Ilmu Komputer & AI editorial
Open AccessOA2026
Large Language Models for Automating Computational Fluid Dynamics (CFD): From Predictive Modeling and Optimization to Execution Scheduling
A systematic review of LLM applications in CFD, covering predictive modeling, optimization, and automated execution scheduling, with key challenges and future directions.
Pei-Zhong Ma; G. Gai; Jian-kang Li; Zheng-Hong Luo; Li-Tao Zhuยท Smart Chemical Engineeringยท 2026ยท DOI 10.53941/sce.2026.100003
The core problem
Computational fluid dynamics (CFD) is a cornerstone of modern engineering and scientific research, enabling the simulation of fluid flows in diverse applications ranging from aerospace to chemical processing. However, traditional CFD workflows are computationally intensive and require significant human expertise for tasks such as mesh generation, solver configuration, and post-processing. Recent advances in artificial intelligence (AI), particularly large language models (LLMs), offer new opportunities to automate and enhance CFD processes. This review systematically examines the integration of LLMs into CFD, focusing on three key areas: predictive modeling, optimization, and automated execution scheduling. By leveraging the natural language understanding and generation capabilities of LLMs, researchers aim to reduce manual intervention, accelerate simulations, and improve accessibility. The authors also discuss the limitations of current approaches, such as insufficient physical credibility and engineering applicability, and propose future directions to address these challenges. The complementary relationship between human intelligence and AI (HI-AI) is emphasized as a pathway to
Innovation
The review reveals significant progress in applying LLMs to CFD across the three identified areas. In predictive modeling, LLMs have been used to generate real-time flow field distributions and to discover turbulence models from direct numerical simulation (DNS) data. For example, the k-ฮต model has been improved by implicitly encoding physical laws through LLM-based learning. In optimization, LLMs facilitate the tuning of numerical model parameters, hyperparameters of machine learning models (e.g., learning rate, optimizer, number of neurons), geometric structures (e.g., topological optimization of equipment), and process operating parameters (e.g., inlet velocity, pressure, temperature). Automated execution scheduling has enabled end-to-end automation, including automatic geometry generation, boundary condition configuration, meshing, and solver invocation. Despite these advances, the review notes persistent weaknesses: LLMs often lack physical credibility, and their engineering applicability is limited by data scarcity and interpretability issues. Performance metrics from surveyed studies indicate that while LLMs can reduce simulation time by up to 50% in certain tasks, accuracy
Computational fluid dynamics (CFD) is a cornerstone of modern engineering and scientific research, enabling the simulation of fluid flows in diverse applications ranging from aerospace to chemical processing. However, traditional CFD workflows are computationally intensive and require significant human expertise for tasks such as mesh generation, solver configuration, and post-processing. Recent advances in artificial intelligence (AI), particularly large language models (LLMs), offer new opportunities to automate and enhance CFD processes. This review systematically examines the integration of LLMs into CFD, focusing on three key areas: predictive modeling, optimization, and automated execution scheduling. By leveraging the natural language understanding and generation capabilities of LLMs, researchers aim to reduce manual intervention, accelerate simulations, and improve accessibility. The authors also discuss the limitations of current approaches, such as insufficient physical credibility and engineering applicability, and propose future directions to address these challenges. The complementary relationship between human intelligence and AI (HI-AI) is emphasized as a pathway to more effective CFD solutions.
The review adopts a systematic approach to survey recent literature on LLM applications in CFD. The authors categorize existing work into three primary domains: (1) predictive modeling, (2) optimization, and (3) automated execution scheduling. For each domain, they analyze the methods employed, the types of LLMs used (e.g., GPT, BERT, and domain-specific models), and the integration strategies with traditional CFD tools. The review also evaluates the performance of these approaches in terms of accuracy, efficiency, and generalizability. Key challenges are identified through a critical assessment of current studies, including issues related to data availability, model interpretability, and physical consistency. The authors synthesize findings to propose a roadmap for future research, emphasizing the need for hybrid approaches that combine LLMs with physics-based simulations. The methodology includes a detailed examination of case studies, such as the use of LLMs for turbulence model discovery and hyperparameter optimization, to illustrate practical implementations and outcomes.
Why it matters
The analysis highlights a complementary relationship between human intelligence and AI (HI-AI) as essential for advancing CFD. LLMs excel at pattern recognition and natural language tasks but struggle with enforcing physical laws and ensuring numerical stability. Therefore, hybrid frameworks that integrate LLMs with traditional solvers and physics-informed neural networks (PINNs) are promising. The authors suggest three future directions: (1) reducing the difficulty of obtaining domain data through transfer learning and synthetic data generation, (2) continuously updating domain models to incorporate new physical insights, and (3) improving model interpretability via explainable AI techniques. Additionally, the review underscores the need for standardized benchmarks to evaluate LLM performance in CFD tasks. The potential of LLMs to democratize CFD by lowering the barrier to entry for non-experts is acknowledged, but caution is advised regarding over-reliance on black-box models. The discussion concludes that while LLMs are not yet ready to fully replace human expertise, they can significantly augment CFD workflows when used appropriately.
Who should read this
CS practitioners and researchers
Opening member contentโฆ