HSG-12M: A Large-Scale Benchmark of Spatial Multigraphs from the Energy Spectra of Non-Hermitian Crystals

This paper introduces Poly2Graph, an automated pipeline for generating HSG-12M, a pioneering 16.7-million-scale dataset of spatial multigraphs derived from non-Hermitian crystal energy spectra, which bridges condensed matter physics and geometry-aware graph learning by preserving vital geometric information often discarded in existing benchmarks.

The Gist

Imagine you are trying to understand the "personality" of a complex machine, like a quantum crystal, just by looking at the sounds it makes. In the world of physics, these "sounds" are called energy spectra. For a long time, scientists have known that if you plot these sounds on a map, they form beautiful, intricate shapes—loops, spirals, and tangled webs. These shapes are like fingerprints; they tell you everything about how the machine works, how it conducts electricity, or if it has hidden superpowers.

However, there was a huge problem: No one had a map of these fingerprints.

Until now.

This paper introduces HSG-12M, a massive new library containing 12 million of these energy fingerprints. But it's not just a library; it's a revolution in how we use AI to discover new materials. Here is the simple breakdown:

1. The Problem: The "Manual Drawing" Bottleneck

Imagine you are an art critic trying to study the style of a painter. But instead of a gallery, you have to visit every single artist's studio, watch them paint one stroke at a time, and manually trace their work onto a piece of paper. That is what scientists were doing with these quantum crystals. They had to calculate the energy shapes by hand, one by one. It was slow, tedious, and impossible to do for millions of examples. Without a big dataset, AI (which needs mountains of data to learn) was stuck.

2. The Solution: The "Magic Printer" (Poly2Graph)

The authors built a tool called Poly2Graph. Think of this as a high-speed, automated 3D printer for data.

• How it works: You feed it the mathematical "recipe" of a crystal (its Hamiltonian).
• What it does: Instead of a human spending hours drawing, Poly2Graph instantly calculates the energy shape and turns it into a digital graph (a network of dots and lines).
• The Speed: It is 100,000 times faster than the old methods. It turned a job that would take centuries into a job that took a few days.

3. The Dataset: HSG-12M (The "Library of Fingerprints")

Using this magic printer, the team created HSG-12M.

• Size: It contains 11.6 million static shapes and 5.1 million shapes that show how the patterns change over time.
• Diversity: It covers 1,401 different types of crystal recipes.
• The "Secret Sauce" (Spatial Multigraphs): This is the most important part.
  • Most AI graph datasets are like simple subway maps: "Station A connects to Station B." There is only one line.
  • HSG-12M is like a real city map. Between Station A and Station B, there might be a highway, a bike path, a river, and a train track. All these are different connections between the same two points.
  • In physics, these "multiple paths" carry vital information. If you squish them into a single line (like most AI does), you lose the secret geometry. HSG-12M keeps all the paths separate, preserving the full 3D shape of the data.

4. Why This Matters: The "Reverse Engineering" Dream

Why do we care about 12 million weird shapes? Because of Inverse Design.

• The Old Way: "I have a crystal. Let me calculate what its energy shape looks like." (Forward problem).
• The New Way (AI): "I want a material that does this specific thing (like conducting electricity with zero resistance). What should the crystal look like?"
  • With HSG-12M, you can show an AI a desired shape (a fingerprint) and ask, "Which crystal recipe makes this?"
  • The AI can then suggest a shortlist of materials for scientists to build. This could lead to super-fast computers, better solar panels, or new medical sensors.

5. The "Universal Translator"

The paper also makes a fascinating claim: Everything is a graph.
The authors show that you can turn not just crystals, but also matrices (math grids) and polynomials (equations) into these same energy fingerprints.

• Imagine a universal translator that turns a complex math equation into a picture of a tangled knot.
• If you can turn any math problem into a graph, you can use the same AI tools to solve problems in physics, chemistry, and pure mathematics.

Summary Analogy

Think of the universe as a giant library of books (materials).

• Before: Scientists could only read a few pages of a few books by hand.
• Poly2Graph: A robot that instantly reads every book and draws a picture of its story.
• HSG-12M: A massive gallery of 12 million of these drawings, organized by genre.
• The Result: Now, if you want to write a story about a "super-material," you can look at the gallery, find the perfect picture, and tell the robot, "Build me the book that matches this picture."

This paper provides the data, the tool, and the theory to finally let AI help us design the materials of the future.

Technical Summary

Here is a detailed technical summary of the paper "HSG-12M: A Large-Scale Benchmark of Spatial Multigraphs from the Energy Spectra of Non-Hermitian Crystals."

1. Problem Statement

The integration of AI into scientific research is often hindered by a lack of large-scale, high-quality, domain-specific datasets. Specifically, in non-Hermitian quantum physics, the energy spectra of 1D crystals under open boundary conditions (OBC) form intricate geometric structures on the complex plane known as Hamiltonian Spectral Graphs.

• The Gap: These graphs serve as "fingerprints" for electronic behavior, containing rich topological and geometric information (loops, arcs, multiplicities) that go beyond standard topological invariants. However, their systematic study has been intractable because extraction traditionally relies on manual plotting and visual inspection, limiting research to small-scale toy examples.
• The Data Limitation: Existing graph machine learning benchmarks overwhelmingly assume simple graphs (at most one edge between nodes) and often lack spatial embedding. They fail to capture spatial multigraphs, where multiple geometrically distinct trajectories exist between nodes, a feature critical for representing physical systems like crystal spectra, urban networks, and biological structures.

2. Methodology

The authors propose a two-pronged solution: a high-performance automated pipeline (Poly2Graph) and a massive resulting dataset (HSG-12M).

