UNIST International Workshop on Machine Learning for Quantum Fields and Geometry 2025
Talks
Talk 1
Understanding diffusion models by Feynman's path integral
Akinori Tanaka (iTHEM, RIKEN)
Thursday, September 4, 2025
108-320, 10:00 - 10:45
Diffusion-based generative models have proven highly effective for image generation tasks, and their mathematical formulation can be reformulated from a quantum mechanical perspective. In this presentation, I will briefly explain this approach and then demonstrate how differences in performance between models with and without noise in the generation process can be analyzed through log-likelihood calculations by introducing a quantum mechanical counterpart to the Planck constant. This talk is based on arXiv:2403.11262.
Talk 2
Neural Ordinary Differential Equations for Mapping the Quark-Gluon Plasma via Holography
Song He (Ningbo University)
Thursday, September 4, 2025
108-320, 11:00 - 11:45
We construct a nonperturbative holographic QCD model that simultaneously reproduces the lattice equation of state and the temperature dependence of transport coefficients across the deconfinement crossover by neural ordinary differential equations (NO). We will demonstrate how this framework works in a typical example. Embedding a Gauss–Bonnet term nonminimally coupled to a dilaton within an Einstein–Maxwell–dilaton framework, we calibrate all potentials at zero chemical potential to lattice thermodynamics and compute the shear viscosity–to–entropy ratio η/s and bulk viscosity–to–entropy ratio ζ/s without small-coupling expansions. The model exhibits a high-temperature plateau in η/s above critical temperature T_c, a pronounced minimum of η/s, and a peak in ζ/s near the pseudo-critical temperature, in agreement with expectations from critical dynamics. This unified holographic construction shows that higher-curvature corrections tied to scalar dynamics capture both equilibrium and real-time properties of the quark–gluon plasma.
Talk 3
Learning 3 dimensional topology with Graph Neural Networks
Seong-Jin Lee (IBS-CGP, Pohang)
Thursday, September 4, 2025
108-320, 13:30 - 14:15
This talk explores the homeomorphism problem in low-dimensional topology in the context of plumbed 3-manifolds. With a brief introduction to geometric deep learning, we address the homeomorphism problem as a graph-learning task and test the efficacy of Graph Neural Networks (GNNs) in two distinct settings: supervised learning and reinforcement learning. Our results highlight the potential of GNNs to provide computational tools for paving the way for future applications to more general classes of 3 and 4 dimensional manifolds.
Talk 4
Machine Learning Ẑ Invariants for Plumbed 3-manifolds
Brandon Robinson (University of Amsterdam)
Thursday, September 4, 2025
108-320, 14:30 - 15:15
The application Machine Learning (ML) techniques in theoretical and mathematical physics offers powerful new methodological approaches to problems in which large amounts of data can be generated. In this talk, I will present the results of an ML based analysis using the q-series expansions of Ẑ-invariants for weakly negative definite plumbed 3-manifolds. Classifier networks trained on the q-series data are able to distinguish classes based on labels such as plumbing graph topology and 3-manifold homology. In each experiment, a systematic study of the logits of the network yields a robust explanation of the learned class-discriminative features, and tests of the effects of changing data representation, network architecture, and data compression are carried out.
Colloquium
Deep Learning Spacetime From Quantum Data
Keun-Yong Kim (GIST)
Thursday, September 4, 2025
108-320, 16:00 - 17:00
According to the holographic principle, gravitational dynamics in "bulk" spacetime are dual to the quantum physics of a system defined on its "boundary". We employ a deep learning approach to infer the bulk spacetime geometry from boundary quantum data, such as optical conductivity and entanglement entropy. In particular, we apply this method to gain insights into the properties of strange metals. Our approach is universal and broadly applicable to a wide range of physics problems involving differential equations and integral formulations. (References: 2502.10245 [hep-th], 2406.07395 [hep-th] 2401.00939 [hep-th,] 2011.13726 [physics.class-ph])
Talk 5
Machine-learning emergent spacetime from linear response
Daichi Takeda (iTHEMS, RIKEN)
Friday, September 5, 2025
108-320, 10:00 - 10:45
In the context of the AdS/CFT correspondence, the process of constructing a gravitational system equivalent to a given quantum system is called bulk reconstruction. This profound question is not only of theoretical interest but also holds the potential to pioneer quantum gravity experiments using quantum systems in the laboratory. In this talk, we propose a method to construct the metric of the equivalent gravitational theory with the help of a neural network (NN), under the assumption that data of thermal retarded Green’s functions of a quantum system have been obtained in a laboratory setting.
