO11) model to detect and track the fireball. Applied to a video, the outer bounding box produced in each frame by YOLO11 is used as an input prompt to Meta’s Segment Anything Model 2 (SAM2), which is shown to accurately predict the boundary of the fireball over time with high resolution. These state-of-the-art computer vision foundation models exhibit impressive visual accuracy in their results but lack an output of values that robustly quantify uncertainty in scientific applications. In this paper, we develop procedures for uncertainty quantification of extracted fireball features. We outline the application of a parallel attention mechanism to calculate uncertainty ranges that complement and better pose model validation data. This higher quality fireball validation data may serve to improve prognostic models describing nuclear detonations in support of nuclear forensic and emergency response activities.

Khristy, Joel [ORNL] (ORCID:0000000209963060)↗
DOE OSTI2023

GenSLMs: Genome-scale language models reveal SARS-CoV-2 evolutionary dynamics

We seek to transform how new and emergent variants of pandemic-causing viruses, specifically SARS-CoV-2, are identified and classified. By adapting large language models (LLMs) for genomic data, we build genome-scale language models (GenSLMs) which can learn the evolutionary landscape of SARS-CoV-2 genomes. By pre-training on over 110 million prokaryotic gene sequences and fine-tuning a SARS-CoV-2-specific model on 1.5 million genomes, we show that GenSLMs can accurately and rapidly identify variants of concern. Thus, to our knowledge, GenSLMs represents one of the first whole-genome scale foundation models which can generalize to other prediction tasks. We demonstrate scaling of GenSLMs on GPU-based supercomputers and AI-hardware accelerators utilizing 1.63 Zettaflops in training runs with a sustained performance of 121 PFLOPS in mixed precision and peak of 850 PFLOPS. We present initial scientific insights from examining GenSLMs in tracking evolutionary dynamics of SARS-CoV-2, paving the path to realizing this on large biological data.

Zvyagin, Maxim↗
DOE OSTI2026

PRIME: An evaluation framework for protein representation inference and generalization in viral mutation space

Background Protein language models (PLMs) have revolutionized protein fitness prediction, yet their application to rapidly evolving viral pathogens is often confounded by extreme sequence homology. This homology leads to “data leakage” in standard random validation splits, yielding inflated performance metrics that fail to translate into real-world biosurveillance utility. Results We present Protein Representation Inference for Mutation Evaluation (PRIME), a framework that integrates domain-specific fine-tuning with a rigorous position-stratified validation protocol to evaluate viral threats. Using a dataset of 347,432 SARS-CoV-2 receptor binding domain (RBD) sequences, we demonstrate that while random training data split yields deceptive R 2 values (> 0.90), they fail to generalize to novel mutational sites. By benchmarking models up to 650 M parameters, we show that domain-specific fine-tuning of the ESM-C 600 M model with correctly stratified data provides an initial demonstration of predictive signal for binding affinity and expression at unseen mutational sites of binding affinity and expression on unseen sites (R 2 ~0.23), a significant advancement over base foundation models which exhibit no predictive power (R 2 <0). PRIME’s embedding-based clustering identified 3.03% of bat coronavirus sequences as candidates for further experimental prioritization based on their functional similarity to human-infective strains in embedding space, offering a perspective complementary to traditional phylogenetic methods. Conclusion PRIME establishes a new benchmark for the application of PLMs in pathogen surveillance. Our findings demonstrate that state-of-the-art models and fine-tuning, when paired with stratified validation, provide biologically meaningful insights into pathogen evolution and zoonotic risk.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2026

Towards Anomaly Detection at the CMS High-Level Trigger System

Traditional trigger strategies in CMS typically rely on model-dependent selections or rigid kinematic cuts, risking the omission of unexpected exotic signatures. To address this, we propose a novel anomaly detection (AD) algorithm for the High-Level Trigger (HLT), designed to serve as a complementary second layer of filtering to the Level-1 AXOL1TL AD algorithm. We employ a transformer-based foundation model trained on a diverse ensemble of Standard Model processes. By combining a joint contrastive and classification objective, and using particle kinematics as inputs, the model learns to map events to a physics-informed latent space where anomalous events are isolated from dominant backgrounds. Preliminary results show that this strategy enhances the signal-to-background ratio across a range of rare SM and BSM scenarios. Furthermore, this work constitutes foundational R&D for the potential implementation of an analogous AD algorithm in the Level-1 trigger system for Phase-2.

