Automated Extraction of Collins-Soper Kernel from Lattice QCD using An Autonomous AI Physicist System

We employ {PhysMaster}, an autonomous agentic AI system integrating theoretical reasoning, numerical computation, and exploitation strategies towards ultra-long horizon automation, to tackle long-standing challenges in non-perturbative lattice analyzes, including low signal-to-noise ratio at large transverse separation, complex systematic uncertainties, and labor-intensive manual workflows. Using the extraction of the CS kernel from quasi-transverse-momentum-dependent wave functions (quasi-TMDWFs) via large-momentum effective theory (LaMET) as a showcase, we demonstrate that \textsc{PhysMaster} automates high-dimensional fitting, renormalization, continuum-chiral extrapolation, and non-perturbative reconstruction in a fully autonomous manner. This framework drastically reduces the duration of the workflow from months to hours without compromising precision, stabilizes signals in the large-$b_\perp$ region to $1~\rm fm$, and produces results consistent with perturbative QCD and state-of-the-art traditional lattice calculations. This work validates the effectiveness of physicist-AI collaboration for first-principles QCD research and establishes a generalizable, reproducible paradigm for automated studies of parton structure and other non-perturbative observables from lattice QCD.