Artificial Intelligence

AI for Nuclear and Particle Physics

Foundation Model for Nuclear and Particle Physics. Introduced FM4NPP, the first generalizable, large-scale AI foundation model for nuclear and particle physics that proves FMs can generalize effectively to sparse detector data. FM4NPP consistently outpaces other methods with improved downstream performance, data efficiency, and interpretability (arXiv 2025).

Unpaired Image-to-Image Translation. Developed UVCGAN (WACV 2023, ~200 citations), an AI image translator designed to work on unpaired images (e.g., cat and dog photos), used for domain shift correction between simulations and experimental data. Applied to ProtoDUNE Liquid Argon Time Projection Chamber (LArTPC) data (IOPscience 2024). Extended work to diffusion models for high-fidelity heavy-ion event simulations (Phys. Rev.C 2024).

Neural Data Compression. Designed AI models that compress vast data into smaller, more manageable size while keeping important details intact. Used for sparse 3D TPC detector data in collaboration with sPHENIX. Progressed from 3D convolutional neural networks (CNNs) (ICMLA 2021) to fast 2D variants (SC23, Best Paper Runner-up) and variable-rate neural compression (arXiv 2024).

AI for Materials Science

X-ray Absorption Spectroscopy (XAS) Data Analysis. XAS measures how materials absorb X-rays. Supervised machine learning was applied to XAS to predict 3D structures of metallic nanoparticles from XANES spectra (JPCL 2017).

X-ray Free Electron Laser (XFEL) Data Sorting and Imaging. Developed machine learning methods for high-repetition XFEL diffraction data analysis. XFELs can capture atoms moving in real time. (Npj Comput. Mater. 2024).

Accelerate Project. Leading efforts to create a foundation model for scanning electron microscopy (SEM) image analysis that enables generalizable, scalable analysis of SEM datasets (Collaboration with Brookhaven’s Center for Functional Nanomaterials).

Materials Insights with Neural Network Potentials. NNP AI models make atomic-level simulations of materials much faster and more scalable. NPPs have revealed how ions move in zinc battery electrolytes and how layered materials relax under strain with results confirmed by synchrotron experiments. These models help scientists design and engineer better materials with new mechanical properties (PRX Energy 2025, cover).

AI for Safeguards and Security

Event-based Camera Object Detection. Achieved state-of-the-art in event-based vision with EvRT-DETR, marking a breakthrough for fast, efficient detection via event-based imaging, which identifies objects using cameras that only record changes rather than full frames at set times (ICCV 2025, accepted).

Safeguards Surveillance. Used deep learning for automated detection, classification, and localization of safeguard-relevant objects. Extended to multi-camera tracking for enhanced nuclear monitoring. Affords support for inspectors via two-time, award-winning (DOE NNSA Joule Award) computer vision tools that speed up international inspections.

Satellite Image Analysis. CNN-based object detection models for satellite imagery, supporting automated geospatial intelligence applications (Partner: National Geospatial-Intelligence Agency).

Other methods for analysis and data processing techniques that nuclear safeguards and security communities rely on include:

• Smarter radiation analysis: Built the first AI models that can interpret radiation data in flexible ways, improving safety and security.
• Tracking uranium enrichment: AI methods that combine real and simulated data to spot unusual patterns and measure uncertainty.
• Preventing material theft: Tools that can detect if nuclear materials are being secretly removed from advanced reactors.
• Privacy-first AI: Using AI for arms control without revealing sensitive information.
• Robots for security: Trained the Spot robot to patrol reactors, check cameras, and detect radiation or other anomalies.
• Reactor monitoring: AI tools that help operators run reactors more efficiently and safely.
• Automated document review: Using AI to glean important information from reports for better risk assessments.
• Tamper-proof video: AI methods that can discern if security camera footage has been altered.