Cruz, Roy [U. Wisconsin, Madison (main)] (ORCID:00↗
DOE OSTI2025

Fine-tuning machine-learned particle-flow reconstruction for new detector geometries in future colliders

We demonstrate transfer learning capabilities in a machine-learned algorithm trained for particle-flow reconstruction in high energy particle colliders. This paper presents a cross-detector fine-tuning study, where we initially pretrain the model on a large full simulation dataset from one detector design, and subsequently fine-tune the model on a sample with a different collider and detector design. Specifically, we use the Compact Linear Collider detector (CLICdet) model for the initial training set and demonstrate successful knowledge transfer to the CLIC-like detector (CLD) proposed for the Future Circular Collider in electron-positron mode. We show that with an order of magnitude less samples from the second dataset, we can achieve the same performance as a costly training from scratch, across particle-level and event-level performance metrics, including jet and missing transverse momentum resolution. Furthermore, we find that the fine-tuned model achieves comparable performance to the traditional rule-based particle-flow approach on event-level metrics after training on 100,000 CLD events, whereas a model trained from scratch requires at least 1 million CLD events to achieve similar reconstruction performance. To our knowledge, this represents the first full-simulation cross-detector transfer learning study for particle-flow reconstruction. These findings offer valuable insights towards building large foundation models that can be fine-tuned across different detector designs and geometries, helping to accelerate the development cycle for new detectors and opening the door to rapid detector design and optimization using machine learning.

43 PARTICLE ACCELERATORS↗
DOE OSTI2024

Decode the Workload: Training Deep Learning Models for Efficient Compute Cluster Representation

Monitoring the status of a high throughput computing cluster running computationally intensive production jobs is a crucial yet challenging system administration task due to the complexity of such systems. To this end, we train autoencoders using the Linux kernel CPU metrics of the cluster. Additionally, we explore assisting these models with graph neural networks to share information across threads within a compute node. The models are compared in terms of their ability to: 1) Produce a compressed latent representation that captures the salient features of the input, 2) Detect anomalous activity, and 3) Make distinction between different kinds of jobs run at Jefferson Lab. The goal is to have a robust encoder whose compressed embeddings are used for several downstream tasks. We extend this study further by deploying these models in a human-in-the-loop production-based setting for the anomaly detection task and discuss the associated implementation aspects such as continual learning and the criterion to generate alarms. This study represents a first step in the endeavor towards building self-supervised large-scale foundation models for computing centers.

Mohammed, Ahmed↗
DOE OSTI2025

A Perspective on Data and Privacy for AI in Healthcare [Industrial and Governmental Activities]

As large language models continue to push the bounds of AI model size, they are also being trained on unprecedented volumes of data. While individual hospitals are estimated to produce petabytes of data per year, only a small fraction is currently being used for developing AI models. Additionally, with such data resources available, healthcare is well-positioned to benefit from the current trends in AI. Moreover, the inherently multi-modal and longitudinal nature of clinical data – from omics to imaging to unstructured notes – provides a fertile ground for the development and application of cutting-edge architectures like foundation models.

Gounley, John [Oak Ridge National Laboratory (ORNL↗
DOE OSTI2026

Low-latency Jet Tagging for HL-LHC Using Transformer Architectures

Transformers are the state-of-the-art model architectures and widely used in application areas of machine learning. However the performance of such architectures is less well explored in the ultra-low latency domains where deployment on FPGAs or ASICs is required. Such domains include the trigger and data acquisition systems of the LHC experiments. We present a transformer-based algorithm for jet tagging built with the HGQ2 framework, which is able to produce a model with heterogeneous bitwidths for fast inference on FPGAs, as required in the trigger systems at the LHC experiments. The bitwidths are acquired during training by minimizing the total bit operations as an additional parameter. By allowing a bitwidth of zero, the model is pruned in-situ during training. Using this quantization-aware approach, our algorithm achieves state-of-the-art performance while also retaining permutation invariance which is a key property for particle physics applications. Due to the strength of transformers in representation learning, our work also serves as a stepping stone for the development of a larger foundation model for trigger applications.

