Generative models, reinforcement learning, and physics-informed learning.
01Generative models
PreprintUnder review
Scalable Continuous Conditional Diffusion via Random Projections and Prioritized Sampling
Jaewon Kim*, Hyo-Seok Hwang*, Junhee Seok
* Equal contribution.
Vicinal objectives in continuous conditional diffusion break down for multi-dimensional labels, where distances concentrate. HYPE-CCDM projects the vicinal objective onto random one-dimensional subspaces and adds distance-aware prioritized sampling, improving conditional fidelity on image generation, face generation, and inverse design.
Paper and code will be linked once available.
02Reinforcement learning
IEEE TNNLS2026
Fokker–Planck Soft Actor–Critic
Hyo-Seok Hwang, Jaewon Kim, Junhee Seok
FP-SAC derives a distribution-level policy objective from the Fokker–Planck equation of a soft policy improvement SDE. Minimizing its PDE residual with normalizing-flow policies yields stable multimodal policies on multigoal, MuJoCo, and Meta-World benchmarks.
NanoDiff: Physics-Guided Diffusion for Image-Based Nanophotonic Device Design
Insoo Kim, Jaewon Kim, Jangwon Seo, Junhee Seok†, Bohyung Han†
† Equal advising.
NanoDiff unifies electromagnetic field prediction and inverse design of nanophotonic devices. A conditional diffusion model generates devices for a target transmittance and peak wavelength, and physics-guided score sampling with a differentiable surrogate keeps designs physically consistent.