Rare AI Lab

PI: Aobo Li· Halıcıoğlu Data Science Institute (HDSI) · Department of Physics · UC SAN DIEGO aol002@ucsd.edu
Nature is full of Rare Physics Events of interest. To search for them, we must encode nature's message into experimental data with rare event search detector, and then decode them to unlock new physics and propel fundamental scientific understanding forward. My group leverage Artificial Intelligence (AI) as a powerful tool to decipher nature's messages.
We're Hiring: Postdoctoral Scholar in Agentic ML Development for Nuclear Physics

The Rare AI Lab at UC San Diego (HDSI) is looking for a postdoc to develop agentic AI systems for dark matter and neutrino searches (LEGEND, KamLAND-Zen, XLZD, and a quantum-enabled axion haloscope). The postdoc will build on our agentic ML framework SIDERIUS and can take part in DOE Genesis Mission projects.

Apply by: December 15, 2026 for full consideration  ·  Start: between January and April 2027  ·  Appointment: 2 years, renewable

Details & Apply on AcademicJobsOnline
My Scientific/programmer profiles:
News:
  • We're hiring a postdoc! The Rare AI Lab is recruiting a Postdoctoral Scholar in Agentic ML Development for Nuclear Physics. Apply by December 15, 2026 for full consideration. Details in this Link.
  • Our AutoML agent framework SIDERIUS is now open source! It turns AutoML functions into reusable agent capabilities with budget-aware exploration, and on the TIDMAD dark matter benchmark it reaches a 63.8% valid-model proposal rate vs. 1.6% for a general-purpose Codex agent. Check out the framework, experiments & examples, and paper.
  • Aobo Li received four DOE Genesis Mission awards! He serves as Lead PI on Denoising Quantum Sensors: Agentic Algorithm Creation and Real-Time AI for Quantum-Enabled Rare Event Discovery and as Co-PI on three additional Genesis Mission projects. Details in this Link.
  • Sonata Simonaitis-Boyd was selected for the 2026 cohort of the LANL-SoCal Hub. Congratulations, Sonata!
  • Aobo Li is chairing the 0vββ AI Summer School; postdoc Yue Ma will give the LLM/Vibe Coding lecture, and PhD student Sonata Simonaitis-Boyd helped organize the event.
  • Aobo Li received the NSF CAREER Award! The project will develop AI algorithms for rare event searches in KamLAND-Zen and XLZD, including LLM-powered AI agents, AI-accelerated simulations, and the Rare AI for Science Ecosystem (RAISE) for training the next generation of Data Physicists.
  • Sonata Simonaitis-Boyd has started her 16-week term as DOE Science Undergraduate Laboratory Internship (SULI) fellowship, congratulations sonata!
  • Yue Ma presented our group's first AI agent result at APS Global Physics Summit: TIDMAD: A Sandbox for Physic-informed LLM Agents in Dark Matter Searches, congratulations Yue!
  • Aobo Li serves as the Program Co-Chair of APS Global Physics Summit 2026 at Denver, Colorado
  • Aobo Li organized the second installment of APS DNP AI in Nuclear Physics Experiment Workshop!
  • Our workshop paper Efficient optimization of COHERENT detector design parameters with the Rare Event Surrogate Model was accepted by NeurIPS 2025 Machine Learning for Physical Science workshop! Congratulations to co-first authors Brian Liu and Sonata Simonaitis-Boyd.
  • Our paper TIDMAD: Time Series Dataset for Discovering Dark Matter with AI Denoising was accepted by NeurIPS 2025 as Spotlight!
  • Our group has submitted 2 papers to NeurIPS 2025! Check out the manuscript here 1. Gravitational Wave Surrogate Model 2. Dark Matter Dataset & Benchmarking.
  • Min Zhong started quantitative research internship at Nomura Securities.
  • Our new work RESuM: A Rare Event Surrogate Model for Physics Detector Design has been accepted by ICLR 2025 as Spotlight! Check out the manuscript at this Link.
  • Aobo Li delivered machine learning bootcamp at Tohoku University, Japan. Check out the video. about using LLM to build machine learning model!
