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

This paper demonstrates that the autonomous AI system PhysMaster can fully automate the complex, multi-step extraction of the Collins-Soper kernel from Lattice QCD data, reducing the workflow duration from months to hours while achieving precision consistent with state-of-the-art traditional methods and stabilizing signals at large transverse separations.

The Gist

The Big Picture: A New "Co-Pilot" for Physics

Imagine trying to solve a massive, 10,000-piece jigsaw puzzle where the pieces are constantly changing shape, the picture is blurry, and you have to do it all by hand while wearing heavy gloves. That is what doing Lattice QCD (a way of simulating the strong force that holds atoms together) has traditionally been like. It takes teams of physicists months of grueling work to get a single clear result.

This paper introduces PhysMaster, an "Autonomous AI Physicist." Think of PhysMaster not just as a calculator, but as a brilliant, tireless research assistant who can read textbooks, write code, run complex simulations, and check its own work, all while you sleep.

The team used this AI to solve a specific, decades-old puzzle: extracting the Collins–Soper (CS) Kernel.

---

The Problem: The "Fuzzy" Signal

To understand the inside of a proton (the particle in the center of an atom), physicists need to know how its internal parts (quarks) move. One key piece of information is the CS Kernel.

• The Analogy: Imagine trying to hear a whisper in a hurricane.
  • The "whisper" is the signal from the particles.
  • The "hurricane" is the noise that gets worse the further you look away from the center.
  • In physics terms, as you measure the distance between particles (transverse separation), the signal gets exponentially weaker and noisier.
  • Traditionally, human physicists had to manually clean up this noise, guess the right mathematical formulas to fit the data, and spend months trying to make the "whisper" audible. If they made a small mistake in the math, the whole result could be wrong.

The Solution: PhysMaster in Action

The authors used PhysMaster to automate this entire process. Here is how the AI tackled the problem, step-by-step:

1. The "Detective" Phase (Pre-Task)

Instead of a human spending weeks reading papers to figure out the rules, PhysMaster instantly scanned its "library" (a database of physics laws and previous studies).

• Analogy: It's like a detective walking into a crime scene, instantly reading every file in the precinct, and immediately knowing the rules of the game before even looking at the evidence.

2. The "Explorer" Phase (Task Execution)

This is the magic part. The AI used a strategy called Monte Carlo Tree Search (MCTS).

• The Analogy: Imagine a hiker trying to find the best path up a mountain in thick fog. A human might pick one path and hope for the best. PhysMaster is like a hiker who can instantly imagine thousands of different paths, try them out in a simulation, and instantly discard the dead ends, focusing only on the routes that look promising based on physics laws.
• It automatically tried different mathematical formulas to fit the noisy data. It didn't just guess; it tested thousands of variations to find the one that made the most physical sense.

3. The "Stabilizer" Phase (Fixing the Noise)

When the signal got too weak (the "hurricane" got too loud), PhysMaster didn't give up. It applied "physics-inspired constraints."

• The Analogy: Imagine you are trying to trace a shaky line on a piece of paper. A human might get frustrated and stop. PhysMaster is like a smart pen that knows, "Based on the laws of physics, this line must curve this way." It gently guides the shaky line back to a smooth, logical path, filling in the gaps where the data was too noisy to see.

The Results: From Months to Hours

The outcome was a game-changer:

• Speed: A process that usually takes months of human labor was completed in hours.
• Accuracy: The AI's results matched the best human-made results perfectly.
• Reach: The AI could see clearly into the "foggy" areas (large distances) where humans usually lose the signal. It stabilized the data up to 1 femtometer (a trillionth of a millimeter), a region that was previously very difficult to study.

The Bigger Meaning: A New Era of Science

This paper isn't just about one specific number; it's about a new way of doing science.

• The Old Way: Physicists spend 90% of their time doing the "plumbing" (cleaning data, fixing code, running fits) and only 10% thinking about the big ideas.
• The New Way: PhysMaster handles the plumbing. It acts as a Co-Scientist.
  • The Human: Provides the big picture, the physical intuition, and the "why."
  • The AI: Handles the "how," doing the heavy lifting, the math, and the endless testing.

Summary

This paper demonstrates that we can now use AI to automate the most difficult, boring, and error-prone parts of theoretical physics. By letting an AI "physicist" handle the complex math and data cleaning, human scientists can focus on discovery. It's like giving every physicist a super-powered assistant that never gets tired, never makes a typo, and can solve a 10,000-piece puzzle in an afternoon.

