From Classical to Quantum: Extending Prometheus for Unsupervised Discovery of Phase Transitions in Three Dimensions and Quantum Systems

This paper extends the Prometheus unsupervised learning framework to 3D classical and quantum many-body systems, successfully detecting critical points, extracting accurate critical exponents, and identifying distinct universality classes—including exotic infinite-randomness criticality—without analytical guidance.

The Gist

Imagine you are a detective trying to solve a mystery: When does a material suddenly change its personality?

In physics, this is called a phase transition. Think of ice melting into water, or a magnet suddenly losing its magnetism when it gets too hot. Usually, scientists know exactly when this happens because they have a "cheat sheet" (mathematical formulas) for simple cases. But for complex 3D objects or weird quantum materials, the cheat sheets don't exist.

This paper introduces a new detective tool called Prometheus. It's an AI that doesn't need a cheat sheet. It just looks at the data and figures out the rules of the game on its own.

Here is how the paper breaks down, using simple analogies:

1. The Old Way vs. The New Way

• The Old Way (Supervised Learning): Imagine teaching a child to recognize cats by showing them 1,000 pictures labeled "Cat" and 1,000 labeled "Dog." The child learns to spot cats, but they can't tell you why a cat is a cat, nor can they spot a "Cat-Dog" hybrid they've never seen. They need a teacher.
• The New Way (Prometheus/Unsupervised): Imagine giving that child a box of 10,000 random animal photos with no labels. The child has to sort them into piles based on what looks similar. Eventually, they realize, "Hey, all the fluffy things with pointy ears go in one pile, and the scaly things go in another." They discover the categories themselves. Prometheus does this with atoms.

2. The First Challenge: Going from 2D to 3D

The team had previously taught Prometheus to solve a flat, 2D puzzle (like a checkerboard). But the real world is 3D (like a Rubik's cube).

• The Analogy: Think of a 2D map vs. a 3D city. In a 2D map, you just look left and right. In a 3D city, you have to look up, down, and around.
• The Result: They upgraded Prometheus to have "3D eyes" (using 3D convolutional neural networks). They fed it data from a giant 3D grid of atoms.
• The Win: Prometheus found the exact temperature where the material changes (the "melting point") with 99.99% accuracy, even though no human mathematician knew the exact answer beforehand. It also figured out the "rules of the game" (mathematical exponents) that describe how the change happens.

3. The Second Challenge: Entering the Quantum World

This is the really weird part. Classical physics is like a ball rolling down a hill. Quantum physics is like a ghost that can be in two places at once.

• The Analogy: Classical physics is like a crowd of people shuffling in a room (thermal heat). Quantum physics is like a single person who can be in the kitchen and the bedroom simultaneously until you look at them.
• The Problem: Standard AI doesn't understand "ghosts" (complex numbers and quantum waves).
• The Solution: They built a Quantum-Aware Prometheus (Q-VAE). Instead of just looking at pictures, this AI looks at "wave functions" (the mathematical description of the ghost). It uses a special "fidelity loss" function, which is like asking, "How close is this ghost to the real one?" rather than "How many pixels are wrong?"
• The Win: It successfully found the "tipping point" for quantum magnets, even though the physics is totally different from the 3D classical version.

4. The Grand Discovery: Finding the "Exotic"

The most exciting part of the paper is what happened when they added disorder (randomness) to the quantum system.

• The Analogy: Imagine a highway where traffic usually stops smoothly at a red light (standard phase transition). But then, imagine the road is full of potholes and random construction zones. The traffic doesn't just stop; it gets stuck in a weird, slow-motion jam that behaves totally differently.
• The Mystery: Physicists predicted that in these messy quantum systems, the "traffic jam" follows a strange rule called Activated Scaling. It's like the time it takes to get through the jam depends on the logarithm of the distance, not the distance itself. It's a very specific, exotic behavior.
• The Miracle: Prometheus didn't know this rule existed. No one told it to look for it. It just looked at the messy data, saw the patterns, and said, "This isn't a normal stop; this is a weird, slow-motion stop!"
• The Result: It calculated the "tunneling exponent" (a number describing this weird behavior) as 0.48, which is almost perfectly the theoretical prediction of 0.5.