  • Alex Migala attended the UCSD NSF A3D3 HDR Hackathon and won both the best-performing model award and most creative model award for the sea-level prediction challenge. Congratulation Alex!
  • Aobo Li was invited to give a talk on "AI in Neutrinoless Double-Beta Decay Experiment" on DNP 2024.
  • Sonata Simonaitis-Boyd volunteered on the UCSD HDR Hackathon challenge at UCSD.
  • Sonata Simonaitis-Boyd published a workshop paper on NeurIPS 2024 Machine Learning for Physical Science workshop, Congratulations Sonata! Link.
  • Alex Migala published a workshop paper on NeurIPS 2024 Machine Learning for Physical Science workshop, Congratulations Alex! Link.
  • Alex Migala presented a poster on Fast Machine Learning 2024 conference and won the best poster award. Congratulation Alex!
  • Sonata Simonaitis-Boyd organized the A3D3 Merchandise design context.
  • Sonata Simonaitis-Boyd organized the first A3D3 Undergraduate Research Symposium.
  • Avi Mehta, Sophie Wang, and Owen Yang won HDSI undergraduate research scholarship, congratulations to all of them!
  • Our new paper TIDMAD: Time Series Dataset for Discovering Dark Matter with AI Denoising has been submitted to NeurIPS 2024 Dataset & Benchmarking Track! Please check out our manuscript at this Link.
  • Eugene Ku has started as an Postbaccalaureate Research at Argonne National Lab. Congratulations Eugene! He will work with Dr. Varuni Sastry on deploying Vision Transformer at scale.
  • Sonata Simonaitis-Boyd will start as a new PhD student at HDSI in fall 2024, Congratulations Sonata! She has also joined A3D3 NSF HDR institute as a trainee affiliates.
  • Aobo Li was invited to deliver plenry talk on the Conference of Science at Sanford Underground Research Facilities (CoSSURF) 2024! The talk title is Pushing Rare Event Search to the Limit with Machine Learning Algorithms.
  • Aobo Li was invited to deliver plenry talk on the The 22nd International Workshop on Advanced Computing and Analysis Techniques in Physics Research! The talk title is Detecting Rare Events Using Artificial Intelligence.Link
  • Aobo Li was invited to A3D3 seminar at A3D3 NSF Harnessing Data Revolution Insittute! The colloquium title is Fast and Slow: AI in Rare Event Search.Link
  • Aobo Li hosted the AI in Nuclear Physics Experiment workshop at 2023 APS DNP-JPS Joint Meeting! 6 machine learning experts are invited to showcase their work.
  • Jessica T. Fry's paper was accepted by NeurIPS 2023 Machine Learning for Physical Science Workshop. The paper tile is "Long Time Series Data Release from Broadband Axion Dark Matter Experiment". Congratz Jessica! Link
  • Aobo Li was invited to give physics colloquium at South Dakota School of Mines and Technology!
  • Majorana Demonstrator has released its calibration data for AI/ML benchmarking purpose. These data are labeled short time series data produced by High-Purity Germanium (HPGe) Detector arrays. These data are accessible in HDF5 format here. For more information, please read the data release note.
  • Katharina Kilgus (U. Tuebingen, GeM group) will start summer summer research projects at UNC, under the joint mentor of Aobo Li and Julieta Gruszko. This trip is funded by German Reinhardt Frank-Stifung Foundation.
  • Aobo Li was invited to give physics division seminar at Argonne National Laboratory!
  • Aobo Li was invited to give panel talk in Carolina Data Science Now! Link
  • Henry Nachman (UNC, GeM group) graduated from UNC Chapel Hill with highest honor, congratulations Henry!
  • Bounds from KamLAND-Zen on neutrinoless double-beta decay begin to probe the heart of neutrino mass inverted hierarchy parameter space! Link    Feature Article in Physics
  • KamNet has helped KamLAND-Zen reach the world's most sensitive search for 0𝑣ββ! Link

Research

Click here for selected publications

Nature

Neutrino

The "ghost particle" that travel through space at almost the speed of light and pass through ordinary matter almost undetected.