Talk 6
AI-based solution to the inverse problem in holography
Sejin Kim (KIAS, Seoul)
Friday, September 5, 2025
108-320, 11:00 - 11:45
Holography is a powerful tool for understanding strongly correlated systems and non-perturbative effects. However, applying holographic methods to real systems is challenging, because the precise holographic dual is often unknown. The inverse problem reconstructing the bulk theory from boundary data, makes it especially difficult to utilize gauge/gravity duality. To address this, we applied generative AI trained on data from entanglement entropy and superconductivity obeying gauge/gravity duality. This approach enables us to solve the inverse problem in holography and predict bulk theories immediately. Moreover, we can verify the AI's predictions by directly comparing them with results from gauge/gravity duality.
Talk 7
Do Androids Dream of Amoebae?
Jiakang Bao (University of Tokyo)
Friday, September 5, 2025
108-320, 13:30 - 14:15
In this talk, I will discuss the application of machine learning in the context of quiver gauge theories. In particular, the concept of the amoebae in mathematics has beautiful connections to the quivers. We will see how the analytic criteria determining the genus of the (lopsided) amoeba can be traced using the techniques of machine learning. I will also mention the possible directions that could help us get more useful results from AI in this context.
Talk 8
FantAIstic Beasts and How to Find Them
Shailesh Lal (BIMSA, Beijing)
Friday, September 5, 2025
108-320, 14:30 - 15:15
We present the R-Matrix Net, a neural architecture designed to solve the Yang–Baxter equation in difference form. The network identifies approximate solutions which are then used to extract exact, analytic families of Hamiltonians satisfying the Reshetikhin criterion identically. Applying this approach, we obtain several new integrable spin chains with three- and four-dimensional local Hilbert spaces, some admitting exact R-matrices. The framework is broadly applicable and not restricted to this particular class of equations. A significant part of the talk will provide a pedagogical but rigorous development of the method. In short: we train neural networks to locate integrable models - and point out some of their fantastic properties.
Talk 9
Sobolev Training for Operator Learning
Namkyeong Cho (Gachon University)
Friday, September 5, 2025
108-320, 16:00 - 16:45
This work investigates the impact of Sobolev Training on operator learning frameworks for improving model performance. Our research reveals that integrating derivative information into the loss function enhances the training process, and we propose a novel framework to approximate derivatives on irregular meshes in operator learning. Our findings are supported by both experimental evidence and theoretical analysis. This demonstrates the effectiveness of Sobolev Training in approximating the solution operators between infinite-dimensional spaces.
Talk 10
Learning Count Distributions from Bags of Graph Instances via Convolution
Chanho Min (Ajou University)
Friday, September 5, 2025
108-320, 17:00 - 17:45
We present a method for learning count-based labels from bags of graph-structured instances. Each instance is first processed with a Graph Neural Network (GNN) to estimate the probability of satisfying a target condition. The resulting probability set for a bag is modeled using a 1D convolutional network to produce a probability distribution over the possible counts of positive instances. This approach allows end-to-end training directly on count-level supervision, enabling effective learning in scenarios where only aggregate count labels are available for sets of complex structured data.
Talk 11
Data-driven discovery of self-similarity using neural networks
Yuji Hirono (University of Tsukuba)
Saturday, September 6, 2025
108-320, 10:00 - 10:45
Self-similarity provides crucial insight into the laws governing complex systems. In this talk, I present a neural network method that uncovers self-similarity directly from data by embedding scaling structures in a parametrized way. We demonstrate the approach on both synthetic and experimental data, showing how it enables model-independent discovery of self-similarity in physical systems.
Talk 12
Quantum advantage pursued by Hamiltonian simulation
Dongwook Ghim (IBS-CTPU, Daejeon)
Saturday, September 6, 2025
108-320, 11:00 - 11:45
Quantum advantage refers to a computational benefit achievable by quantum processors that is unattainable by its classical counterparts. While numerous quantum algorithms have been proposedwith the goal of demonstrating such an advantage, whether it has truly been achieved remains a subject of ongoing debate. This talk will introduce a Hamiltonian simulation on quantum processors as an alternative yet promising approach toward the quantum advantage. In particular, focusing on two key examples—ground state preparation and excited state spectroscopy— this talk will highlight how challenging physical problems can be addressed through quantum algorithms implemented on quantum hardware. If time permits, I will also explore how machine learning techniques can be applied to the design of physics-inspired quantum algorithms.