Boosted decision tree reweighting of simulated neutrino interactions for $O(1)$ GeV neutrino cross section measurements

This paper presents a generic method using Boosted Decision Trees to multi-dimensionally reweight existing neutrino Monte Carlo simulations to match a target model, allowing for efficient reuse of legacy data without the need for new event generation.

The Gist

Imagine you are a master chef trying to recreate a legendary, secret recipe for a famous soup. You have a "Source" recipe (an old, slightly outdated cookbook), but you want your soup to taste exactly like the "Target" recipe (the modern, gold-standard version used by top chefs today).

The problem is, you can't just start from scratch. The old cookbook is all you have, and if you tried to cook every single variation of the soup from scratch to match the new one, you’d run out of ingredients and time before you even finished lunch.

This paper describes a clever mathematical "shortcut" to make that old recipe taste like the new one without actually cooking new soups.

The Problem: The "Old Cookbook" vs. The "New Standard"

In the world of particle physics, scientists use computer programs called Monte Carlo generators to simulate what happens when neutrinos (ghostly particles) hit atoms.

Think of these generators as "digital cookbooks." For years, scientists have used an old version (let's call it GENIE v2). Now, a new, much more accurate version has come out (GENIE v3). However, running these massive simulations is incredibly "expensive"—it takes huge amounts of supercomputer time and electricity. Scientists want to use their old, existing data but make it "act" like the new, better data.

The Solution: The "Smart Weighting" Trick

Instead of throwing away the old data and starting over, the researchers used a machine-learning tool called a Boosted Decision Tree (BDT).

Think of the BDT as a highly skilled food critic.

1. The critic tastes a spoonful of the "Old Recipe" soup.
2. The critic compares it to the "Gold Standard" soup.
3. The critic says: "This spoonful has too much salt, but not enough pepper."
4. Instead of adding salt or pepper (which you can't do to a soup that's already cooked), the critic assigns a "Weight" to that spoonful. They might say, "This spoonful is only 50% as good as it should be, so let's count it as half a serving," or "This spoonful is perfect, let's count it as two servings."

By giving every single simulated event a "weight" (a multiplier), the researchers can mathematically transform the old data so that, when you look at the big picture, the distributions of particles look exactly like the new, high-quality model.

How They Organized the Chaos

Neutrino collisions are messy. They can spit out protons, neutrons, and muons in all sorts of combinations. If you tried to fix everything at once, the math would explode.

To solve this, the researchers used "Event Categorization." Imagine instead of trying to fix the whole soup at once, you separate it into bowls:

• Bowl A: Soup with one carrot.
• Bowl B: Soup with two carrots.
• Bowl C: Soup with one carrot and a potato.

They trained a separate "critic" (the BDT) for each bowl. This made the job much easier and more precise.

Why Does This Matter?

This paper proves that this "weighting" trick actually works. They tested it on a specific measurement called Transverse Kinematic Imbalance (which is basically checking if the "debris" from a collision flies off in a balanced way).

The results were a success:

• The "Old Recipe" (after being weighted) looked almost identical to the "Gold Standard."
• It even worked for things they didn't specifically train the critic to look for, proving the "critic" actually understood the underlying physics, not just memorized the answers.

The Big Picture: This allows scientists to breathe a sigh of relief. They can take years of old, expensive computer simulations and "upgrade" them to modern standards instantly. It saves massive amounts of computing power and allows them to make much more accurate discoveries about the fundamental building blocks of our universe.

Technical Summary

Technical Summary: Boosted Decision Tree Reweighting of Simulated Neutrino Interactions

1. Problem Statement

In neutrino physics, estimating interaction cross sections requires high-fidelity Monte Carlo (MC) simulations to model detector efficiency, background predictions, and resolution effects. However, the simulation process is computationally expensive, particularly the detector response stage. When neutrino interaction models (generators) are updated (e.g., moving from GENIE v2 to GENIE v3), researchers typically face a dilemma: either spend massive computational resources regenerating full MC samples or work with legacy data that may contain systematic biases due to outdated physics assumptions.

