======================================
5/22 ( Room 8101 at KIAS )
======================================
11:00-12:00 Sejin Kim (KIAS)
13:00-14:00 Benjamin Suzzoni (UNIST)
14:00-15:00 Jeongsup Moon (CJ)
15:00-16:00 Rak-Kyeong Seong (UNIST)
16:00-17:00 Ashutosh Tripathi (APCTP)
17:00-18:40 Yeongwoo Song (KAIST)
======================================
5/23 ( room 519, 5th floor 공학관 at Kookmin University )
======================================
10:00 Excursion to Gyeongbokgung Palace
(Outing where you can freely participate)
13:30-14:30 Talk and Round Table Meeting I by Yunseok Seo
14:30-15:30 Talk and Round Table Meeting II by Kyung Kiu Km
======================================
Titles and Abstracts
======================================
Ashutosh Tripathi (APCTP)
Deep learning holographic QCD: gravity duals from transport properties of the quark-gluon plasma
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 dependence of the shear-viscosity-to-entropy-density ratio. 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 neural networks to solve the bulk field equations, we reconstruct the scalar potential and nonminimal couplings that reproduce lattice-QCD thermodynamics and a broad class of phenomenological $\eta/s(T)$ profiles, including those with a minimum near the phase transition. 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 replaces ad hoc holographic model building with a systematic, high-precision procedure that directly connects heavy-ion phenomenology to its dual gravitational description.
Benjamin Suzzoni (UNIST)
Neural-Network Field Theory Correspondance for Wess-Zumino-Witten Theories
Neural-Network Field Theories (NNFTs) have emerged as a powerful and practical tool for constructing free field theories using a family of infinitely large neural networks. It was recently shown that these can also be used to recreate the free boson and free fermion in two dimensions, together with their Virasoro symmetry. In this talk, I will start by presenting an overview of what NNFTs are and how they are constructed. We will spend some time discussing how to reconstruct 2d CFTs, and how their description in terms of NNFTs differs greatly from their higher-dimensional cousins. We will wrap up with a brief presentation of how this toolset can be used to construct 2d CFTs on group manifolds, i.e. Wess-Zumino-Witten theories (paper to appear soon).
Sejin Kim (KIAS)
Learning the Inverse Ryu–Takayanagi Formula with Transformers
We train a Transformer to invert the Ryu–Takayanagi formula: given a boundary entanglement-entropy curve, the network outputs the bulk blackening function f(z) that produced it. The training set — one hundred thousand bulk–boundary pairs — is manufactured entirely by holography itself, by adding Wiener noise to BTZ-like profiles and applying the Ryu–Takayanagi map. After a single round of training, the model reconstructs unseen smooth black holes, faithfully interpolates a continuous family of perturbed geometries (MSE ~10⁻⁷), and even recovers horizonless geometries outside the trained distribution. We discuss what this implies for combining physics-informed (PINN) and data-driven generative methods in holography, and sketch natural extensions to higher dimensions and time-dependent entropies.
Jeongsup Moon (CJ)
Deep Learning Approaches to Simulation in the Biological and Physical Sciences
For half a century, scientific modeling rested on a simple recipe: write the equation, discretize it, throw computers at it. That recipe still works in principle. In practice, it has been quietly displaced in field after field by models that never integrate the equations at all. An examination of this shift across biology and several areas of physics reveals a common thread.
Structural biology offers the clearest opening case. AlphaFold 2, ESMFold, and AlphaFold 3 read three-dimensional shape and binding directly from sequence, without any explicit physical force ever being computed. This same shift appears in different costumes across the physical sciences: neural wavefunctions that out-resolve coupled-cluster on small molecules; learned weather emulators (GraphCast, GenCast, Aurora) that beat ECMWF's operational solver at a fraction of the cost; generative pipelines for inorganic crystals (GNoME, MatterGen, MACE-MP-0); and closed-loop systems such as AlphaQubit decoding surface codes on Sycamore, DINGO inferring gravitational-wave parameters in under a minute, and a reinforcement-learning policy steering plasma on the TCV tokamak.
The pattern is not that data has replaced theory. Rather, learned models can absorb the parts of a problem that were always too costly to solve from first principles, and turn that effort into something reusable. Yet, critical breaking points remain, distribution shift, conservation drift, and the gap between predicted-stable and synthesizable matter, leaving the boundary of when to reach for a learned model open to ongoing exploration.
Yeongwoo Song (KAIST)
How Neural Networks Learn Physics: From Explicit Priors to Emergent Representations
Neural networks can often predict the dynamics of a physical system with impressive accuracy, but accurate prediction alone does not necessarily imply that a model has learned the underlying physical law. This talk asks where, and in what form, physical principles are represented inside neural networks. I will discuss this question through two complementary approaches in AI for physics. The first approach explicitly incorporates physical priors. This method combines Hamiltonian structure, and meta-learning to acquire representations that can rapidly adapt across distinct dynamical systems. This suggests that neural networks can learn reusable representations aligned with the shared structure of Hamiltonian mechanics, rather than merely fitting system-specific trajectories. The second approach asks whether physics representations can emerge without explicitly imposing physical laws. We demonstrate that large language models can forecast the dynamics of a given system, purely from the given prompt, without any parameter updates. More importantly, probing the internal residual stream activations of the language model reveals internal features correlated with quantities related to physics concepts. Together, these studies frame a broader perspective on how neural networks learn physics: from models guided by explicit physical priors to general models that appear to form emergent internal representations of physical concepts. I will discuss that the next shift in physics-informed AI is not only to improve domain specific predictions, but also to identify, validate, and control the physical representations that make such tasks possible.
Rak-Kyeong Seong (UNIST)
Birational Transformations on Dimer Integrable Systems
We show that when two toric Calabi-Yau 3-folds and their corresponding toric varieties are related by a birational transformation, they are associated with a pair of dimer models on the 2-torus that define dimer integrable systems, which themselves become birationally equivalent. These integrable systems defined by dimer models were first introduced by Goncharov and Kenyon. We illustrate this equivalence explicitly using a pair of dimer integrable systems corresponding to the abelian orbifolds of the form C^3/Z_4 x Z_2 with orbifold action (1,0,3)(0,1,1) and C/Z_2 x Z_2 with action (1,0,0,1)(0,1,1,0), whose spectral curves and Hamiltonians are shown to be related by a birational transformation.
Kyung Kiu Kim (Kookmin University)
Defect Mass Deformation of ABJM Theory
We study particular gravity duals to the inhomogeneously mass-deformed ABJM (ImABJM) theory. The final goal of this paper is to obtain the ABJM theory deformed by a delta function-type mass. The resultant theory can be interpreted as the ABJM theory with defects. The mass operators of the mass-deformed ABJM (mABJM) theory are identified with these defect operators. Corresponding gravity duals are supergravity solutions obtained by solving a relevant BPS equation. This BPS equation is a linear partial differential equation known as the Helmholtz equation. We solve the differential equation using various numerical methods, including machine learning. We investigate appropriate physical quantities in the supergravity background.
Yunseok Seo (Kookmin University)
Gap opening in WSM by D-instantons
We investigate D-insatnton effects on the holographic Weyl semimetal in top-down approach. From the free energy of the D7 brane embedding solutions, we get phase diagram in terms of the electron mass, instanton number, and temperature in the unit of the weyl parameter. We calculate non-linear conductivities from the regularity condition of the probe D7 brane and investgate anomalous Hall phenomena in the boundary system. From the study of the phase diagram, we suggest the gaped phase induced by the instanton to a topological insulator.


