Selected Work

PaNO: Readout-Aligned Simulation for Photonic Devices

#AI for Photonics #Neural Operator #MMI #Port Readout

Starting from the mismatch between a field surrogate's global error and its port readouts, organizing a neural operator along the propagation axis so the local quantities a designer actually cares about sit at the center of evaluation and design.

S-Crescendo: Simulating Higher-Order Nonlinear Circuits

#AI for EDA #Transformer #S-Domain #Nonlinear Systems

Building training data from first- to third-order RC networks, then letting the model follow modal structure in the S-domain to reach ninth-order circuits it never saw in training.

Physics-Constrained CNN and Bayesian Inference for Beam-Parameter Inversion

#AI for Optics #Bayesian Inference #CNN #Uncertainty Quantification

Pairing a neural network with a physical equation and a probability distribution, turning laser beam-parameter measurement from black-box fitting into a checkable measurement system.

Diffusion Reproduction and One-Step Distillation on CIFAR-10

#Diffusion Models #Distillation #Generative Modeling #CIFAR-10

Reproducing DDPM/EDM from scratch, then distilling a 1000-step diffusion model into a single-step generator, with an ablation that pins down why the teacher's quality is the hard ceiling.

SIM1: Reproducing a Physics-Aligned Data Scaler for Deformable Worlds

#Physics AI #Deformable Manipulation #World Models #Simulation

Reconstructing SIM1's end-to-end loop from paper to simulation replay to data export, and finding that physically runnable rollouts are still far from task-successful ones.