Robert de Mello Koch (Huzhou University)
Talk title: A two-phase perspective on deep learning dynamics
Abstract: Deep neural networks exhibit several striking phenomena whose relationship is not immediately obvious: grokking, in which generalization emerges long after training error has vanished; double descent, in which test error becomes non-monotonic near the interpolation threshold; and the information bottleneck, in which learned representations undergo a characteristic compression. Remarkably, numerical experiments indicate that these phenomena are governed by the same characteristic time scales. This suggests a unified picture in which learning proceeds through two distinct phases, separated by a dynamical crossover associated with the emergence of generalization. I will explain this two-phase perspective and then describe how a circuit representation of deep networks makes it possible to formulate an associated notion of complexity. In this language, learning can be viewed as a process that searches for and ultimately favors lower-complexity circuits. I will conclude with several possible directions for developing this viewpoint into a broader theory of generalization in deep learning.
Xin Gao (Sichuan University)
Talk title: From orientifold Calabi-Yau geometry to machine-learning screening
Abstract: Orientifold Calabi-Yau threefolds provide key geometric data for Type IIB compactifications, including O-plane structures, equivariant Hodge splittings, and tadpole constraints. I will describe our construction of orientifold Calabi-Yau hypersurfaces in toric varieties using proper divisor exchanges and multi-divisor reflections in MPCP-resolved geometries. I will then present a Multi-Head Attention classifier trained on resolved polytope vertex data to identify Calabi-Yau hypersurfaces admitting proper NID divisor-exchange involutions. The results show that such orientifold structure is statistically learnable and that machine learning can serve as an efficient screening tool before exact geometric verification.
Koji Hashimoto (Kyoto University) online
Talk title: Wasserstein space of quantum chaos
Abstract: Based on the conjecture [arXiv:2604.17649] that optimal transport provides holography, we study quantum chaotic systems to find their nobel features when seen from the Wasserstein space of the optimal transport: (1) The chaotic dimensional reduction, (2) Scrambling gives folded Wasserstein space, (3) Wasserstein branch is a quantum scar. The dimensional reduction is consistent with holography and the conjecture. This talk is based on the collaboration with Norihiro Tanahashi and Kentaroh Yoshida [https://arxiv.org/abs/2605.20995].
Song He (Ningbo University)
Talk title: Learning quark-gluon plasma properties holographically with neural ODEs
Abstract: We use neural ODEs and state-of-the-art lattice QCD thermodynamics to calibrate a holographic Einstein-Maxwell-Dilaton model with Gauss-Bonnet corrections. The model quantitatively reproduces the equation of state at zero and finite baryon chemical potential and predicts the phase structure in the T-$\mu_B$ plane, a temperature-dependent $\eta/s$, and a peaked $\zeta/s$, while remaining consistent with thermodynamic constraints.
Yuji Hirono (Tsukuba University)
Talk title: Extracting physics from data with machine learning
Abstract: Recent advances in machine learning are opening new ways to extract physical information from data. In this talk, I will present two examples. The first uses machine learning to construct an observable for the chiral magnetic effect while suppressing background contributions. The second identifies scaling variables and reveals self-similarity through data collapse.
Keun-Young Kim (Gwangju Institute of Science and Technology)
Talk title: Deep learning bulk spacetime from boundary quantum data
Abstract: According to the holographic principle—one of the most influential frameworks in contemporary physics—gravitational physics in a bulk spacetime is dual to the quantum physics of a system defined on its boundary. We employ a deep learning approach to reconstruct the bulk spacetime geometry from boundary quantum data, such as conductivity and entanglement entropy. This method offers novel insights into the properties of quantum matter through their dual spacetime interpretation. Because our deep learning approach generalizes to a wide range of problems involving differential equations and integral formulations, it holds broad utility for diverse applications across physics and engineering. We illustrate this generalizability with concrete examples drawn from introductory (freshman-level) physics problems.