AI for Quantum

Measure and Program Quantum Systems for Machine Learning (ML). Created methods that learn to optimize quantum measurements for ML tasks, rather than using fixed measurement strategies. These quantum neural networks can adaptively choose what to measure, dramatically improving performance (ICASSP 2025).

Quantum Reinforcement Learning. Developed quantum reinforcement learning that uses adaptive non-local observables, learning which quantum properties provide the most useful information for decision-making. Applied to controlling charged particle beams in accelerators and power-grid systems. Introduced the “Quantum Rainbow,” a hybrid algorithm combining variational quantum circuits with the Rainbow Deep Q-Network (ICASSP 2025; IEEE-QCE 2024).

Transfer Learning in Variational Quantum Circuits. Examined how knowledge learned by quantum circuits on one task can transfer to related tasks, making quantum machine learning (QML) more practical by reducing expensive training.

Addressing Quantum Hardware Limitations with Multi-chip Ensembles. Built ensemble methods that combine multiple small quantum processors to solve larger problems more reliably, addressing current quantum computing limitations while maintaining theoretical advantages.

Quantum Applications in Biomedical Signal Processing. Exploring how quantum computing can help with brain research with tools that analyze brainwave signals (EEG) across different tasks and datasets, which could improve brain–computer interfaces and medical diagnosis. Created quantum models to study resting-state fMRI analysis brain scans, affording new ways to understand how the brain functions at rest (IEEE-QCE 2025, Best Paper).

Quantum Entanglement Certification and Verification. Created practical methods for certifying quantum entanglement through nonlocality measurements, crucial for verifying when QML provides genuine advantages over classical computing approaches.

Fundamental Quantum Machine Lear Methods. Focused on quantum solvers for high-dimensional partial differential equations and special-unitary parameterization techniques for improved variational quantum circuit training (Collaboration with Imperial College London).

Privacy-preserving Quantum Machine Learning. Novel approaches to ensure privacy in QML applications, including federated learning with differential privacy guarantees and quantum teacher ensemble methods. Techniques provide robust privacy protection while maintaining high accuracy, addressing critical concerns for deploying ML models on sensitive data (ICASSP 2024a,b,c).

Quantum Model Interpretability and Explainability. Methods that make QML models more interpretable and explainable, crucial for understanding how quantum models make decisions and building trust in quantum AI systems (SiPS 2024).

AI for Environmental and Biological Sciences

Representation Learning for Scientific Data. Developed representation learning and surrogate models for earth system science with AI emulators to replace time-consuming simulations, followed by adding uncertainty quantification to improve confidence in model predictions so scientists will trust them. Continuous field reconstruction from sparse data was extended to super-resolution then SCENT, a scalable, continuity-informed framework for complex problems (NeurIPS 2022; JAMES 2022; arXiv 2025).

Responsible and Explainable AI

Transparency and Interpretability of AI Models. Interactive visual analytics systems/tools that visualize how AI models behave with detailed results. Newer methods highlight which data are most important in shaping an AI model’s predictions, helping users better understand why a model makes certain decisions and enhancing trust in overall results (ICLR2025; 2021).

Scalable Anomaly Detection. Chimbuko (GitHub) is a scalable anomaly detection and visualization framework with an on/offline visualization module that automatically spots unusual behavior in data, in real time during large-scale computing runs to help diagnose problems before the system lags or crashes (IJHPCA 2025; ISAV 2020).

Scientific Literature Mining for Accelerated Experimental Discovery. Methods for natural language processing (NLP) and AI that have been applied to extract knowledge from scientific publications and augment studies of COVID-19 drug discovery, bioengineering, plant biology, biomedical sciences, nuclear physics, chemistry, and medical isotopes. These techniques include document layout analysis, NLP-based named entity recognition for domain-relevant terminology, inference of entity relationships and experimental ffects, information extraction from tables and figures, scientific knowledge benchmarks for modern LLMs, and optimal LLM prompting strategies (J. Comput. Biol. 2025; ACL 2024).