Thermal and airflow surrogate model for Datacenter design
This example demonstrates the use of a Deep Learning model (3D UNet) for training a surrogate model for datacenter airflow to enable real-time datacenter design. The aim of this workflow is to train a Deep Learning model that can predict the temperature and airflow distribution within a hot aisle of a typical datacenter. For any given geometry of the hot aisle (height, width, and length) and the number of IT racks inside it, the trained model can predict the temperature, velocity, and pressure distribution inside the hot aisle instantaneously. Such an approach can be very useful from a datacenter design perspective where iterating through various design combinations is crucial to obtain optimal cooling and minimize hotspots.
Design study using the AI surrogate model
The model is trained on OpenFOAM simulation data. Based on the variables, i.e., Length, Height, Width, and Number of Racks, several hot aisle configurations are generated. These configurations are solved with OpenFOAM assuming maximum flow rate and rack exit temperature (max load condition). Steady state simulations are used and the resulting OpenFOAM data is exported in VTK format for training of the AI surrogate. The dataset is then normalized using the mean and standard deviation statistics of the dataset. The normalized dataset, along with a sample OpenFOAM configuration, can be downloaded using the NGC link: [Link to be added].
Note: Access to NVAIE is required to download the dataset and the reference OpenFOAM configuration.
Note: The OpenFOAM configuration provided is only representative. Several key aspects have been masked to protect the IP. Users should not expect to generate the training data exactly using this setup, and one will have to change it using their own geometries and boundary conditions.
A UNet model is used in this problem. The hex-dominant mesh used in this problem makes this model an attractive choice offering good speed and accuracy. Since the model is primarily trained to capture the changes in geometry, we use the Signed Distance Field of the interior of the hot aisle to capture the parameter variation. Additionally, we add sinusoidal embeddings to enable the model to capture sharp features in the flow field. Finally, to make the different datacenter sizes uniform (for ingestion into UNet), we pad the geometry for the maximum hot aisle dimensions. This padding is removed before computing the loss.
The model can be trained by executing the below commands:
python train.py
To train on multiple GPUs,
mpirun -np <#GPUs> python train.py
Once the model is trained, you can use the inference.py script to compute the model inference. For generating the Signed Distance Field and geometry for the inference, we make use of the utilities from PhysicsNeMo-Sym.
Training of Physics-Informed model
We also train a variant where we add the physics losses to the data loss. The physics-informed training can be executed using the below commands:
python train_physics_informed.py
Addition of such physics data losses proves very beneficial in the low-data regime where the physics losses can compensate for the lack of enough data.
Comparison of data+physics driven training with pure data driven training
This example was developed as a part of collaboration between NVIDIA and Wistron.