Abstract
Predicting quantum phase transitions is a fundamental challenge in computational physics, often requiring extrapolation beyond known data regimes. This paper presents a Gaussian Process (GP) regression framework that utilizes the Bayesian Information Criterion (BIC) to evolve kernel complexity. By training on smooth precursor functions within a single phase, the model successfully predicts sharp transitions into unseen phases, demonstrating a robust method for physical extrapolation.
Introduction: The Extrapolation Frontier
In machine learning for the physical sciences, models are predominantly used for interpolation—estimating values between known data points. However, the "holy grail" of physical modeling is extrapolation: predicting a system's behavior under conditions entirely outside the training set. This is particularly difficult when systems undergo phase transitions, where properties change non-analytically. Standard smooth models often fail to detect the "cliff" of a transition before reaching it.
Methodology: Bayesian Learning and Kernel Evolution
The researchers employ Gaussian Process (GP) regression, a Bayesian approach that provides a probability distribution over possible functions rather than a single "best guess." This naturally accounts for uncertainty, which is critical in extrapolation tasks.
The core innovation lies in the handling of kernels—functions that define similarity between data points. The authors use a greedy search algorithm to construct complex "composite" kernels by combining simpler ones (Linear, Matérn, etc.). To prevent overfitting, they apply the Bayesian Information Criterion (BIC), which penalizes excessive model complexity unless it is physically justified by the data.
Instead of modeling discontinuous jumps directly, the GP is trained on functions that evolve smoothly across transitions, such as energy dispersions. Sharp physical properties are then derived from these smooth predictions.
Key Results and Physical Insights
The method was tested on two classic condensed matter models: the Lattice Polaron Model and the Mean-Field Heisenberg Model.
- Predicting the Unseen: In the polaron model, a GP trained entirely within one phase successfully predicted the existence and location of two subsequent phase transitions. This suggests that "precursor" information about future transitions is encoded within the smooth behavior of a single phase.
- Quantitative Accuracy: When data from multiple phases were available, the GP's predictions of phase boundaries became nearly indistinguishable from rigorous, computationally expensive quantum calculations.
- Kernel Selection: For the Heisenberg model, the BIC-guided selection of RBF and Rational Quadratic kernels was essential. Simpler kernels failed to capture the evolution of the order parameter, highlighting that physical complexity requires appropriately complex mathematical structures.
Discussion and Conclusion
While powerful, the method faces challenges as kernels become more complex, leading to "rugged" optimization landscapes. The authors suggest multiple optimization restarts to ensure robustness. Future work aims to apply this framework to first-order transitions and integrate it into experimental workflows to guide researchers toward critical "tipping points" in new materials. By leveraging the Bayesian nature of GPs, this approach transforms machine learning from a tool for data fitting into a predictive engine for physical discovery.
Summary updated