Efficient Routing-based Synthesis for Digital Microfluidic Biochips via Reinforcement Learning

The use of digital microfluidic biochips (DMFBs) has highlighted their superiority in automatically executing biochemical assays by controlling tiny nano/picoliter droplets, which are moved in parallel to enhance throughput. Routing-based synthesis for DMFBs yields faster assay execution times compared to module-based synthesis when on-chip resource constraints are stringent. However, without predefining modules, it is very challenging to handle all the droplets directly on the chip for successfully executing the desired biochemical assay, especially in dynamic environments. Through modeling routing-based synthesis into two kinds of real-time decision tasks, i.e., transportation and mixing, this paper proposes a new routing-based synthesis framework that uses deep reinforcement learning (DRL) to train transportation and mixing agents respectively. Additionally, we design effective partial observations and curriculum learning (CL) schemes for both kinds of agents to improve their generalization ability and accelerate the training process. Compared to the state-of-the-art heuristic routing-based synthesis methods, more efficient synthesis processes of the given assays can be achieved using the proposed method of combining DRL and CL. For example, the average completion time on several real-world bioassay benchmarks (PCR, INVITRO, and PROTEIN) was reduced by 12.9% 18.5% approximately.