JoWUA

Abstract

It has been suggested in the present paper that a Quantum-Inspired Reinforcement Learning (QIRL) framework can be proposed to adaptive semiconductor wafer probing in intricate multi-die settings, where traditional deterministic policies and classical reinforcement learning frameworks have difficulties in combinatorial explosion, sparse reward gradients, and uneven distributions of defects. The proposed approach introduces an amplitude-based superposition mechanism for probabilistic policy encoding, enabling the simultaneous exploration of multiple probing trajectories and improving decision-making efficiency under uncertainty. The environment of the high-fidelity digital twin simulation was created based on the realistic conditions of the wafer, which included the heterogeneous die layout, stochastic defect patterns (random, clustered and edge-based), and probe degradation dynamics. QIRL framework was strictly tested with comparison to the baseline procedures that were deterministic raster scanning, tabular Q-learning, Deep Q-Network (DQN), and Proximal Policy Optimization (PPO). The effectiveness of the proposed model is shown to be reduced to 28 % less probing moves than raster scanning and 17 % less than classical Q-learning through experimentation. Also, the switching costs are minimized up to 22% compared to PPO, which means an increase in the efficiency of path optimization. QIRL has a defect detection accuracy of 94/97 on average with a false positive rate of less than 3 meaning that in clustered defect situations, QIRL is able to find defects with a high accuracy of up to 97. Moreover, the framework also decreases the wear of probes by about 20 %, which leads to the increased lifespan of the equipment and reliability of the testing. All in all, the presented QIRL model is an excellent balance between the probing efficiency, accuracy of defect localization, and hardware durability. The findings indicate that it is scalable and can be used in next-generation semiconductor testing, and its potential is high when it comes to implementation in hybrid quantum-classical computing systems.