My physics research primarily revolves around computational physics, specifically simulating particle collisions. As members of the MLhad collaboration, we focus on the process of hadronizatoin, where quarks—the elementary building blocks—combine to form particles. Despite its fundamental nature, this process remains a physical black box, with no theoretical framework for its description, making it necessary to utilize machine learning (ML), particularly generative models.
Beyond my work in ML for Science, I have cultivated a deep fascination for ML research, leading me to serve as an Independent AI researcher on a diverse array of projects in generative art and LLM, resulting in publications. Over the past few years, I have become increasingly excited about LLM, AI reasoning, safe and scalable deep learning.
I am consistently enthusiastic about tackling new and meaningful projects, and I actively embrace the opportunity to expand my knowledge and skills.
Ph.D. in Physics, 2020 - Present
University of Cincinnati
BSc in Physics, 2019
Ruhr University Bochum
Responsibilities include (listed tasks resulted in publication(s) and talks at conferences):
Responsibilities include:
Responsibilities include:
In this paper, we present two novel training paradigms on how to train generative models on statistical data accessible in real-world experiments, to generate fine-grained particle information, which do not exist in measurement.
In this paper we introduce a model of hadronization based on invertible neural networks and a new training method for normalizing flows that improves the agreement between simulated and experimental data. We also demonstrate how to analyse statistical and modeling uncertainties in the generated distributions.
In this paper, we ‘hack’ generative models by injecting a small-scale trainable module between the intermediate layers of the model. This approach pushes their outputs away from the original training distribution towards a new objective.
The paper introduces a novel unsupervised text style transfer method using few-shot abstractive summarization. This approach aligns vector space embeddings of source and target texts, selects nearest neighbors based on semantic similarity, and reranks the summarization. This method significantly improves style transfer quality and achieves soa results in automatic evaluation metrics.
In this paper, a novel unbinned test statistic based on the Wasserstein distance is introduced to detect charge conjugation parity (CP) symmetry violation, demonstrated with CP asymmetric distributions in B0 and D0 decays. This statistic matches the sensitivity of traditional tests while providing more detailed insights into localized CP asymmetry.
The paper introduces a Monte Carlo-veto algorithm for efficient uncertainty estimation in collider-event simulations, applicable to the Lund string-fragmentation model. This approach enables analysis of different input parameters using a single simulated event set, streamlining uncertainty assessments in physics measurements.
In this paper, we calculate the leading and next-to-leading-logarithmic electroweak corrections for the charm-top-quark contribution to the CP violation parameter epsilon_K, which leads to the reduction of the theory prediction. A more accurate prediction in epsilon_K is crucial in particle physics as it measures CP violation in kaon decay, tests the Standard Model, and probes new physics.