Laatu, Lauri [Imperial Coll., London]↗
DOE OSTI2025

Data and Code for Understanding Generative AI Content with Embedding Models

This repository contains code for the experiments in the paper "Understanding Generative AI Content with Embedding Models". Constructing high-quality features is critical to any quantitative data analysis. While feature engineering was historically addressed by carefully hand-crafting data representations based on domain expertise, deep neural networks (DNNs) now offer a radically different approach. DNNs implicitly engineer features by transforming their input data into hidden feature vectors called embeddings. For embedding vectors produced by foundation models -- which are trained to be useful across many contexts -- we demonstrate that simple and well-studied dimensionality-reduction techniques such as Principal Component Analysis uncover inherent heterogeneity in input data concordant with human-understandable explanations. Of the many applications for this framework, we find empirical evidence that there is intrinsic separability between real samples and those generated by artificial intelligence (AI).

Vargas, Max [Pacific Northwest National Laboratory↗
DOE OSTI2024

Sub-pilot-scale Production of High-Value Products from U.S. Coals

Investigators from the University of Utah, University of Wyoming and Marshall University pursued a program to study the conversion of raw coal to high-value products of carbon fiber and silicon carbide. Team members also developed an initial framework for a data portal that can incorporate laboratory data on coal processing and product quality, and also work with tools for machine learning for data analysis, data visualization and economic assessment. Experimental R&D efforts focused on the conversion of raw coal to coal tar and other byproducts, and the resulting tar intermediates were upgraded to form anisotropic and isotropic pitch materials. These pitch materials were produced from coal using both thermal (pyrolysis) and chemical (mild solvolysis liquefaction) decomposition of raw coal. Four different coals were studied: Utah bituminous coal (Sufco), Wyoming PRB coal (Black Thunder), Illinois bituminous coal (Illinois #6), and West Virginia bituminous coal (Flying Eagle). Both metallurgical-grade coking coals and lower-grade steam coals were investigated, and controlled secondary gas-phase reactions were used during a two-stage pyrolysis process to induce cracking and condensation reactions among the pyrolytic tar species. This approach successfully improved the performance of the lower grade coals for yielding pitch materials, with properties more consistent with a commercial-grade pitch that had previously demonstrated success for quality carbon fiber production. The use of waste plastic materials was also studied, to help improve physical and chemical characteristics of the intermediate tars and final pitch product; in particular, for lowering the pitch softening point to an acceptable level for melt spinning carbon fiber. Mild solvolysis liquefaction was also used as a method for producing pitch for carbon fiber production. As expected, significantly higher pitch yields were obtained using this approach, and waste plastic materials were also successfully used to reduce pitch softening point to an acceptable level. The plastic materials were also utilized to create a solvent for the mild solvolysis process, and this plastic-derived solvent was shown to provide results consistent with more expensive commercial chemical solvents, and could thus avoid the need for costly recovery and recycle of a liquefaction solvent. Additional experimental R&D focused on the production of silicon carbide (β-SiC) from the residual char byproduct from pitch production, and also on the production of carbon fiber from the anisotropic pitch. SiC was successfully synthesized using a mixture of residual char and sandstone at a ratio of 1:1. Reaction temperature and residence time were optimized and yielded a product purity of 81%. For carbon fiber production, the most successful pitch samples were obtained from the mild solvolysis liquefaction approach, combined with the use of a plastic (HDPE)-derived solvent. Fiber properties improved over time as laboratory fiber production methodologies improved, and final yields of carbon fiber were obtained with a diameter of 12.14 ± 1.10 um, Modulus of 173.73 ± 15.25 GPa, and Tensile Strength of 1.04 ± 0.10 GPa. A proof-of-concept Modern Community Research Data Portal (MCRDP) was developed and deployed for coal and coal-derived pitch characterization, with the full support of (i) remote web-based access, (ii) distributed analysis, (iii) interactive visualization and exploration, (iv) shared and long-term data access, (v) advanced query capabilities and (vi) real-time collaboration. The Coal to Products Data Portal “coaltoproducts.org” provides researchers with space to store and share data within a project, tools for analyzing and understanding data for scientific investigation, and the ability to publish data to the broader community for reproducibility. The portal leverages the Material Commons 2.0 (MC) platform developed by the Center for PRedictive Integrated Structural Materials Science (PRISMS) of the University of Michigan, to achieve long-term longevity of data collections and, more importantly, collaborative science. A number of data visualization tools were also assessed and implemented for interrogating the experimental and modeling data. The machine learning portion of this project analyzed datasets from two different coal conversion processes performed on a diverse set of coal samples from both the coal pyrolysis experiments and the solvent liquefaction experiments. The work was initiated by exploring standard regression models on the pyrolysis data, aiming to understand the impact of sample characteristics and processing conditions on key product metrics. Over the course of the project, the focus expanded to include a variety of machine learning tools, delving into both supervised and unsupervised learning methods. Models tested on the pyrolysis data included linear, ridge, lasso, elastic-net, Gaussian process, random forest regression, and AutoSklearn, and the approach was continually refined to enhance predictive accuracy and model interpretability. Similar techniques were applied to the liquefaction data with an additional focus on feature engineering. Along with mesophase content, additional outputs of interest were the pitch yield, softening point, and QI content. Insights derived from these analyses are crucial in determining the factors influencing the quality and yield of coal-derived products. As the work progressed, the research evolved from foundational model comparisons to analyses of random forests, decision paths, and feature importance scores. A thorough market analysis was performed to examine the prospects of coal-based carbon fibers. The best opportunities for coal come from its lower and more stable price relative to petroleum, particularly for subbituminous coals, which is the primary advantage that a coal refinery may have over a petroleum refinery. Before a commercial CTP production facility can be modeled, however, several things need to be understood regarding the nature of the would-be coal refinery. These include the technology to be deployed, the size of facility, the volume(s) of co-product(s), and the waste and emissions profile of the plant. The volume of co-products and waste may be substantial and will require separate market analysis to ensure viability. In the near-term, the importance of coal tar pitch, in the form of carbon pitch, to the aluminum and steel industries is likely to overshadow the alternative use of this material as an input for carbon fiber. The importance of steel and aluminum in building materials, and the need for carbon materials in their manufacturing, will ensure that demand for these products remains for the long run. In addition, carbon fiber may also be the best substitute for steel and aluminum well into the future. While society will eventually be able to shift production of much of its electricity needs to renewables, it will not be able to shift away from fossil fuels for production of high-strength construction and vehicular materials. Demand for carbon fiber is expected to increase quickly, but the volume of carbon fiber and the amount of coal that would be needed to produce even a sizeable share of this market may still be relatively small compared to current coal production. Thus, other coal-based products like graphene, graphite, carbon foams, resins, and carbon-based building products will play important roles in sustaining coal production as coal-fired power generation continues to decline.