Why This Matters

This paper proves that AI isn't just a tool for sorting known things. It can be a discovery engine.

1. It works where math fails: When equations are too hard to solve, Prometheus can still find the answers.
2. It works across universes: It can handle both "normal" hot/cold physics and "spooky" quantum physics with the same brain.
3. It finds the unknown: It didn't just find a point on a graph; it identified a new type of behavior that scientists had to guess existed, but hadn't proven yet.

In short: Prometheus is like a blindfolded explorer who can feel the texture of a new planet, map its mountains, and tell you exactly where the "magic" happens, all without ever having seen a map before.

Technical Summary

1. Problem Statement

The paper addresses the limitations of current machine learning (ML) approaches in condensed matter physics, specifically regarding phase transition discovery.

• Supervised Learning Limitations: Existing supervised methods require labeled data (known phase structures) to classify phases. They cannot discover new phases, identify critical points, extract critical exponents, or determine order parameters without prior knowledge.
• The "Unknown" Regime: Many physically interesting systems (e.g., 3D classical systems without exact solutions, quantum many-body systems, and disordered systems) lack analytical solutions. In these regimes, researchers rely on computationally expensive numerical methods (Monte Carlo, Exact Diagonalization) that often require manual tuning and hypothesis testing.
• Core Questions: The authors aim to answer two fundamental questions:
  1. Can an unsupervised framework scale to higher dimensions (3D) where exact analytical solutions are unavailable?
  2. Can the framework generalize from classical thermal transitions to quantum phase transitions (driven by quantum fluctuations at $T=0$)?

2. Methodology: The Extended Prometheus Framework

The authors extend their previous "Prometheus" framework, which utilizes Variational Autoencoders (VAEs) for unsupervised discovery. The core principle is that VAEs, when compressing high-dimensional data into a low-dimensional latent space, naturally organize the latent variables according to the dominant sources of variation. Near a phase transition, the order parameter is the dominant variation.

A. Architectural Extensions

1. 3D Classical Systems (3D Ising Model):

   • Architecture: A 3D Convolutional VAE designed to process volumetric spin configurations ($L \times L \times L$).
   • Encoder: Uses 3D convolutional layers to capture local spin correlations and domain walls in three dimensions.
   • Decoder: Mirrors the encoder using transposed 3D convolutions to reconstruct the spin configuration.
   • Input: Binary spin configurations $\sigma \in \{-1, +1\}$.

2. Quantum Systems (Transverse Field Ising Model - TFIM):

   • Challenge: Quantum ground states are complex-valued vectors in a Hilbert space, and global phase differences are physically irrelevant. Standard MSE loss fails here.
   • Architecture: A Quantum-Aware VAE (Q-VAE) using fully connected layers (to handle non-local entanglement).
   • Input Representation: Complex wavefunctions $|\psi\rangle$ are flattened into real vectors $x = [\text{Re}(c), \text{Im}(c)]$.
   • Loss Function: Replaces standard reconstruction loss with a Fidelity-Based Loss:
     $$\mathcal{L}_{quantum} = (1 - |\langle \psi | \psi_{recon} \rangle|^2) + \beta D_{KL}$$
     This ensures the model learns the physical state regardless of global phase.

B. Discovery Pipeline

The framework operates fully unsupervised through the following steps:

1. Order Parameter Discovery: Identifies the latent dimension $d^*$ with the maximum variance across the control parameter (temperature $T$ or field $h$). This dimension is hypothesized to be the order parameter.
2. Critical Point Detection: Uses an ensemble of methods:
   • Peak in latent susceptibility ($\chi_\phi$).
   • Minimum in reconstruction fidelity (or peak in error).
   • Maximum gradient of the latent mean.
   • Binder cumulant crossing (for finite-size scaling).
3. Exponent Extraction: Performs Finite-Size Scaling (FSS) analysis on the latent data to extract critical exponents ($\beta, \gamma, \nu, \eta$) via data collapse optimization.
4. Universality Class Identification: Compares extracted exponents against known theoretical classes using $\chi^2$ statistics.
5. Exotic Scaling Detection: Specifically tests for Activated Scaling (logarithmic divergence) vs. Power-Law scaling in disordered systems using Bayesian Information Criterion (BIC).

