Semester project - optical computing

SLM mask design with diffusion models for optical computing

Project Overview

When: 06/2025 – 09/2025
Duration: 3 months
Context: EPFL Master
Laboratory: Laboratory of Applied Photonic Devices (LAPD), EPFL
Stack & Skills: Pytorch Optical computing HPC SLURM Python Diffusion models Variational Auto Encoders AI Machine Learning

Conducted full-time at EPFL's Laboratory of Applied Photonic Devices (LAPD), this project develops an AI-driven pipeline for optical computing on Spatial Light Modulators (SLMs). A dual-decoder Variational Auto-Encoder jointly embeds SLM inputs and their optical outputs, and a latent diffusion model learns to generate new valid masks directly from data, bypassing slow physics simulations. Two optical setups were explored : a compact zigzag configuration on a gold-coated slab and a multi-plane SLM bench used for data acquisition. Several conditioning strategies were developed, witht the goal of driving the generator toward user-specified optical functions. The final proof of concept demonstrates that a neural network can design SLM masks whose predicted light patterns match the target, an early step toward real-time AI-assisted optical design.

Pipeline overview showing the VAE, diffusion model, and conditioning strategy. — End-to-end pipeline: VAE, latent diffusion model, and conditioning strategy.

Example of optical output generated by the AI pipeline using SLM masks. — Optical output generated by the trained pipeline from an SLM mask.