01 COAL, LIGNITE, AND PEAT↗
DOE OSTI2025

DNABERT-S: pioneering species differentiation with species-aware DNA embeddings

SUMMARY: We introduce DNABERT-S, a tailored genome model that develops species-aware embeddings to naturally cluster and segregate DNA sequences of different species in the embedding space. Differentiating species from genomic sequences (i.e. DNA and RNA) is vital yet challenging, since many real-world species remain uncharacterized, lacking known genomes for reference. Embedding-based methods are therefore used to differentiate species in an unsupervised manner. DNABERT-S builds upon a pre-trained genome foundation model named DNABERT-2. To encourage effective embeddings to error-prone long-read DNA sequences, we introduce Manifold Instance Mixup (MI-Mix), a contrastive objective that mixes the hidden representations of DNA sequences at randomly selected layers and trains the model to recognize and differentiate these mixed proportions at the output layer. We further enhance it with the proposed Curriculum Contrastive Learning (C2LR) strategy. Empirical results on 28 diverse datasets show DNABERT-S's effectiveness, especially in realistic label-scarce scenarios. For example, it identifies twice more species from a mixture of unlabeled genomic sequences, doubles the Adjusted Rand Index (ARI) in species clustering, and outperforms the top baseline's performance in 10-shot species classification with just a 2-shot training. AVAILABILITY AND IMPLEMENTATION: Model, codes, and data are publically available at https://github.com/MAGICS-LAB/DNABERT_S.