3. Key Contributions

1. Extension to 3D Classical Systems: Successfully applied unsupervised learning to the 3D Ising model (up to $L=32$), a system with no exact analytical solution.
2. Quantum Generalization: Developed the Q-VAE architecture, demonstrating that the same unsupervised principles apply to quantum phase transitions driven by quantum fluctuations.
3. Discovery of Exotic Criticality: For the first time, an unsupervised ML method detected Infinite-Randomness Fixed Point (IRFP) behavior in the disordered TFIM, identifying Activated Dynamical Scaling without prior knowledge of the functional form.
4. Unified Framework: Established a single methodology capable of spanning classical thermal transitions, quantum transitions, and disordered systems.

4. Key Results

A. 3D Ising Model (Classical)

• Critical Temperature ($T_c$): Detected $T_c/J = 4.511 \pm 0.005$, deviating from the literature value ($4.5115$) by only 0.01%.
• Order Parameter: The leading latent dimension correlated with physical magnetization with Pearson coefficient $r = 0.997$.
• Critical Exponents: Extracted exponents ($\beta, \gamma, \nu, \eta$) with ~72% accuracy relative to literature values.
• Universality: Correctly identified the 3D Ising universality class via $\chi^2$ analysis ($p=0.72$), rejecting other classes (Mean-field, 2D Ising, etc.).

B. Clean Transverse Field Ising Model (Quantum)

• Critical Point: Detected the quantum critical point at $h_c/J = 1.00 \pm 0.02$ (Exact: 1.0), achieving 2% accuracy.
• Order Parameter: Discovered ground state magnetization ($\langle \sigma^z \rangle$) with correlation $r = 0.97$.
• Exponents: All extracted quantum critical exponents fell within one standard deviation of exact values.

C. Disordered TFIM (Exotic Criticality)

• Activated Scaling: The framework automatically distinguished between conventional power-law scaling and Activated Scaling ($\ln \xi \sim |h-h_c|^{-\psi}$) in the presence of strong disorder.
• Tunneling Exponent: Extracted the tunneling exponent $\psi = 0.48 \pm 0.08$, consistent with the theoretical prediction $\psi = 0.5$.
• Significance: Statistical analysis ($\Delta \chi^2 = 12.3, p < 0.001$) strongly favored the activated scaling model over power-law, demonstrating the ability to identify qualitatively different types of critical behavior.

5. Significance and Impact

• Beyond Pattern Recognition: The paper demonstrates that unsupervised ML can move beyond simple classification to genuine scientific discovery. It can identify the type of critical behavior (e.g., activated vs. power-law) and extract quantitative physical parameters without human guidance.
• Validation in the "Unknown": By successfully analyzing the 3D Ising model (no exact solution) and the Disordered TFIM (complex physics), the framework proves its utility in regimes where analytical methods fail and traditional numerical methods require heavy manual intervention.
• Information-Theoretic Principle: The results suggest a deep connection between deep learning and statistical physics: VAEs discover order parameters because they represent the dominant source of variation in the data, a principle that holds true for both thermal and quantum fluctuations.
• Future Applications: The framework provides a template for exploring unknown phase diagrams in frustrated magnets, topological phases, and experimental data (e.g., neutron scattering, cold atoms), acting as a hypothesis generator for physicists.

In conclusion, the Prometheus framework establishes that unsupervised learning is a robust tool for exploring complex physical systems, capable of scaling to higher dimensions, generalizing to quantum mechanics, and discovering exotic phenomena like infinite-randomness criticality.