Yongheng Xu

?

Postdoctoral Fellow

Short bio

I am a particle physicist with a broad interest in new physics at the intersection of theory and experiment, focusing on the phenomenology of beyond-standard-model physics in detectors. I was trained as a theorist during my master's studies beginning from 2016, and defended my thesis on the phenomenology of future electron colliders in 2019. I then transitioned to experimental work, joining the T2K neutrino experiment as a PhD student at Lancaster University in the UK, supervised by Dr. Laura Kormos. I defended my PhD thesis on statistical inference and data analysis of the T2K experiment in May 2023. I then joined UCLA as a postdoctoral researcher, working with Prof. Alvine Kamaha on the data analysis of the LUX-ZEPLIN experiment and the setup of a local lab. Currently, I am a postdoctoral research fellow in the theoretical physics section, working on my DSTrain project, developing AI-based simulation tools for particle detectors.

Research interests and hobbies

Beyond my work on the DSTrain project, I am an active member of the GAMBIT collaboration, where I contribute to the global statistical analysis of neutrino experiments, integrating experimental results with theoretical models to explore new physics scenarios. My involvement includes refining global fits and developing efficient computational techniques to assess constraints on neutrino oscillation parameters.

Additionally, I plan to continue my participation in the data analysis and simulation efforts of the LUX-ZEPLIN (LZ) dark matter direct detection experiment, which utilizes a dual phase liquid xenon TPC. My work in LZ focuses on improving the sensitivity to new physics in low-energy electron recoil signals, which are crucial for dark matter searches and rare event detection. This involves optimizing signal extraction techniques, refining background modeling, and developing robust statistical inference frameworks. By leveraging my expertise in both phenomenology and experimental techniques, I aim to bridge the gap between theoretical predictions and observational data in dark matter direct detection experiments.

When I don��t do particle physics, I spend my time travelling , cycling or doing nothing at all.

DSTrain project

Simulating dark matter and neutrino experiments with generative AI

Particle physics asks the most fundamental questions��what the universe is made of, how its forces work, and why matter exists at all��and the answers ripple far beyond academia. To probe nature at its smallest scales, we need exquisitely sensitive experiments that turn abstract theory into measurable reality: detectors that can register a handful of photons or electrons, timing systems that resolve billionths of a second, and analysis chains that sift rare signals from oceans of background. These experiments don��t just test ideas; they create the data that anchor our models, reveal new phenomena, and drive technological innovation in sensors, computation, and materials. In short, advancing particle physics requires advancing the experiments themselves��because every breakthrough begins as a trace in a detector.

Accurate simulation is a pillar of particle physics experiments. It links theory to measurable detector responses, guides experiment design, and converts limited beam or underground runtime into maximal scientific return. Better, faster simulations enable richer likelihoods, tighter systematics, and broader parameter scans��turning scarce data into decisive insight.

This project builds a machine-learning�Cdriven simulation stack for neutrino and dark-matter experiments, complementing today��s computationally demending, sequential GEANT4 �� microphysics �� electronics �� waveform pipelines with fast, data-centric generative models. The aim is an AI-powered event simulation tool that matches (and in some aspects surpasses) conventional tools in fidelity while cutting Monte Carlo costs by orders of magnitude��enabling richer likelihoods, broader parameter scans, and routine toy-experiments that were previously impractical. A flow chart of the proposed tool is shown below.

Methodologically, the utlimate goal is to leverage modern generative AI��diffusion models, transformers, GANs/VAEs��to learn detector response directly from data, conditioning on interaction ��prompts�� such as particle type and kinematics. Energy deposits in liquid xenon detectors manifest themselves as prompt scintillations (S1s) and delayed scintillations (S2s), which are believed to be combinations of single photons (SPhEs) and single electron (SEs). Currently, we are building a quick waveform-stacking pipeline for noble-liquid experiments, in which the foundational SE and single-photon-electron SPhE components are produced by validated machine-learning generators and later stacked up to form simualted events. The system will provide clear conditioning controls (particle type, kinematics, detector state), a stable truth schema, and a plug-and-play interface so generated mock data can be readily inserted into analysis chains.