As a mesoscopic approach, the lattice Boltzmann method has achieved considerable success in simulating fluid flows and associated transport phenomena. The calculation, however, suffers from a massive amount of computing resources. A predictive model, to reduce the computing cost and accelerate the calculations, is proposed in this work. By employing an artificial neural network, composed of convolution layers and convolution long short-term memory layers, the model is an equivalent substitution of multiple time steps. A physical informed training loss function is introduced to improve the model predictive accuracy; and for the two-dimensional driven cavity problem, the mean square error of the prediction is less than $1.5\times {10}^{-6}$. For non-stationary flow, a time-dependent computing structure based on the current model is established. Nine iterative model calculations are performed consecutively for a two-dimensional driven cavity model, and the results are validated by comparing with the original (serial) lattice Boltzmann algorithm. Generally, in the case of training Reynolds number, for velocity and speed, the mean and the maximum absolute errors are lower than 0.012 and 0.12. Similarly, in the generalizing case, the mean and the maximum absolute errors are lower than 0.017 and 0.012. Besides, the current model's efficiency is about 15 times higher than that of the original lattice Boltzmann method.

作为一种细观方法，格子玻尔兹曼方法在模拟流体流动和相关的传输现象方面取得了相当大的成功。然而，计算受到大量计算资源的影响。在这项工作中提出了一种预测模型，以降低计算成本并加速计算。通过使用由卷积层和卷积长短期记忆层组成的人工神经网络，该模型是多个时间步长的等效替代。引入了物理通知训练损失函数以提高模型预测精度；对于二维驱动腔问题，预测的均方误差小于$1.5\times {10}^{-6}$. 对于非平稳流，建立基于当前模型的时间相关计算结构。对二维驱动腔模型连续进行了9次迭代模型计算，并通过与原始（串行）格子Boltzmann算法进行比较来验证结果。一般在训练雷诺数的情况下，对于速度和速度，均值和最大绝对误差低于0.012和0.12。同样，在泛化情况下，均值和最大绝对误差低于 0.017 和 0.012。此外，当前模型的效率比原始格子 Boltzmann 方法的效率高约 15 倍。