TL;DR
This review explores how Machine Learning (ML) surrogate models are revolutionizing Computational Fluid Dynamics (CFD) for built environments. By drastically reducing simulation times, ML-powered CFD enables faster design optimization and real-time control for energy efficiency and occupant comfort, overcoming the computational bottlenecks of traditional methods.
What Problem Does This Paper Address?
Simulating airflow, thermal environments, and air quality in and around buildings using traditional CFD is computationally expensive. This high cost hinders rapid scenario analysis, design optimization, and real-time control, which are crucial for sustainable building design and operation. While traditional surrogate models exist, they often struggle with the complexity and high dimensionality of built environment flow fields. The paper addresses this by reviewing how ML-based surrogate models offer a promising solution to accelerate CFD simulations while maintaining acceptable accuracy.
Methods
The paper reviews various ML techniques applied to CFD in built environments. Key methods discussed include:
- Traditional Surrogate Models: Polynomial regression, response surface modeling, and Kriging models are briefly covered as a baseline.
- Machine Learning (ML) Surrogate Models: This is the core focus. ML models learn patterns from data to predict outcomes, bypassing computationally intensive solvers.
- Supervised Learning: Models trained on input features (e.g., geometric parameters) and corresponding outputs (e.g., flow field data).
- Unsupervised Learning: Discovering hidden structures in unlabeled data (e.g., for dimensionality reduction).
- Semi-supervised Learning: Combining small labeled datasets with large unlabeled ones.
- Neural Network Architectures: The review delves into specific ANN architectures:
- MLP (Multi-Layer Perceptron): A foundational feedforward network, though limited for complex spatial data.
- CNN (Convolutional Neural Network): Excels at extracting spatial features, useful for flow field patterns.
- RNN (Recurrent Neural Network) / LSTM (Long Short-Term Memory): Ideal for time-series data, capturing temporal dynamics of airflow.
- PINN (Physics-Informed Neural Network): Embeds physical governing equations (like Navier-Stokes) into the neural network's loss function, offering a balance between data-driven and physics-based approaches.
- Hybrid Approaches: Combining ROM (Reduced Order Models) with DSM (Direct Surrogate Models), and integrating ML with traditional CFD solvers or using ML to accelerate specific CFD stages (mesh preprocessing, numerical solving, post-processing).
Datasets
The paper discusses the generation and use of datasets for training ML models. These datasets are primarily derived from:
- High-fidelity CFD simulations: Generating large amounts of data for various scenarios, building geometries, and boundary conditions.
- Experimental measurements: Used for validation and sometimes as input data.
The paper highlights the challenge of acquiring high-quality, diverse, and physically representative datasets, especially for transient or turbulent flows and complex geometries. It notes that existing datasets are often static and may not capture temporal correlations effectively. No specific new benchmark datasets are introduced in this review.
Key Insights and Ideas
- ML as a CFD Accelerator: ML isn't just about prediction; it's fundamentally changing how CFD is performed by accelerating its core components: meshing, solving, and visualization.
- Bridging the Gap: ML surrogate models are crucial for bridging the gap between the need for high-fidelity CFD and the demand for rapid predictions in engineering design and operation.
- Beyond Pure Data-Driven: Physics-informed ML (like PINNs) is emerging as a powerful approach to overcome the limitations of purely data-driven models, especially regarding generalization and physical plausibility.
- Hybridization is Key: Combining ML with traditional CFD methods or using hybrid ML architectures (e.g., CNN-LSTM, POD-ANN) often yields superior results compared to standalone models.
- Temporal Dynamics are Crucial: For built environments, capturing time-varying phenomena (like wind gusts or changing occupancy) is vital, making RNNs and LSTMs particularly relevant.
Key Results
- Speedup: ML-based surrogate models can achieve tens to hundreds of times efficiency gains compared to traditional CFD methods, predicting flow fields in milliseconds to seconds.
- Accuracy: While speed is a major advantage, ML models can maintain fidelity comparable to traditional CFD in many scenarios. Specific examples mention accuracy rates of 88-96% and error rates significantly reduced compared to conventional methods.
- Mesh Generation: ML techniques have demonstrated an 85% reduction in mesh generation time and a 4- to 8-fold decrease in computational costs.
- Numerical Solution: ML-enhanced CFD solvers have shown up to an 80-fold acceleration in computation for 2D turbulent flow simulations.
- External Flow Prediction: DNNs performed optimally in predicting wind pressure, and hybrid GNN frameworks achieved computational efficiencies 280 times higher than traditional LES.
Limitations and Future Work
- Data Availability: Acquiring high-quality, diverse, and representative datasets remains a significant challenge, especially for complex, transient, or multi-physics scenarios.
- Multi-Physics Coupling: Most ML models are trained on single physical fields. Real-world built environments involve complex interactions (thermal, radiation, humidity, buoyancy, pollutant diffusion), which are difficult to model with current ML approaches.
- Generalization Capability: ML models can struggle to extrapolate to unseen geometries or boundary conditions due to their inductive nature and lack of explicit physical constraints.
- Temporal Consistency: Maintaining temporal consistency in transient simulations is essential but challenging for current ML models, particularly RNNs and TCNs.
- Trade-off: The inherent trade-off between prediction speed and high-fidelity accuracy remains a critical issue.
Future Work:
- Physical Information-Driven Techniques: Further development of techniques like active learning and physics-informed data augmentation to reduce computational costs and expand data coverage.
- Hybrid Models: Continued exploration of hybrid ML models (e.g., PINN-XGBoost, CNN-LSTM) to leverage the strengths of different approaches.
- Online Deployment: Research into strategies like migration learning for efficient online deployment of trained ML models.
- Interdisciplinary Collaboration: Emphasizing collaboration between ML researchers, CFD specialists, and built environment practitioners to translate theoretical advances into practical tools.
Summary updated