Model order reduction of CFD spray field based on large language model

Computational Fluid Dynamics (CFD) captures complex physical processes with high precision, but its high computational cost hinders real-time prediction in digital twin applications. Conventional reduced-order models and regression methods are constrained b y fixed input–output dimensions, which restricts their flexibility in adapting to variable and dynamic mappings between operating conditions and flow field characteristics. Benefiting from advances in large language models (LLMs), this study proposes a fast prediction framework integrating Proper Orthogonal Decomposition (POD) and an LLM. POD is first employed to decompose high-dimensional spray fields into low-dimensional modes and coefficients to construct a reduced-order model (ROM). The mapping from operating parameters to modal coefficients is learned by an LLM via LoRA fine-tuning, with a weighted cross-entropy loss specially designed for numerical regression tasks. The proposed method outperforms Gaussian Process Regression (GPR), Decision Tree (DT), and Multi-Layer Perceptron (MLP) in prediction accuracy, spray morphology recovery, and gaseous penetration length prediction. The predicted spray fields are physically consistent with strong generalization ability in extrapolation cases, validating the high potential of Pretrained-Tuning frame of LLM on regression tasks in specialized domains, and provides a novel, efficient pathway for real-time modeling and fast reconstruction of complex physical fields in engineering digital twins.