Dark Matter

Mysterious and elusive, dark matter is postulated to constitute about 85% of the matter in our universe.

Particle Physics Experiments

The LEGEND Experiment

Next-generation NLDBD experiment conducted at the INFN Gran Sasso National Laboratory (LNGS) in Italy, aiming at unambiguous discovery of NLDBD utilizing high purity Germanium detectors.

The KamLAND-Zen Experiment

The monolithic Xenon liquid scintillator detector experiment, carried out at the Kamioka Observatory in Japan, currently holds the world's most stringent limit in the search for neutrinoless double-beta decay (NLDBD).
Rendering of the XLZD liquid xenon time projection chamber

The XLZD Experiment

Next-generation liquid xenon observatory formed by the XENON, LZ, and DARWIN collaborations, using a 60-tonne dual-phase time projection chamber to search for WIMP dark matter, neutrinoless double-beta decay, neutrinos, and other rare processes.

AI/ML Research

Current funded initiatives apply these research directions through agentic AI, surrogate modeling, foundation models, quantum-sensor denoising, low-latency embedded AI, and Bayesian uncertainty quantification. These efforts include four DOE Genesis Mission Phase I awards, with Aobo Li serving as Lead PI on one project and Co-PI on three others. Learn more.

AI Agent

AI Agent

develop multi-agent AI system that is keen to the critical need of neutrino and dark matter detection, such as experiment monitoring, event reconsutrction, and physics data analysis. These agent will autonomously conduct research and validate result to produce rigorous scientific result.

Surrogate Model

Surrogate Model

building uncertainty aware surrogate model to drastically accelerate expensive GEANT4 simulation. This will drastically reduce the cost for designing the next-generation detector as well as conducting background modeling that was not previously possible.

Time Series Analysis

Time Series Analysis

uncover faint signals from the dark matter or gravitational wave. These signals are buried within ultra-long time series data from quantum sensor readouts. We aim to develop advanced AI algorithms to decipher these signal from the ample quantum noise.

Fast Machine Learning on FPGA

Fast Machine Learning on FPGA

develop and deploy AI algorithms onto FPGA chips to unveil partices in detector within a few millisecond

Physics

KamLAND-Zen

With the help of KamNet, KamLAND-Zen 800 yields worlds most stringent limit on the effective Majorana mass and the first search for Majorana neutrino in the inverted ordering region!

Majorana Demonstrator

The Majorana Demonstrator, a predecessor experiment of LEGEND, recently unveiled its final results in 2023; LEGEND-200 is presently in the initial stages of collecting its first set of physics data.

Group Members

AI Tutorials

Introduction

Welcome to this AI/ML tutorial repository, developed during my faculty career. This tutorial series aims to provide interdisciplinary training, specifically designed to help physicists cross disciplinary borders into the world of Artificial Intelligence.

No prior AI/ML expertise is needed. We start from zero. During these tutorials, you will have the opportunity to move beyond theory and play with real data derived from cutting-edge physics detectors. You will learn how to build neural network models from scratch using this data.

Note: This is a growing repository. We plan to release more physics detector datasets and create additional tutorials in the future.