Zhou, Zhihan↗
DOE OSTI2025

A machine-learning-driven data labeling pipeline for scientific analysis in MLExchange

This study introduces a novel labeling pipeline to accelerate the labeling process of scientific data sets by using artificial intelligence (AI)-guided tagging techniques. This pipeline includes a set of interconnected web-based graphical user interfaces (GUIs), where Data Clinic and MLCoach enable the preparation of machine learning (ML) models for data reduction and classification, respectively, while Label Maker is used for label assignment. Throughout this pipeline, data can be accessed through a direct connection to a file system or through Tiled for access through Hypertext Transfer Protocol (HTTP). Our experimental results present three use cases where this labeling pipeline has been instrumental for the study of large X-ray scattering data sets in the area of pattern recognition, the remote analysis of resonant soft X-ray scattering data and the fine-tuning process of foundation models. These use cases highlight the labeling capabilities of this pipeline, including the ability to label large data sets in a short period of time, to perform remote data analysis while minimizing data movement and to enhance the fine-tuning process of complex ML models with human involvement.

Chavez, Tanny (ORCID:0000000193172896)↗
DOE OSTI2026

Opportunities in AI/ML for the Rubin LSST Dark Energy Science Collaboration

The Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) will produce unprecedented volumes of heterogeneous astronomical data (images, catalogs, and alerts) that challenge traditional analysis pipelines. The LSST Dark Energy Science Collaboration (DESC) aims to derive robust constraints on dark energy and dark matter from these data, requiring methods that are statistically powerful, scalable, and operationally reliable. Artificial intelligence and machine learning (AI/ML) are already embedded across DESC science workflows, from photometric redshifts and transient classification to weak lensing inference and cosmological simulations. Yet their utility for precision cosmology hinges on trustworthy uncertainty quantification, robustness to covariate shift and model misspecification, and reproducible integration within scientific pipelines. This white paper surveys the current landscape of AI/ML across DESC's primary cosmological probes and cross-cutting analyses, revealing that the same core methodologies and fundamental challenges recur across disparate science cases. Since progress on these cross-cutting challenges would benefit multiple probes simultaneously, we identify key methodological research priorities, including Bayesian inference at scale, physics-informed methods, validation frameworks, and active learning for discovery. With an eye on emerging techniques, we also explore the potential of the latest foundation model methodologies and LLM-driven agentic AI systems to reshape DESC workflows, provided their deployment is coupled with rigorous evaluation and governance. Finally, we discuss critical software, computing, data infrastructure, and human capital requirements for the successful deployment of these new methodologies, and consider associated risks and opportunities for broader coordination with external actors.

Aubourg, Eric [APC, Paris] (ORCID:000000025592023X↗
DOE OSTI2026

FATHOMS-RAG: A Framework for the Assessment of Thinking and Observation in Multimodal Systems that use Retrieval Augmented Generation

Retrieval-augmented generation (RAG) has emerged as a promising paradigm for improving factual accuracy in large language models (LLMs). We introduce a benchmark designed to evaluate RAG pipelines as a whole, evaluating a pipelines ability to ingest several modalities of information. We present (1) a curated dataset of 93 questions designed to evaluate a pipeline's ability to ingest textual data, tables, images, multimodal data, and cross-document multimodal data; (2) a phrase-level recall metric for correctness; (3) a nearest-neighbor embedding classifier in an attempt to classify pipeline hallucinations; (4) a comparative evaluation of 2 pipelines built with open-source retrieval mechanisms and 4 closed-source foundational models; and (5) a third-party human evaluation of the alignment of our correctness and hallucination metrics. We find that closed-source pipelines significantly outperform open-source pipelines in both the correctness and halucination metrics, with a wider performance gap in questions relying on multimodal and cross-document information. We also find after a human evaluation of our correctness and hallucination metric compared with our questions and pipeline responses, average agreement was 4.62 for correctness 4.53 for hallucination detection on a 1-5 Likert scale with 5 being strongly agree with our determination.

