计算机科学
强化学习
可扩展性
布线(电子设计自动化)
微流控
比例(比率)
工程类
理论(学习稳定性)
噪音(视频)
控制工程
实时计算
钥匙(锁)
模拟
分布式计算
钢筋
作者
Hongchuan Guo,Tingjing Zhang,Zhengkun Yi,Hailong Yao,Tsung-Yi Ho,Bo Yuan
标识
DOI:10.1109/tcad.2026.3720026
摘要
By automatically controlling discrete droplets, digital microfluidic systems have demonstrated significant advantages in biochemical assays, such as viral testing and drug screening. A primary challenge in large-scale digital microfluidic systems lies in performing real-time droplet routing for dozens or even hundreds of droplets, using real-time environmental feedback, while ensuring that all droplets reach their destinations as quickly as possible without violating both static and dynamic fluidic constraints. To address this challenge, this paper proposes an imitation learning (IL) - reinforcement learning (RL) - curriculum learning (CL) -based framework for large-scale adaptive droplet routing in digital microfluidic systems. Specifically, by performing behavior cloning on an exact solver for small-scale droplet routing problems, an initial policy is pre-trained, which is then refined through a multistage RL and CL process, enabling the trained policy to handle large-scale droplet routing problems. As an adaptation of our IL–RL–CL framework, we combine Independent Proximal Policy Optimization (IPPO) with a transformer-based communication mechanism for information sharing and coordination. Compared to the state-of-the-art RL method, the superior performance of the proposed method is demonstrated across varying numbers of droplets, i.e., the improvement on success rate is 8%∼29%, and the average completion time is reduced by 14%∼27%. Besides, we show that our method achieves a high success rate when routing one hundred droplets, whereas the previous method only works with up to ten droplets. Moreover, the well-trained policy can be directly applied to any routing tasks with a varying number of droplets.
科研通智能强力驱动
Strongly Powered by AbleSci AI