SCULPT: An Interactive Machine Learning Platform for Analyzing Multi-Particle Coincidence Data from Cold Target Recoil Ion Momentum Spectroscopy

The paper introduces SCULPT, an interactive web-based machine learning platform that utilizes advanced techniques like UMAP and adaptive confidence scoring to analyze high-dimensional multi-particle coincidence data from COLTRIMS experiments, thereby enabling efficient discovery of rare events and correlations in atomic and molecular physics.

The Gist

Imagine you are a detective trying to solve a crime, but instead of a few witnesses, you have millions of them, and they are all speaking a different language at the same time. This is the challenge scientists face when studying how molecules break apart.

The Problem: A Chaotic Crowd
In experiments called "Cold Target Recoil Ion Momentum Spectroscopy" (COLTRIMS), scientists shoot particles at molecules to see how they shatter. When a molecule like water breaks apart, it doesn't just split into two pieces; it can explode into five or more pieces (ions and electrons) all at once.

Every single "explosion" creates a massive amount of data. For one event, the computer records the speed and direction of every piece. If you add up all the angles, energies, and speeds, you end up with a list of 50 or more numbers for every single event. When you have millions of these events, it's like trying to find a specific pattern in a hurricane of data. Traditional methods are like looking at the hurricane through a keyhole; you only see one or two dimensions at a time, missing the bigger picture of how the pieces relate to each other.

The Solution: SCULPT
The authors of this paper present a new software tool called SCULPT (Supervised Clustering and Uncovering Latent Patterns with Training). Think of SCULPT as a smart, interactive 3D map generator that helps scientists navigate this data hurricane.

Here is how it works, using simple analogies:

1. The "Magic Map" (UMAP)

Imagine you have a giant, messy pile of colored marbles. Some are red, some blue, some green, but they are all mixed up in a 50-dimensional box that you can't see. You want to sort them by color.
SCULPT uses a technique called UMAP to flatten this 50-dimensional box down into a simple 2D map (like a flat sheet of paper).

• The Magic: It doesn't just squish the data; it intelligently arranges the marbles so that similar ones (those that broke apart in similar ways) end up next to each other, while different ones stay far apart. Suddenly, you can see distinct "islands" of colors that were previously hidden in the chaos.

2. The "Trust Meter" (Confidence Scoring)

When you look at a map, how do you know the islands are real and not just a trick of the light?
SCULPT includes a Trust Meter. It doesn't just show you the map; it calculates a score to tell you, "Hey, these groups are very distinct," or "Be careful, these groups might be overlapping."

• It checks the map using several different rules (like checking if the islands are tight together or if they are clearly separated from the empty space).
• It combines these checks into a single score. If the score is high, the scientist knows, "Okay, I can trust this grouping." If it's low, they know to try a different angle.

3. The "Filter" (Cleaning the Data)

Sometimes, the data is too noisy, like trying to hear a whisper in a crowded stadium.
SCULPT lets scientists act like a sound engineer. They can use filters to:

• Zoom in: Focus only on the loudest voices (the most common events).
• Tune the frequency: Ignore the background noise and only listen to specific types of sounds (specific energy levels or angles).
  This helps them isolate rare events that might be hidden in the crowd.

4. The "Auto-Pilot" (Genetic Programming)

Sometimes, scientists don't know which numbers to look at to solve the puzzle.
SCULPT has a feature that acts like an auto-pilot for discovery. It can automatically mix and match different numbers (like combining "speed" with "angle") to see if a new, hidden pattern emerges. It's like a chef who keeps trying new spice combinations until they find the perfect recipe that makes the flavors pop.

The Real-World Test: The Water Molecule

To prove it works, the team used SCULPT to analyze data from D2O (a heavy version of water).

• The Goal: They wanted to separate the different ways the water molecule could break apart. There were 8 different "quantum states" (different ways the molecule could be vibrating or spinning before it broke).
• The Result: Traditional methods struggled to separate these 8 states because their data looked very similar. SCULPT, however, successfully mapped them out. It found that some states were hiding inside the same "island" on the map. By using the Trust Meter and re-mapping specific sections, the software peeled them apart, revealing all 8 distinct states clearly.

Why This Matters

SCULPT is like giving scientists a high-tech microscope for data. Instead of spending weeks manually sorting through millions of numbers, they can interactively explore the data, find hidden patterns, and trust the results immediately. It turns a mountain of confusing numbers into a clear, navigable landscape, allowing researchers to spot rare and important events that were previously invisible.

The software is open and web-based, meaning any scientist can use it without needing to be a computer expert, making the complex world of molecular physics much more accessible.

Technical Summary

Technical Summary: SCULPT – An Interactive Machine Learning Platform for COLTRIMS Data Analysis

Problem Statement
Cold Target Recoil Ion Momentum Spectroscopy (COLTRIMS), or "reaction microscopy," enables kinematically complete measurements of multi-particle fragmentation processes, generating high-dimensional datasets (often exceeding 50 features per event, including 3D momentum vectors, mass-to-charge ratios, and derived quantities like Kinetic Energy Release). While powerful, the analysis of these tens of millions of events faces significant challenges with traditional methods:

1. Projection-based limitations: Reducing data to 1D or 2D histograms often obscures critical multi-dimensional correlations.
2. Sequential filtering bias: Applying cuts on specific parameters can introduce bias and hide unexpected patterns.
3. Manual feature engineering: Reliance on domain expertise to construct observables may miss non-intuitive relationships.
4. Tool incompatibility: Existing machine learning (ML) tools lack specific physics-informed constraints, configurable molecular profiles, and interactive exploration capabilities tailored to tabulated momentum spectroscopy data.

