Wafer　defect　inspection　is　one　of　the　crucial　semiconductor　processing　technologies　because　it　can　help　to　identify　the　surface　defects　in　the　process　and　eventually　improve　the　yield.　Manual　inspection　using　human　eye　is　subjective　and　long-term　fatigue　can　lead　to　erroneous　classification.　Deep　learning　technology　such　as　convolutional　neural　network　(CNN)　is　a　promising　way　to　achieve　automated　wafer　defect　classification.　The　training　of　CNN　is　time　consuming　and　it　is　nontrivial　to　fine　tune　its　hyperparameters　to　achieve　good　classification　performance.　In　this　study,　Arithmetic　Optimization　Algorithm　(AOA)　is　proposed　to　optimize　the　CNN　hyperparameters,　such　as　momentum,　initial　learn　rate,　maximum　epochs,　L2　regularization,　to　reduce　the　burden　brought　by　trial-and-error　methods.　The　hyperparameters　of　a　well-known　pretrained　model,　i.e.,　GoogleNet,　are　optimized　using　AOA　to　perform　wafer　defects　classification　task.　Simulation　studies　report　that　the　AOA-optimized　GoogleNet　achieves　promising　accuracy　of　91.32%　in　classifying　wafer　defects.　©　The　2023　International　Conference　on　Artificial　Life　and　Robotics　(ICAROB2023).