When performing evolutionary optimization for computationally expensive objective, surrogate-assisted evolutionary algorithm(SAEA) is an effective approach. However, due to the limited availability of data in these scenarios, it can be challenging to create a highly accurate surrogate model, leading to reduced optimization effectiveness. To address this issue, we propose an Interpretable Convolution Network(ICN) for offline surrogate-assited evolutionary optimization. ICN retains the non-linear expression ability of traditional neural networks, while possessing the advantages of clear physical structure and the ability to incorporate prior knowledge during network parameter design and training process. We compare ICN-SAEA with tri-training method(TT-DDEA) and model-ensemble method(DDEA-SA) in several benchmark problems. Experimental results show that ICN-SAEA is better in searching optimal solution than compared algorithms.