Nakwoo Kim (Kyung Hee University)
Talk title: Machine-learning toric Sasaki-Einstein metrics
Abstract: We solve for Sasaki-Einstein metrics in five dimensions with a neural network and validate the solver with both positive and negative results. The unknown is a single convex function on a two-dimensional polygon, the symplectic potential of the six-dimensional Ricci-flat Kahler cone whose link is our target metric. As a test of the integrity of our pipeline, we demand not only that the solver find metrics that exist, but that it fail recognizably when none does. In particular, on the cone over the third del Pezzo surface it recovers the Kahler-Einstein metric of Doran et al. (2007) with no symmetry imposed, and the $D_6$ symmetry of the hexagon emerges unprompted from the equation alone. On the other hand, on the cone over the second del Pezzo at the regular Reeb vector, where no solution exists, its out-of-sample residual instead sticks at a nonzero floor under an eleven-fold increase in parameters, a fingerprint of non-existence. The toric frame additionally supplies a linear ansatz, an expansion in lattice-invariant polynomials, which drives the same residual to between $10^{-12}$ and $10^{-17}$. We use the polynomial solutions as the reference the network is measured against.
Seong-Jin Lee (Institute for Basic Science - Center for Geometry and Physics)
Talk title: Explainable AI diagnoses 4d N=1 gauge theory for local defects
Abstract:
Yuanche Liu (University of Science and Technology of China)
Talk title: “Learning” Feynman integrals
Abstract: Feynman-integral calculations involve two complementary challenges: reconstructing exact analytic structures from numerical data and finding a basis in which those structures become manifest. I will present two AI-assisted approaches. First, symbolic regression is used to infer compact exact expressions from numerical canonical differential equations, with every proposal verified algebraically. Second, I will introduce CANON, a stateful AI-agent framework that coordinates IBP reduction, differential equations, and several strategies for generating uniform-transcendental candidates. The search proceeds sector by sector, while accepted results, failed candidates, and diagnostics are recorded in a persistent scientific state. From these two examples we can see how AI influence theoretical physics research in this era.
Rak-Kyeong Seong (Ulsan National Institute of Science and Technology)
Talk title: Machine learning, brane tilings and toric Calabi-Yau 3-folds
Abstract:
Benjamin Suzzoni (Ulsan National Institute of Science and Technology)
Talk title: Conformal defects in neural network field theories
Abstract: Neural network field theory has become a fruitful formalism for constructing quantum field theories by trading the path integral with a better behaved statistical integral over neural network parameters. A couple of years ago, it was shown that this formalism can be used to engineer conformal field theories too. In this talk, I will present an extension of this formalism, valid in the presence of a defect, which we call Neural Network Defect Conformal Field Theory (nn-dCFT). I will outline the motivations for engineering such field theories and demonstrate how to construct them from neural networks. The talk will conclude with a couple of examples and further directions.
Norihiro Tanahashi (Kyoto University)
Talk title: Holography and optimal transport: Emergent Wasserstein spacetime from quantum states
Abstract: Optimal transport and the Wasserstein distance are prominent tools to quantify the space of probability distributions. Combining them with the manifold hypothesis in machine learning, we ask whether holography can be understood as a dimensional reduction of the space of quantum distributions. We find that the 1-Wasserstein distance between Husimi Q-representations makes the energy eigenstates of a harmonic oscillator collapse onto a one-dimensional emergent space, and that the Lindblad time evolution promotes it to a spacetime possessing an event horizon. For a Lindbladian subsystem of the SYK model, the same construction gives a trajectory consistent with the AdS$_2$ black hole, and in these examples the 1-Wasserstein distance is identified with a generalized Krylov complexity. These results suggest that the holographic principle may be positioned on a broader basis by using optimal transport.
Jiahua Tian (East China Normal University)
Talk title: Conformal fields from neural networks
Abstract: In this talk I will discuss a framework for constructing conformal fields from ensembles of neural networks. Using the embedding-space formalism, Lorentz-invariant homogeneous neural networks in (D+2) dimensions give rise to conformal fields in D dimensions, with correlation functions computed directly from integrals over network parameters. I will present examples with exactly calculable correlators and conformal block decompositions, explain the emergence of generalized free fields in the infinite-width limit, and discuss how deep networks generate sequences of conformal field theories.