Methodology
The authors present SCULPT (Supervised Clustering and Uncovering Latent Patterns with Training), a web-based platform built on Python and the Plotly Dash framework. It integrates advanced ML techniques with physics-informed analysis through a modular architecture:

• Data Processing & Physics Features: The system allows users to define molecular profiles (particle types, masses, charges) without code modification. It automatically calculates physics features such as individual particle kinetic energies, Kinetic Energy Release (KER), electron energy sums/sharing, and inter-particle angles.
• Dimensionality Reduction:
  • UMAP: The primary tool for non-linear dimensionality reduction. It constructs a topological representation of high-dimensional data, preserving both local neighborhood relations and global structure. It is used to reveal non-linear correlations between momentum components and derived physics quantities.
  • Deep Autoencoders: A symmetric neural network (encoder-decoder) implemented in PyTorch is available for feature learning, compressing input data into a low-dimensional latent space to capture essential variance in fragmentation dynamics.
• Feature Discovery: The platform employs Genetic Programming using symbolic regression to discover interpretable mathematical combinations of input features, aiming to find non-trivial relationships that standard physics parameters might miss.
• Adaptive Confidence Scoring: A novel system provides quantitative reliability assessments for clustering results. It calculates a weighted confidence score ($C$) based on user-selected metrics:
  • Tier 1 (High Reliability): Silhouette Score (cohesion/separation) and Hopkins Statistic (clustering tendency).
  • Tier 2 (Moderate Reliability): Cluster Stability (via noise injection), Physics Consistency (variance ratios of key physics parameters), and Calinski-Harabasz Index.
  • Tier 3 (Low Reliability): Davies-Bouldin Index (heavily down-weighted due to sensitivity to non-spherical cluster shapes common in UMAP).
  • The system also tracks the Noise Ratio via DBSCAN clustering and applies asymptotic scaling bonuses for exceptional results while capping scores for severe failures.
• Interactive Filtering & Visualization: Users can filter data via feature selection, density-based filtering (removing sparse outliers in UMAP or feature space), and physics parameter filtering (e.g., specific KER ranges). Interactive selection tools (lasso, box) allow users to isolate events and re-analyze subsets.

Key Contributions

1. Physics-Informed ML Integration: SCULPT bridges the gap between generic ML tools and specific experimental requirements by integrating UMAP with automated physics feature calculation and configurable molecular systems.
2. Adaptive Reliability Assessment: The introduction of a weighted, multi-tier confidence scoring system addresses the lack of objective quality metrics in exploratory ML analysis of scientific data.
3. Iterative, Hypothesis-Driven Workflow: The platform enables a dynamic workflow where users can iteratively refine data subsets, adjust features, and re-cluster to isolate specific quantum states or reaction pathways, guided by both visual inspection and quantitative scores.
4. Automated Feature Discovery: The inclusion of genetic programming allows for the automated discovery of mathematical feature combinations that enhance cluster separation beyond standard physical observables.

Results: D2O Double Ionization Case Study
The platform was validated using a dataset of $\sim$1.9 million coincidence events from the photo-double ionization of D$_2$O (61 eV), resulting in D$^+$ + D$^+$ + O + 2e$^-$. The ground truth consisted of eight distinct dication quantum states.

• Initial Analysis: An initial UMAP projection using KER, total electron energy, total energy, and ion-ion angles identified five clusters. Two clusters contained single quantum states, while three were mixed.
• Iterative Refinement: By isolating mixed clusters and re-running UMAP with optimized feature sets, the authors successfully separated the remaining states. For instance, a cluster containing three states was split into two, and subsequently into individual states, with the confidence score increasing from 0.70 to 0.79 as the separation became physically distinct.
• Metric Performance: The confidence scoring system proved effective; high scores correlated with successful physical separation, while low scores (e.g., 0.14) correctly indicated poor separation when suboptimal features were chosen. The system successfully handled cases where quantum states had overlapping kinematic signatures (e.g., 3B$_1$ and 1B$_1$ states sharing similar KER), using the Hopkins statistic to validate underlying clustering tendency despite modest silhouette scores.

Significance and Claims
The paper claims that SCULPT represents a significant advance in analyzing multi-particle coincidence data by enabling researchers to extract detailed information from high-dimensional data more efficiently than established tools.

• Accessibility: The web-based, modular design makes advanced ML analysis accessible to the broader atomic and molecular physics community, reducing the time required for complex multi-dimensional analyses.
• Rare Event Detection: The platform facilitates the on-the-fly identification and isolation of rare events exhibiting non-linear correlations, potentially guiding experimental exploration and improving efficiency.
• Future Potential: The authors suggest the platform lays the foundation for AI-driven discovery in multi-particle quantum dynamics. They propose future integration of "digital-twin" simulations and agent-based capabilities to autonomously identify non-standard dissociation behaviors (such as the "slingshot mechanism" in water dication states) that violate the axial-recoil approximation, which are difficult to detect via traditional manual analysis.
• Broader Impact: While focused on COLTRIMS, the authors posit that the approach of combining dimensionality reduction with domain-specific constraints could benefit other fields dealing with large, multi-parameter tabulated datasets, such as quantum information science, pharmaceuticals, and aerospace.

The work is presented as a tool to support, rather than replace, physical intuition, providing a data-driven pathway to validate and interpret complex experimental results.