A specific challenge in neutrino-nucleus interactions is the high dimensionality of the final-state particle content (varying numbers of protons, neutrons, and pions) and the complex kinematic correlations between them. Simply reweighting a single global variable is insufficient to capture these multi-dimensional differences.

2. Methodology

The authors propose a generic, multi-dimensional reweighting method using a Boosted Decision Tree (BDT) algorithm. The core objective is to transform a "source" MC sample (legacy GENIE v2) so that its reconstructed particle content and kinematics match a "target" model (modern GENIE v3 AR23 tune).

Key methodological components include:

• Event Categorization: To manage high dimensionality, the authors divide events into seven distinct categories based on "visible" particle topology (e.g., $1p0n$ for one proton and zero neutrons, $2pNn$ for two protons and $N$ neutrons). This focuses the reweighting on particles that exceed detector detection thresholds (50 MeV for protons, 10 MeV for neutrons in MINERvA).
• Variable Selection: Instead of attempting to match every theoretical parameter, the BDT is trained on detector-focused observables: the momenta ($p_x, p_y, p_z$) and calorimetric energy ($P_T^p$) of above-threshold particles.
• BDT Reweighter Algorithm: The method uses a gradient-boosting-like approach. It iteratively builds decision trees that partition the multi-dimensional space into "leaves." The algorithm maximizes a symmetrized $\chi^2$ to identify regions where the source and target distributions differ most. Each event is assigned a weight based on the ratio of the target to the source density in its respective leaf.
• Normalization: Since BDTs adjust the shape of distributions but not the total magnitude, a normalization constant is applied per category to match the total cross section of the target model.

3. Key Contributions

• Efficient Reuse of Legacy Data: The method provides a way to "upgrade" old simulation datasets to modern physics models without the need for full re-generation.
• Detector-Centric Approach: By training on reconstructible quantities rather than theoretical model parameters, the method ensures that the reweighted samples are directly useful for experimental measurements.
• Handling Discontinuities: The method successfully manages unphysical "spikes" in legacy generators (such as the 25 MeV energy subtraction artifact in GENIE v2) by assigning zero weights to those unphysical regions.
• Validation Framework: The paper provides a rigorous validation through "truth-to-reconstruction" migration studies and efficiency calculations.

4. Results

• Distribution Matching: The reweighted source sample ($v2'$) showed significant improvement in matching the target sample ($v3$). This was evidenced by the reduction of the Kolmogorov–Smirnov (K-S) test statistic ($D_{KS}$) across all trained and untrained (but correlated) variables.
• TKI Variable Reproduction: Even though Transverse Kinematic Imbalance (TKI) variables (like $\delta p_T, \delta \phi_T, \delta \alpha_T$) were not used in the training, the reweighting successfully reproduced their distributions. This demonstrates that the BDT correctly captured the underlying kinematic correlations.
• Efficiency and Unfolding:
  • The reweighting accurately recreated the target model's detector efficiency profiles.
  • In a realistic "unfolding" test (correcting for detector smearing), the bias introduced by using a reweighted migration matrix was found to be minimal and largely covered by existing systematic uncertainty bands.
• Sensitivity Analysis: The authors demonstrated that the quality of reweighting is highly dependent on the completeness of the training variables; including leading proton kinematics is essential for accurate TKI reproduction.

5. Significance

This work is highly significant for the neutrino physics community (including experiments like DUNE, MicroBooNE, and MINERvA) because it offers a computationally efficient path to maintaining state-of-the-art simulations. It bridges the gap between theoretical model evolution and experimental data analysis, allowing researchers to mitigate model-dependent systematic uncertainties by providing a robust, machine-learning-driven method to align legacy simulations with the most current physics understanding.