Model-Based OPC With Adaptive PID Control Through Reinforcement Learning

Model-based optical proximity correction (MB- OPC) relies on a feedback loop, in which correction result, measured as edge placement error (EPE), is used for decision of next correction. A proportional-integral-derivative (PID) control is a popular mechanism employed for such feedback loop, but current MB-OPC usually relies only on P control. This is because there is no systematic way to customize P, I, and D coefficients for different layouts in different OPC iterations.We apply reinforcement learning (RL) to construct the trained actor that adaptively yields PID coefficients within the correction loop. The RL model consists of an actor and a critic. We perform supervised pre-training to quickly set the initial weights of RL model, with the actor mimicking standard MB-OPC. Subsequently, the critic is trained to predict accurate Q-value, the cumulative reward from OPC correction. The actor is then trained to maximize this Q-value. Experiments are performed with aggressive target maximum EPE values. The proposed OPC for test layouts requires 5 to 7 iterations, while standard MB-OPC (with constant coefficient-based control) completes in 20 to 28 iterations. This reduces OPC runtime to about 1/2.7 on average. In addition, maximum EPE is also reduced by about 24%.