I design inverse-designed silicon photonic devices and multi-dimensional optical interconnects for next-generation on-chip computing and AI acceleration — working across algorithm, device and system.
I am a direct-PhD candidate at Fudan University and currently a visiting PhD researcher at A*STAR Institute of Microelectronics (IME), Singapore. My research focuses on inverse-designed silicon photonic devices, multi-dimensional optical interconnects, and integrated photonic computing for AI acceleration.
My work covers the full design stack — algorithm-level inverse design, chip-scale fabrication and packaging, and system-level high-speed coherent transmission and on-chip optical computing — aimed at scaling on-chip bandwidth, dimensionality and compute density on silicon.
Photonic tensor computing scaled in parallelism by simultaneously multiplexing in time, space, frequency, mode and wavelength — unifying computation and interconnect in one architecture.
Non-volatile in-memory photonic computing on phase-change materials. Mode-insensitive PCM compute units enable hyper-multiplexed convolution; an in-situ training scheme breaks PCM write-cycle limits.
High-bandwidth silicon-photonic devices and multi-dimensional interconnect systems — process-robust inverse design, MDM/WDM components, high-speed microring modulators and on-chip high-speed link experiments.
800G silicon-photonic transceiver design and system packaging, with forward exploration for 1.6T — modulator doping optimization, high-density modulator/PD array integration, routing and multi-channel packaging.
An edge-guided inverse-design algorithm produces digital-metamaterial mode multiplexers that simultaneously achieve high capacity, broad bandwidth and fabrication robustness.
A photonic edge–metro architecture seamlessly stitches together distributed compute for generative-AI workloads, demonstrating end-to-end optical cloud inference.
A thin-film lithium-niobate modulator delivers electro-optic modulation continuously from near-infrared into the mid-infrared, opening a path toward mid-IR coherent links and sensing.
Ultra-broadband digital metamaterial multiplexers stitch multiple wavelength bands and spatial modes into a single chip — drastically lifting aggregate on-chip bandwidth.
An end-to-end deep-learning model jointly optimizes photonic-assisted fiber + millimeter-wave transmission for multiple users — encoder, channel and decoder learned together.
A compact dual-band wavelength demultiplexing splitter is obtained by inverse design, with a systematic study of how optimizer hyperparameters shape device performance.