Content Overview

  • Datasets: Seven AI-ready physics datasets, ranging from BASIC (e.g., Majorana Demonstrator, CUORE) through INTERMEDIATE (e.g., EXO-200, NEXT) to ADVANCED (nEXO signal denoising and TIDMAD for dark matter discovery).
  • Tutorials: Covers the essentials ("AI In a Nutshell"), coding cookbooks, using ChatGPT for ML, and two additional Machine Learning Course Series.
Suggested Learning Paths
  • The Essential Path (Quick Start):
    BASIC DATASET → 1. AI In a Nutshell (run Jupyter Notebook) → 3. Using ChatGPT (update Jupyter Notebook)
  • The Full Exploration (Deep Dive):
    BASIC DATASET → 1. Nutshell → 2. Cookbook → 3. ChatGPT → 4. Practical Series → 5. Bootcamp → ADVANCED DATASET.

Physics Datasets

AI-ready datasets from neutrinoless double-beta decay (0νββ), neutrino mass, and dark matter experiments, many of them prepared for the 0νββ AI Summer School.

Majorana Waveform Example
Fig 1. Sample Waveforms

Majorana Demonstrator Data Release

BASIC DATASET High Purity Germanium Detector Signals

This dataset contains over 3 million data points derived from the Majorana Demonstrator experiment, which utilizes High Purity Germanium Detectors to search for Neutrinoless Double-Beta Decay. Each data point represents a real time-series waveform generated by the detector.

  • Input Data: 1D Numpy vector with 4000 samples (Time Series).
  • Labels: 4 Binary Classification labels + 1 Energy Regression label.
  • Goal: Train ML models to use the waveform as input to predict the analysis labels.
  • Format: HDF5 (See Tutorial 1 for reading instructions).
Download Dataset (Zenodo) Read Description (arXiv)
CUORE single-pulse and pile-up waveforms
Fig 2. Single-Pulse (top) vs Pile-Up (bottom) Event

CUORE Pulse Shape Data Release

BASIC DATASET Cryogenic Calorimeter Pulses

This dataset contains 10,000 triggered events from the CUORE (Cryogenic Underground Observatory for Rare Events) experiment. Each event is a 10-second waveform sampled at 1 kHz, beginning with a 3-second pretrigger baseline. The dataset is evenly split between clean single-pulse events and pile-up events containing two or more pulses.

  • Input Data: 1D array with 10,000 samples (Waveform), eventId.
  • Labels: 1 Binary Classification label (clean single-pulse: 0, pile-up: 1). 90% of events are labeled for training; 10% are reserved for testing.
  • Goal: Identify pile-up events from the raw waveform.
  • Format: HDF5. Waveforms are raw; normalization parameters are provided to apply during data loading.
Download Dataset (Zenodo) Read Description (arXiv)
SuperNEMO simulated two-electron event display
Fig 3. Simulated Two-Electron Event

SuperNEMO Simulation Dataset

BASIC DATASET Electron Tracking Detector Events

Simulated data from the SuperNEMO Demonstrator, a detector that reconstructs individual electron trajectories. It covers four physical processes in the source foil: the 0νββ decay signal in 82Se, and three backgrounds: two-neutrino double-beta (2νββ) decay in 82Se and decays of the foil contaminants 208Tl and 214Bi.

  • Input Data: Numpy array of reconstructed event features (ev_no, E1, E2, tX, tY, tZ, tR, dY, dZ, theta, phiR, phiS).
  • Labels: 1 Multiclass Classification label (0nubb, 2nubb, Bi214, Tl208).
  • Goal: Separate the 0νββ signal from the three background processes.
  • Format: HDF5.
Download Dataset & Description (Zenodo)
EXO-200 event waveforms across wire and APD channels
Fig 4. Waveforms on Wire and APD Channels

EXO-200 Public Data Release

INTERMEDIATE DATASET Liquid Xenon TPC Multichannel Waveforms

A subset of the Th-228 calibration data from the EXO-200 experiment, designed for training and testing AI/ML algorithms. Each event includes raw waveforms from the charge (wire) and light (APD) channels, together with reconstructed quantities such as charge energy, light energy, rotated energy, and charge-cluster information.