Technical Summary

1. Problem Statement

The extraction of the Collins–Soper (CS) kernel, $K(b_\perp, \mu)$, from Lattice Quantum Chromodynamics (QCD) is a critical task for understanding the 3D structure of nucleons via Transverse-Momentum-Dependent (TMD) parton distribution functions. However, traditional manual lattice analyses face three major bottlenecks:

• Signal-to-Noise Ratio (SNR) Degradation: Non-local correlation functions decay exponentially with increasing transverse separation ($b_\perp$), leading to unstable extractions in the large-$b_\perp$ region ($>0.5$ fm).
• Complex Systematic Uncertainties: The process requires intricate multi-step extrapolations (infinite-momentum, continuum, and chiral limits) and renormalization procedures, which introduce significant model dependence and human error.
• Labor-Intensive Workflows: The end-to-end process, involving data processing, high-dimensional fitting, and model validation, typically takes months of manual effort by human researchers.

2. Methodology: The PhysMaster System

The authors employ PhysMaster, an autonomous agentic AI system designed for theoretical and computational physics. PhysMaster integrates Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS) and a layered academic data universe (LANDAU) to automate the entire research workflow.

The methodology follows a three-stage autonomous workflow:

A. Pre-task Stage (Planning & Knowledge Retrieval)

• Task Decomposition: PhysMaster breaks down the CS kernel extraction into sub-tasks: correlator fitting, renormalization, extrapolation, and kernel extraction.
• Knowledge Base Construction: It builds a task-specific knowledge base by retrieving relevant literature (e.g., LaMET formalism, perturbative QCD predictions) and integrating prior lattice results into the LANDAU (Layered Academic DAta Universe) system. This ensures the AI adheres to physical constraints (symmetries, conservation laws).

B. Task Execution Stage (Exploration & Reconstruction)

• MCTS-Driven Exploration: The system uses Monte Carlo Tree Search to navigate the high-dimensional space of fitting parameters and model selection. It balances exploration (trying new parametrization forms) with exploitation (adhering to known physics constraints).
• Hierarchical Agent Collaboration:
  • Supervisor Agent: Manages global scheduling, evaluates intermediate results against LANDAU knowledge, and assigns scalar rewards to guide the search.
  • Theoretician Agent: Performs analytical derivations, writes code, and executes numerical fitting.
• Physics-Motivated Parametrization: To stabilize signals in the large-$b_\perp$ region, PhysMaster imposes a parametrization on the quasi-TMD wave functions (quasi-TMDWFs) constrained by perturbative QCD at small $b_\perp$ and lattice data at intermediate $b_\perp$. This replaces noisy data tails with smoothly continued values.

C. Post-task Stage (Validation & Reporting)

• Validation: The extracted CS kernels are compared against traditional lattice results and perturbative QCD predictions.
• Uncertainty Quantification: Systematic uncertainties are quantified, and structured reports are generated.

3. Key Contributions

• End-to-End Automation: The paper demonstrates the first fully autonomous extraction of a non-perturbative QCD observable (CS kernel) from raw lattice correlators to final results without human intervention in the engineering steps.
• Workflow Acceleration: The system reduces the duration of the analysis workflow from months to hours while maintaining or improving precision.
• Stabilization of Large-$b_\perp$ Signals: By imposing physics-inspired constraints via MCTS, the system successfully stabilizes signals up to $b_\perp \approx 1$ fm, a region where direct extraction methods typically fail due to noise.
• Reproducibility: The framework establishes a reproducible paradigm for automated lattice QCD studies, removing subjective choices in fitting ranges and parametrization forms.

4. Results

Using state-of-the-art lattice data (specifically the C32P23 ensemble from Ref. [21]), PhysMaster achieved the following:

• Consistency: The extracted CS kernel $K(b_\perp, \mu=2\text{ GeV})$ aligns with benchmark data from traditional manual calculations (LPC 25) within uncertainties.
• Signal Stability: The physics-constrained parametrization (Eqs. 7 & 8) effectively suppressed unphysical oscillations in the large-$b_\perp$ region, providing robust non-perturbative constraints up to 1 fm.
• Full Pipeline Execution: The system successfully executed the full chain:
  1. Computed ratios of nonlocal to local correlators.
  2. Identified plateau regions and performed one/two-state fits.
  3. Applied renormalization (removing linear and logarithmic divergences).
  4. Stabilized correlator tails using Bayesian priors.
  5. Performed Fourier transformations to momentum space.
  6. Extrapolated to the infinite-momentum limit ($P_z \to \infty$) to extract the final kernel.

5. Significance and Future Outlook

• Paradigm Shift: This work validates the "Physicist–AI collaboration" model, where AI handles labor-intensive numerical and analytical execution, freeing human researchers to focus on physical interpretation and theoretical innovation.
• Scalability: The framework is generalizable and can be extended to other non-perturbative observables, nucleon TMDs, and different lattice ensembles with minimal modification.
• Limitations & Future Work: Currently, PhysMaster takes existing lattice data as input; it does not yet generate raw lattice data. Future developments aim to incorporate neural network-based parametrizations to reduce model dependence and eventually achieve a fully end-to-end workflow including raw data generation.

In conclusion, this paper represents a milestone in AI-driven scientific discovery, demonstrating that autonomous agents can not only replicate but also optimize complex, high-stakes calculations in fundamental physics, paving the way for a new era of automated theoretical research.