Abstract Predicting ocean transport has many practical applications ranging from search and rescue operations to predicting the spread of oil, debris, and biogeochemical tracers, yet trajectory prediction remains a challenge for existing ocean modeling techniques. General circulation models require high resolution observational data in order to be properly initialized, but these data do not exist for the ocean. Statistical models are difficult to tune with existing data and are often too simple to accurately encapsulate turbulent flows. Here we investigate a data-driven approach to ocean transport prediction wherein the goal is to first learn from available data instead of prescribed laws of physics and then apply this information to new data. More specifically, we explore whether simple artificial neural networks (ANNs) are capable of learning to predict 2D particle trajectories using only previous velocity observations. ANNs are trained in two ways: first, a so-called “one-to-one ANN” uses a particle’s most recently observed velocity to predict its velocity six hours later, and second, a “time series ANN” uses the past 24 hours’ worth of velocity observations to predict the next 24 h. We present a proof-of-concept considering particles in a hierarchy of simulated flow regimes ranging from uniform, steady flow to more complex scenarios with interacting scales of motion and then substantiate our approach on trajectories in modeled flows generated by a high-resolution Hybrid Coordinate Ocean Model for a mesoscale eddy in the northern Gulf of Mexico. We also assess ANN sensitivity to the prediction window over which forecasts are made, the number of training particles used, and the size of the network. ANNs successfully predict 24 h trajectories within the temporal bounds of the training data with forecast errors around half those of both rudimentary persistence and classical ARIMA models. Predicting beyond the domain of the training data leads to forecast errors comparable to ARIMA models. Our results suggest that ANNs offer promising potential as a data-driven approach to forecasting material transport in the ocean.

中文翻译：

---

使用人工神经网络预测海洋流动中的粒子轨迹

摘要 预测海洋运输有许多实际应用，从搜索和救援行动到预测石油、碎片和生物地球化学示踪剂的扩散，但轨迹预测仍然是现有海洋建模技术的挑战。一般环流模型需要高分辨率观测数据才能正确初始化，但这些数据并不存在于海洋中。统计模型很难用现有数据进行调整，而且通常太简单而无法准确地封装湍流。在这里，我们研究了一种数据驱动的海洋运输预测方法，其目标是首先从可用数据而不是规定的物理定律中学习，然后将这些信息应用于新数据。进一步来说，我们探讨了简单的人工神经网络 (ANN) 是否能够仅使用先前的速度观察来学习预测 2D 粒子轨迹。人工神经网络通过两种方式进行训练：首先，所谓的“一对一人工神经网络”使用粒子最近观察到的速度来预测六小时后的速度，其次，“时间序列人工神经网络”使用过去 24 小时' 价值的速度观测来预测接下来的 24 小时。我们提出了一个概念验证，考虑到模拟流态层次结构中的粒子，从均匀、稳定的流动到具有相互作用的运动尺度的更复杂的场景，然后证实我们在由高分辨率混合坐标生成的建模流中的轨迹的方法墨西哥湾北部中尺度涡旋的海洋模型。我们还评估了 ANN 对进行预测的预测窗口、使用的训练粒子数量和网络大小的敏感性。人工神经网络成功地预测了训练数据时间范围内的 24 小时轨迹，预测误差约为基本持久性和经典 ARIMA 模型的一半。超出训练数据域的预测会导致与 ARIMA 模型相当的预测误差。我们的研究结果表明，人工神经网络作为一种数据驱动的方法来预测海洋中的物质运输具有广阔的潜力。人工神经网络成功地预测了训练数据时间范围内的 24 小时轨迹，预测误差约为基本持久性和经典 ARIMA 模型的一半。超出训练数据域的预测会导致与 ARIMA 模型相当的预测误差。我们的研究结果表明，人工神经网络作为一种数据驱动的方法来预测海洋中的物质运输具有广阔的潜力。人工神经网络成功地预测了训练数据时间范围内的 24 小时轨迹，预测误差约为基本持久性和经典 ARIMA 模型的一半。超出训练数据域的预测会导致与 ARIMA 模型相当的预测误差。我们的研究结果表明，人工神经网络作为一种数据驱动的方法来预测海洋中的物质运输具有广阔的潜力。