Hildebrand, Samuel [ORNL] (ORCID:0009000465963104)↗
DOE OSTI2025

LLMs for Mfg.—On the State of Large Language Models and Applications to Manufacturing

Additive Manufacturing (AM), referred to as 3D printing, has emerged as a key pillar of Industry 4.0 enabling layer-by-layer fabrication of intricate geometries from CAD models. In parallel, Large Language Models (LLMs), deep learning models for natural language generation trained on vast text corpora, have demonstrated unprecedented capabilities in understanding and generating human-like text. The convergence of these trends opens new opportunities at the intersection of AM and AI/ML, where LLMs can assist engineers and researchers in design, manufacture planning, and knowledge discovery. Recent academic work has begun to explore LLM applications in AM and adjacent fields, such as material science, mechanical engineering, and design for additive manufacturing. This exploration ranges from intelligent process planning to domain-specific knowledge retrieval. This survey provides a comprehensive review of current developments, focusing on peer-reviewed literature contributions that apply, adapt, and advance LLMs in general and domain-specific domains. We analyze state-of-the-art (SOTA) techniques, such as fine-tuning foundational models for specific domains, retrieval-augmented generation (RAG) pipelines, knowledge graph integration, and delve into the architectures and evaluation methods employed. The goal of this survey is to inform researchers and practitioners of the current capabilities and limitations of LLMs in general and in domain-specific applications, and to outline how these models are being tailored to meet the requirements of these applications.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Data readiness pipeline patterns for scientific AI at scale: Insights from climate, fusion, life sciences, and materials

This article examines how data readiness for AI principles apply to large scientific datasets used to train foundation models. We analyze archetypal workflows across four representative domains—climate, nuclear fusion, life sciences, and materials—to identify common preprocessing patterns and domain‐specific constraints. We introduce a two‐dimensional readiness model that combines canonical preprocessing patterns with a five‐level operational readiness scale, both tailored to high‐performance computing (HPC) environments. This construct helps outline key challenges in transforming large‐scale scientific data into formats suitable for scalable AI training. Together, these dimensions form a conceptual maturity matrix that characterizes scientific data readiness and guides infrastructure development toward standardized, cross‐domain support for scalable and reproducible AI for science. Finally, we evaluate this maturity matrix in the context of case studies including ClimaX (climate), AFLOW (materials), OpenFold (proteomics), and DIII‐D fusion disruption‐prediction workflows, from which we distill lessons learned and provide recommendations to guide practitioners in developing robust AI‐readiness pipelines. Finally, we discuss remaining cross‐cutting challenges that persist across scientific domains.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

yProv4ML: Effortless provenance tracking for machine learning systems

The rapid growth in interest in deep learning and foundation models (FMs) in particular, has attracted the attention of a diverse range of researchers thanks to their generalization ability. However, the advent of these techniques has also brought to light the lack of transparency and rigor in the way development is pursued. In particular, the inability to determine the number of epochs and other hyperparameters in advance presents challenges in identifying the best model. To address this challenge, machine learning frameworks such as MLFlow can automate the collection of this type of information. However, these tools capture data using proprietary formats and pose little attention to lineage. This paper proposes yProv4ML, a framework that captures provenance information generated during machine learning processes in PROV-JSON format, with minimal code modification.

Machine learning↗
DOE OSTI2025

A Representation Fusion Framework for Decoupling Diagnostic Information in Multimodal Learning

Modern medicine increasingly relies on multimodal data, ranging from clinical notes to imaging and genomics, to guide diagnosis and treatment. However, integrating these heterogeneous data sources in a principled and interpretable manner remains a major challenge. We present MODES (Multi-mOdal Disentangled Embedding Space), a representation fusion framework that explicitly separates shared and modality-specific factors of variation, offering a structured latent space for multimodal information that improves both prediction and interpretability. By leveraging pre-trained unimodal foundation models, MODES mitigates the dependency on extensive paired datasets, crucial in data-scarce clinical settings. We introduce a masking strategy that optimizes representation dimensionality by eliminating low-information dimensions, to achieve compact, information-rich representations. Our framework demonstrates superior performance in predicting diagnoses and phenotypes compared to unimodal and conventional fusion models. MODES also enables robust diagnostic inference in missing data scenarios, offering an opportunity toward interpretable and efficient multimodal diagnostics in personalized healthcare.

60 APPLIED LIFE SCIENCES↗