Prometheus
Physics-Informed Machine Learning Framework
Unsupervised discovery of phase transitions in physical systems using variational autoencoders. From classical 2D Ising models to exotic quantum critical phenomena, achieving unprecedented accuracy in automated order parameter discovery.
YCRG Physics LabCritical temperature accuracy in 2D Ising model with 0.85 correlation to theoretical order parameters
Detection of infinite-randomness fixed point in quantum systems with activated scaling ψ = 0.48±0.08
Improvement over standard PCA methods through physics-informed constraints and symmetries
Our framework doesn't just identify phase transitions. It discovers the underlying order parameters and critical exponents without prior knowledge of the physics. From classical spin systems to quantum many-body states, Prometheus reveals the hidden structure of complex physical phenomena through unsupervised learning.
Project Prometheus
The Prometheus framework represents a breakthrough in unsupervised machine learning for physics discovery. Our research demonstrates that physics-informed variational autoencoders can automatically discover order parameters and phase transitions without prior knowledge of the underlying physical phenomena.
Through comprehensive validation across classical and quantum systems, we've achieved unprecedented accuracy in critical temperature prediction and order parameter discovery, including the first machine learning detection of exotic quantum critical phenomena.
Unlike traditional methods that require prior knowledge of order parameters, Prometheus learns to identify critical points directly from raw system configurations. The framework maps high-dimensional physical state spaces into compact latent representations, where phase boundaries emerge as discontinuities in the learned manifold structure.
Systems Studied
2D Ising Model
Classical spin system benchmark
3D Ising Model
Three-dimensional spin systems
J₁-J₂ Heisenberg
Frustrated magnetic systems
Disordered TFIM
Quantum critical phenomena
How Prometheus Works
Prometheus uses physics-informed variational autoencoders to compress high-dimensional spin configurations into low-dimensional latent spaces. Phase transitions manifest as sharp discontinuities in these learned representations, allowing automated detection without human supervision.
The framework incorporates physical symmetries and constraints directly into the neural network architecture, dramatically improving sample efficiency and interpretability compared to generic unsupervised methods. This physics-informed approach achieves an 89% improvement over standard PCA methods.
Physics-Informed VAE
Variational autoencoders with physical symmetries and constraints built into the architecture
Order Parameter Discovery
Automatic extraction of order parameters and critical exponents from raw configurations
Multi-System Validation
Validated across classical 2D/3D Ising, frustrated magnets, and quantum systems
Published Papers
Beam-Plasma Collective Oscillations in Intense Charged-Particle Beams: Dielectric Response Theory, Langmuir Wave Dispersion, and Unsupervised Detection via Prometheus
Applies the Prometheus framework to beam-plasma systems, using dielectric response theory and Langmuir wave dispersion analysis to detect collective oscillation modes in intense charged-particle beams.
Brandon Yee, Wilson Collins, Michael Iofin, Jiayi Fu
System
Beam-Plasma
Detection
Langmuir Waves
Submitted
CERN BL4S
From Classical to Quantum: Extending Prometheus to Uncover Quantum Critical Behavior in Disordered Transverse Field Ising Chains
First unsupervised ML detection of infinite-randomness fixed point in quantum systems, successfully extracting activated dynamical scaling ψ = 0.48 ± 0.08 in disordered TFIM.
Brandon Yee, Wilson Collins, Maximilian Rutkowski
Quantum Critical
h₂/J = 1.00 ± 0.02
Activated Scaling
ψ = 0.48 ± 0.08
Correlation
r = 0.97
Prometheus: Unsupervised Discovery of Phase Transitions in Three-Dimensional Spin Systems Using Variational Autoencoders
Extension to 3D Ising model with critical temperature T₂/J = 4.511 ± 0.005 and ≥70% accuracy in critical exponent extraction across multiple lattice sizes.
Brandon Yee, Wilson Collins, Maximilian Rutkowski
Critical Temp
T₂/J = 4.511 ± 0.005
Exponent Accuracy
≥70%
Configurations
50,000
Unsupervised Discovery of Intermediate Phase Order in the Frustrated J₁-J₂ Heisenberg Model via Prometheus Framework
Extends Prometheus to frustrated magnetic systems, discovering intermediate phase order in the J₁-J₂ Heisenberg model without prior knowledge of the complex phase diagram.
Brandon Yee, Wilson Collins, Maximilian Rutkowski
Model
J₁-J₂ Heisenberg
Award
CSEF 3rd Place
Framework
Prometheus
Prometheus: Unsupervised Discovery of Phase Transitions Through Physics-Informed Variational Autoencoders
Demonstrates unsupervised discovery of phase transitions in 2D Ising model using physics-informed VAE with 99.7% critical temperature accuracy and 0.85 correlation with theoretical order parameters.
Brandon Yee, Wilson Collins, Caden Wang, Mihir Tekal
Accuracy
99.7%
Correlation
0.85
Critical Exponent
β = 0.124±0.008
Access Research & Code
Prometheus is open research. Explore our papers, reproduce our results, and build on the framework for your own physics discovery.