Chuyuan Liu Letter Fall 25

Dear Colleagues

I am writing to strongly support Chuyuan Liu's application for a postdoctoral fellowship at Cornell. Chuyuan is determined and independent and exceptionally talented. He is a skilled and creative programmer — easily the strongest technical member we have at CMU, student or postdoc. He has demonstrated the ability to design and implement modern machine learning architectures, to develop technically demanding codebases, and to independently execute complex analyses. His level of initiative and intellectual depth makes him an outstanding candidate for this position.

My group focuses on searches for new physics with di-Higgs (HH) production in the 4b final state. Chuyuan has been deeply involved in 4b analyses since joining in 2020. This channel is particularly challenging: it suffers from a combinatorial problem in b-jet assignment when forming Higgs boson candidates, and it faces large, hard-to-model backgrounds from QCD multi-jet production.

ZZ/ZH -> 4b

When Chuyuan joined, my group had just begun exploring solutions to these problems. He was the only graduate student on the effort, working alongside postdoc Patrick Bryant and myself. Our plan was to develop and test strategies in the ZZ/ZH → 4b analysis and later apply them to HH → 4b. As a junior graduate student, Chuyuan quickly began making important contributions—and he has not stopped since.

To address the combinatorics of b-jet pairing, we developed a custom classifier architecture for the 4b final state that simultaneously processes all jet pairings. This was a major learning exercise for my group both technically—mastering state-of-the-art ML tools—and scientifically, as it improved our understanding of the structure of the problem. For example, we learned it is a mistake to identify a single “best” jet pairing: the pairing optimal for reconstructing signal events is often sub-optimal for rejecting background. Ultimately, we built a model analogous to a graph neural network with an attention mechanism to combine information from all pairings. This network became the workhorse of the ZZ/ZH → 4b analysis. It was used to define the final discriminating variable, and to derive the background model and associated systematic uncertainties Chuyuan played a critical role in every aspect of its development, from initial prototyping to optimizing the user interface to allow the network to be adapted to other analyses.

The second major challenge in the 4b analysis is assessing systematic uncertainties on the data-driven multijet background. The QCD background is determined in a signal-free control region using a variant of the ABCD method, The background prediction in the signal region requires an extrapolation that is a significant source of systematic uncertainty. Conventional approaches for assessing this extrapolation uncertainty validate the background prediction in a third, statistically independent validation region. This strategy can address how accurately the background model extrapolates to a different region of phase space, but does not directly test the extrapolation into the signal region. In addition, it necessarily suffers from a lack of statistical power in the phase space with the highest signal-to-background ratio; the selection that makes the validation region background-dominated depletes the most sensitive phase space.

To overcome this, we introduced a new method using synthetic datasets obtained through hemisphere mixing. These datasets serve as high-fidelity proxies of the 4b background, enabling validation of extrapolations directly in the relevant phase space. Once again, Chuyuan was deeply involved in developing this technique—from debugging early implementations to deriving final systematic uncertainties from observed non-closure.

This work was published in European Physical Journal C, with Chuyuan as the only contributing student.

Fully Hadronic VHH -> 4b

Once the major pieces of the ZZ/ZH analysis were in place, Chuyuan approached me about extending our effort to VHH production in the fully-hadronic jj+4b final state. I was excited to see his initiative, but was initially hesitant. Would he really have time to see this and the ZZ/ZH analysis through? Won't the sensitivity be dominated by the leptonic V decays? In the end, Chuyuan did the work to convince me that this was worth pursuing, but I made him aware that he would largely be on his own.

Chuyuan sought out collaboration with Chris Palmer and Chyanit Asawatangtrakuldee. Chris and Chyanit's groups would cover the leptonic V decay modes and Chuyuan was to be responsible for the fully-hadronic final state. Chuyuan then independently ported our improvements to this more challenging final state; he adapted the graph network to also handle V->jj decays and validated the synthetic dataset in the more complex final state. I followed his work and gave him some high-level advice, but Chuyuan led and executed the analysis.

The result was a major success. It was the first VHH result published in this final state~\cite{2}. The fully hadronic channel proved to be among the most sensitive, and Chuyuan was honored to be the first to present the result publicly for at a conference.

Code development

The ZZ/ZH and VHH analyses were originally carried out in a C++ framework. Before tackling HH → 4b, Chuyuan pushed our group to migrate to a coffea/Python-based framework. This migration required new reproducibility tools and yielded major performance gains: workflows that previously took an hour and ~200 Condor jobs in C++ now run locally on a single LPC node in just 15 minutes.

Chuyuan also led the adoption of modern software practices—continuous integration, unit testing, and systematic result replication. Today, our codebase runs over 30 integration checks for every merge request, from unit tests of base functions to end-to-end tests of the full analysis pipeline. Here again, I was initially reluctant to invest in this infrastructure, but it has transformed our productivity and collaboration. Thanks to Chuyuan’s leadership, our software is more robust, our group more efficient, and our students better prepared for careers both within and beyond HEP.

HH->4b

We are now analyzing HH → 4b, applying the tools first developed for ZZ/ZH and extended by Chuyuan in VHH. Our group is part of the CMS-wide effort and is responsible for the Run-2 dataset. Our methods have been adopted as the baseline strategy for the combined Run-2 + Run-3 analysis. Chuyuan is once again playing a key role in advancing this result, which is currently undergoing CMS review. A draft for submission to Physical Review D is expected later this year.

In summary, Chuyuan has tackled some of the hardest problems in the 4b final state, developed novel solutions, and seen them through to published physics results. He has repeatedly extended these ideas into new contexts, taken intellectual risks, and executed them at a high level. Within my group, he already operates at the level of a postdoc. I give him my strongest possible recommendation for the Cornell fellowship.

and he is a crucial member of my group and our community at CMU. I would have no hesitation hiring chyuan as a postdoc in my group.