  • Input Data: Waveforms (226 channels × 600 timesteps).
  • Labels: Charge_Cluster_Number (Multiclass Classification label).
  • Goal: Predict the number of charge clusters in each event directly from the raw waveforms.
  • Format: HDF5.
Download Dataset (Zenodo)
NEXT simulated signal and background tracks in 3D
Fig 5. Signal vs Background 3D Tracks

NEXT Simulation Dataset

INTERMEDIATE DATASET High-Pressure Xenon Gas TPC Tracks

A simulation of ~650k signal (0νββ) and ~500k background (Bi-214) events in the NEXT detector, a high-pressure gaseous xenon time projection chamber. Each event is a set of voxelized (x, y, z) energy depositions: signal events carry a two-electron signature, while background events show a single electron from a gamma interaction.

  • Input Data: Pandas dataframe with columns event_id, x, y, z, energy (3D point cloud).
  • Labels: 1 Binary Classification label (signal or background).
  • Goal: Distinguish two-electron 0νββ tracks from single-electron background tracks.
  • Format: HDF5.
Tip: The full dataset is large. Start with 10% of it (0nubb_part_1.tar + Bi_part_1.tar) before training on everything.
Download Dataset (Zenodo) Read Description (arXiv)
Simulated nEXO charge waveforms on 32 tile channels
Fig 6. Noisy Charge Waveforms on One Tile

nEXO Charge Readout Simulation

ADVANCED DATASET Liquid Xenon TPC Signal Denoising

83.5 million simulated double-beta decay events recorded by the charge collectors of a single-phase liquid xenon TPC, following the approximate design of the nEXO experiment. Charge is collected on tiles of 16 × 16 perpendicular electrode strips, and each event is confined to one tile: 32 channels of 1D charge waveforms over 1377 time steps (0.5 μs each). Every event comes in a clean ("noiseless") and a noisy version, along with its energy in MeV.

  • Input Data: Noisy event waveforms (EArrays, convertible to Numpy and Pandas; code provided).
  • Labels: Clean event (denoiser_img_clean) and energy.
  • Goal: Denoise the charge waveforms to recover the clean signal, and reconstruct the event energy.
  • Format: HDF5.
Download Dataset (Zenodo) Read Description (arXiv)
TIDMAD Denoising Example
Fig 7. Noisy Input vs Ground Truth

TIDMAD: Dark Matter Discovery with AI Denoising

ADVANCED DATASET Quantum-enabled Axion Detector Data

Generated by a cutting-edge quantum-enabled lumped-element axion detector, this dataset targets the search for dark matter (85% of the universe's matter). It is an ultra-long time series dataset (>1 TB) containing 8-bit integer samples.

The Logic: We inject a clean dark-matter-like signal (Ground Truth) and read it out via a quantum sensor (Noisy Input). The Task: Build a Neural Network to denoise the readout and recover the injected signal.

  • Structure: HDF5 format. Split into Training (20 files), Validation (20 files), and Science (208 files).
  • Benchmark 1: Denoising Score (calculated on Validation Set).
  • Benchmark 2: Dark Matter Limit (Physics result calculated on Science Set).
High Difficulty: This dataset is extremely large (>1TB). Beginners are strongly suggested to complete other tutorials first. If starting here, please begin with a single training file.
Download & Scripts Benchmark Paper

Tutorial Series

Awards

Teaching

Appointment

Assistant Professor, Chancellor's Joint Initiative

Halıcıoğlu Data Science Institute & Department of Physics
UC San Diego
Sept. 2023 - Present

CoSMS Fellow

UNC Chapel Hill

Advisor: Julieta Gruszko

Sept. 2020 - Aug. 2023

Education

Boston University

Doctor of Philosophy
Major: Physics

Advisor: Chritopher Grant

Thesis: The Tao and Zen of neutrinos: neutrinoless double beta decay in KamLAND-Zen 800
Sept. 2015 - Sept. 2020

University of Washington, Seattle

Bachelor of Science
Major: Physics

Advisor: Jason Detwiler

Sept. 2011 - June 2015
IP Address