The Sea Journal of the Korean Society of Oceanography ISSN:1226-2978(Print) 2671-8820(Online) 한국해양학회지 바다

In this study, a data-driven artificial intelligence (AI) model of the Lorenz system was constructed using a multi-layer perceptron (MLP) neural network, and a four-dimensional variational (4D-Var) data assimilation system was developed using this model. The AI model was trained to predict the state of the Lorenz system at a time interval ten times larger than the integration time interval of the original dynamical model, and it successfully reproduced chaotic characteristics of the Lorenz system, including the attractor structure, deterministic non-periodic flow, bifurcation phenomena, and sensitivity to initial conditions. The tangent linear and adjoint models of the AI model were derived using automatic differentiation, and the accuracy of the adjoint gradients was verified numerically. The AI 4D-Var model showed analysis accuracy comparable to that of the classical 4D-Var model. By combining a NumPy implementation based on extracted model weights (AI-NP) with network reduction, it achieved computational performance up to approximately 3.5 times faster than that of the classical 4D-Var model. With an appropriately sized network, a residual-learning method that adds a neural-network correction to the physics-based Euler forward-difference scheme also showed accuracy comparable to that of the classical 4D-Var model. In ensemble experiments comprising 64 cases, which combined eight initial conditions extracted from the Lorenz attractor with eight background-perturbation magnitudes, performance generally improved as network size increased. In particular, the AI-NP model with hidden-layer size h = 64 achieved a mean analysis error of 1.18 and a success rate of 90.6%, showing performance comparable to that of the classical 4D-Var model, which achieved a mean analysis error of 1.52 and a success rate of 87.5%. These results suggest that an AI-based 4D-Var model with sufficient network capacity can perform comparably to the classical 4D-Var model in a low-dimensional chaotic system. This study is a proof of concept in a low-dimensional chaotic system; future work should extend the validation to partial-observation settings and more complex systems.