The prediction of outcomes in university physics education is crucial for evaluating and improving the effectiveness of teaching methods. Traditional methods often face challenges in accurately predicting student performance, thus necessitating the exploration of new approaches. In this paper, we propose a novel application of machine learning in predicting the outcomes of university physics education. Specifically, we introduce a neural network based on particle swarm optimization (PSO) with the integration of attention mechanisms to enhance the prediction accuracy of the model. We first propose a multi-scale convolutional neural network (CNN) with attention mechanisms as the core model. The multi-scale convolution allows for the extraction of features at different scales, enhancing the richness and diversity of the captured features. The integration of attention mechanisms further improves the feature selection process, enabling the network to focus on the most informative aspects of the input data. Additionally, we address the issue of random initialization in neural networks by introducing an improved particle swarm optimization algorithm. Traditional PSO algorithms may suffer from premature convergence or slow exploration of parameter space, leading to suboptimal solutions. Therefore, we propose an improved PSO algorithm aimed at enhancing the performance of the neural network.