Graph-based Reinforcement Learning for Dynamic Photolithography Scheduling

WSC 2025, Seattle, USA · Collaborative work with Samsung Display

Authors: Sang-Hyun Cho, Sohyun Jeong, Jimin Park, Boyoon Choi, Paul Han, and Hyun-Jung Kim
Venue: Proceedings of the 2025 Winter Simulation Conference (WSC), Seattle, USA
Collaboration: Samsung Display

TL;DR

Photolithography is the tightest bottleneck in display and semiconductor production, and it is a dual-resource problem: a job needs both an eligible machine and the right mask (with transfer and setup times) at the same time. We propose a graph-based reinforcement learning framework that schedules photolithography operations in real time, explicitly tracking mask locations, machine statuses, and transfer times.

The problem

As OLED/TFT panels add more deposited layers, reentrant flows (jobs revisiting the same equipment) make photolithography scheduling increasingly complex. Jobs arrive dynamically with priorities, machine-eligibility constraints, and a required mask per machine. Prior RL approaches (e.g. deep Q-networks) must be retrained whenever mask counts or product types change, limiting them in highly dynamic lines.

The approach

  • Graph state — the scheduling state is a graph of job nodes and machine-mask pair nodes, connected by four edge types: job ↔ eligible machine-mask, jobs of the same lot type, machine-mask pairs on the same machine, and pairs sharing the same mask. Node/edge features encode priorities, remaining step targets, mask/machine availability, and transfer/setup times.
  • GNN policy — a GATv2 graph encoder feeds a policy that assigns probabilities to (job, machine–mask actions), trained end-to-end with policy gradients.
  • Size-agnostic — because the state is a graph, the same model handles varying numbers of machines, masks, and lot types without architectural changes or retraining, adapting to real-time changes on the floor.
  • Objective — a weighted combination of three KPIs: throughput, step-target fulfillment, and weighted priority.

Results

Across varied scenarios, the method delivers consistent and robust scheduling performance, making it a practical solution for real display-manufacturing lines.