A. Poly2Graph Pipeline

Poly2Graph is an open-source, end-to-end automated pipeline that converts arbitrary 1D crystal Hamiltonians into spectral graphs. It integrates non-Bloch band theory, algebraic geometry, and morphological image processing.

1. Input: Takes a Bloch Hamiltonian $H(z)$ or its characteristic polynomial $P(z, E) = \det[H(z) - EI]$.
2. Root Solving (The Bottleneck): For a discretized energy domain $\Omega$, it solves for the roots $\{z_i(E)\}$ of $P(z, E)=0$. To achieve a $10^5\times$ speedup over previous methods, it uses custom optimized root-solvers based on Frobenius companion matrices with parallel eigen-solvers (leveraging GPU acceleration via TensorFlow/PyTorch backends).
3. Spectral Potential & DOS: It computes the Spectral Potential (Ronkin function) $\Phi(E)$ and derives the Density of States (DOS) $\rho(E)$ as the Laplacian of $\Phi(E)$. Physically, the spectral graph corresponds to the "ridges" of this potential landscape.
4. Adaptive Resolution: To balance accuracy and cost, it employs a two-stage approach:
   • Coarse Identification: Computes DOS on a low-resolution grid to create a mask of the spectral graph region.
   • Refined Calculation: Recalculates DOS at high resolution (e.g., $1024 \times 1024$) only within the masked region.
5. Graph Extraction: The high-resolution DOS image is binarized and skeletonized (thinned to 1-pixel width). The skeleton is converted into a NetworkX MultiGraph, where nodes are junctions/leaves and edges store their full geometric path as ordered $(Re(E), Im(E))$ coordinates.

B. HSG-12M Dataset Construction

Using Poly2Graph, the authors distilled 177 TB of spectral potential data into a structured dataset.

• Scale: Contains 11.6 million static graphs and 5.1 million dynamic (temporal) graphs.
• Diversity: Covers 1,401 characteristic polynomial classes (Hamiltonian families), ranging from 1-band to 3-band models with varying hopping ranges.
• Nature of Data: These are Spatial Multigraphs. Unlike simple graphs, they preserve edge multiplicity and the continuous geometric shape of edges in the complex plane.
• Variants: Includes balanced subsets (HSG-one-band, HSG-two-band, HSG-three-band), an imbalanced topology-focused subset (HSG-topology), and a temporal dataset (T-HSG-5M) capturing continuous deformations of graphs as Hamiltonian parameters change.

3. Key Contributions

1. First Large-Scale Spatial Multigraph Benchmark: HSG-12M is the first dataset of its kind, addressing the critical gap in graph representation learning where existing benchmarks ignore edge multiplicity and spatial geometry.
2. Poly2Graph Tool: A high-performance, open-source tool that automates the mapping of algebraic objects (polynomials/matrices) to graphs, enabling researchers to generate custom spectral datasets.
3. Universal Algebra-to-Graph Link: The paper establishes that spectral graphs are universal topological fingerprints not just for Hamiltonians, but for any bivariate Laurent polynomial, univariate polynomial, vector, and matrix (via Toeplitz decomposition). This bridges graph learning with linear algebra and polynomial theory.
4. New Challenges for GNNs: The dataset introduces new difficulties for Graph Neural Networks (GNNs), specifically learning from spatial multiedges and handling extreme class diversity (1,401 classes).

4. Benchmark Results

The authors benchmarked 8 popular GNNs (GCN, GAT, GIN, GraphSAGE, etc.) on the HSG-12M variants.

• Performance Trends:
  • Edge Attributes Matter: Edge-aware models (e.g., GINE) significantly outperformed edge-agnostic models (e.g., GIN). On HSG-12M, GINE achieved ~46% accuracy vs. GIN's ~6%, proving that multi-edge geometry (length, curvature, DOS) carries irreducible signal.
  • GraphSAGE Dominance: Under fixed parameter budgets, GraphSAGE consistently achieved the best performance, suggesting its inductive bias is well-suited for spatial multigraphs or that attention mechanisms (GAT) struggle with the noise/multiplicity of dense edges.
  • Inverse Design Potential: While Top-1 accuracy was moderate (54% for GraphSAGE on the full set), **Top-10 accuracy was extremely high (95%)**. This indicates the models can effectively retrieve a small candidate set of Hamiltonian families, making the dataset highly valuable for inverse material design (finding a structure that yields a desired spectral graph).
• Challenges Identified: Performance degrades as graph complexity and class diversity increase. The gap between Top-1 and Top-10 suggests that current models lose fine-grained geometric details (e.g., curvature, torsion) when using simple edge summaries.

5. Significance and Impact

• For Condensed Matter Physics: Provides a data-driven pathway to discover exotic phases and design materials (e.g., acoustic metamaterials, photonic crystals) with targeted quantum properties by treating spectral graphs as topological fingerprints.
• For Graph Learning: Forces the community to move beyond simple graphs. It highlights the need for algorithms that can handle spatial multigraphs, preserving both connectivity topology and geometric edge information.
• For AI4Science: Demonstrates a scalable pipeline for converting physical laws (Hamiltonians) into machine-readable graph data, setting a precedent for other scientific domains where algebraic structures underpin physical phenomena.
• Open Science: Both the Poly2Graph code and the HSG-12M dataset are released under permissive licenses (MIT and CC BY 4.0), fostering reproducibility and further exploration in geometry-aware graph learning.

In summary, this work transforms a niche area of non-Hermitian physics into a massive, structured benchmark, simultaneously advancing the understanding of crystal spectra and pushing the boundaries of graph neural network capabilities.