Ashutosh Tripathi (Asia Pacific Center for Theoretical Physics)
Talk title: Deep learning holographic QCD: Gravity duals from quark-gluon plasma properties
Abstract: We introduce a data-driven approach to the holographic inverse problem for the quark-gluon plasma based on physics-informed neural networks. Whereas two-derivative Einstein gravity yields the universal prediction $\eta/s=1/4\pi$, Bayesian analyses of heavy-ion collision data indicate a nontrivial temperature ($T$) dependence of the shear-viscosity-to-entropy-density ratio ($\eta/s$). Motivated by this, we consider effective theories of gravity coupled to scalar matter and, more generally, an infinite tower of higher-curvature corrections. By training deep neural networks to solve the bulk field equations, we reconstruct the scalar potential and non-minimal coupling that reproduce lattice-QCD thermodynamics and a broad class of phenomenological $T$-dependent $\eta/s$ profiles, including those with a minimum near the phase transition, as well as the bulk viscosity profiles. As a first concrete realization, we truncate the higher-derivative sector to Einstein-Dilaton-Gauss-Bonnet gravity. Remarkably, the ultraviolet behavior of the reconstructed potential, specifically the sign of its effective mass squared, is determined by the temperature slope of $\eta/s$. This framework offers a systematic, data-driven alternative to ad hoc holographic model building, directly connecting heavy-ion phenomenology to its dual gravitational description.
Yuan Xin (Shanghai Institute for Mathematics and Interdisciplinary Sciences)
Talk title: Conformal bootstrap and vibe coding
Abstract: Numerical conformal bootstrap is one of the sharpest tools for studying CFTs. Despite major community efforts to simplify workflow and automate low-level tasks, its technical complexity is still a bottle-neck for many explorative projects. Working with AI agents is therefore a promising approach that will benefit both conformal bootstrap research and AI developments. In this talk, I will introduce an ongoing effort on solving some research-level conformal bootstrap problems with AI agents. The new workflow drastically simplifies the process of setting up and modifying bootstrap problems, while still requiring substantial human guidance, validation, and physical judgement. I will discuss what has worked, what has not, and several promising near-term applications of the new workflow for future.
Juntao Wang (Beijing Institute of Mathematical Sciences and Applications)
Talk title: Positive tensor-network Kähler metrics on gCICY threefolds
Abstract: In this talk, I will introduce a tensor-network approach to computing Ricci-flat Kähler metrics on Calabi-Yau manifolds. The construction gives a structured positive subfamily of algebraic Kähler metrics: a low-degree algebraic metric can be embedded exactly into a higher-degree tensor-network representation, while at fixed local and bond dimensions the number of trainable parameters grows only linearly with the algebraic degree. Under an immersion condition on the source sections, every parameter value defines a global positive Kähler metric in the same Kähler class. I will illustrate the method on three generalized complete-intersection Calabi-Yau (gCICY) threefolds. On the first example, we compare the tensor-network metric with a parameter-matched neural Kähler-potential correction using the same section data and a common starting metric. On the second, a degree-eight tensor network achieves lower Ricci-flatness errors than an unrestricted degree-four Hermitian baseline while using fewer parameters. Finally, I will discuss the extension to a type-gCICY. I will also examine how the accuracy depends on algebraic degree, bond dimension, and initialization, and show that increasing formal capacity does not necessarily lead to monotonic improvement.
Yang Zhang (University of Science and Technology of China)
Talk title: AI methods for the analytic computation of Feynman integrals
Abstract: In recent years, remarkable progress has been made in the analytic computation of Feynman integrals, for instance, on the cutting edge of two‑loop six‑point and three‑loop five‑point integrals. Nevertheless, several bottlenecks still persist in analytic computations, most notably the construction of a uniform transcendental integral basis and the derivation of the alphabet. In this talk, we will discuss novel AI‑based approaches to addressing these longstanding challenges.