Domino multiscale geometry features capability and inference script

CHANGELOG.md

@@ -59,6 +59,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   cross entropy loss.
 - Option to offload checkpoints to further reduce memory usage
 - Added StormCast model training and simple inference to examples
+- Multi-scale geometry features for DoMINO model.
 
 ### Changed
 

examples/cfd/external_aerodynamics/domino/README.md

@@ -20,6 +20,9 @@ by sampling random neighboring points within the same sub-region. The local-geom
 encoding and the computational stencil are aggregrated to predict the solutions on the
 discrete points.
 
+A preprint describing additional details about the model architecture can be found here
+[paper](https://arxiv.org/abs/2501.13350).
+
 ## Dataset
 
 In this example, the DoMINO model is trained using DrivAerML dataset from the
@@ -69,17 +72,36 @@ To enable retraining the DoMINO model from a pre-trained checkpoint, follow the
 5. Download the validation results (saved in form of point clouds in `.vtp` / `.vtu` format),
    and visualize in Paraview.
 
+## DoMINO model inference on STLs
+
+The DoMINO model can be evaluated directly on unknown STLs using the pre-trained
+ checkpoint. Follow the steps outlined below:
+
+1. Run the `inference_on_stl.py` script to perform inference on an STL.
+
+2. Specify the STL paths, velocity inlets, stencil size and model checkpoint
+ path in the script.
+
+3. The volume predictions are carried out on points sampled in a bounding box around STL.
+
+4. The surface predictions are carried out on the STL surface. The drag and lift
+ accuracy will depend on the resolution of the STL.
+
 ## Guidelines for training DoMINO model
 
 1. The DoMINO model allows for training both volume and surface fields using a single model
  but currently the recommendation is to train the volume and surface models separately. This
   can be controlled through the config file.
+
 2. MSE loss for both volume and surface model gives the best results.
+
 3. The surface and volume variable names can change but currently the code only
  supports the variables in that specific order. For example, Pressure, wall-shear
   and turb-visc for surface and velocity, pressure and turb-visc for volume.
+
 4. Bounding box is configurable and will depend on the usecase. The presets are
  suitable for the AWS DriveAer-ML dataset.
+
 5. Integral loss factor is currently set to 0.0 as it adversely impacts the training.
 
 The DoMINO model architecture is used to support the Real Time Wind Tunnel OV Blueprint

examples/cfd/external_aerodynamics/domino/src/conf/config.yaml

@@ -26,6 +26,15 @@ hydra: # Hydra config
     dir: ${output}
   output_subdir: hydra  # Default is .hydra which causes files not being uploaded in W&B.
 
+# The directory to search for checkpoints to continue training.
+resume_dir: ${output}/models
+
+data_processor: # Data processor configurable parameters
+  kind: drivaer_aws # must be either drivesim or drivaer_aws
+  output_dir: /lustre/rranade/modulus_dev/data/volume_data/
+  input_dir: /lustre/datasets/drivaer_aws/drivaer_data_full/
+  num_processors: 12
+
 data: # Input directory for training and validation data
   input_dir: /lustre/rranade/modulus_dev/data/volume_data/
   input_dir_val: /lustre/rranade/modulus_dev/data/volume_data_val/
@@ -36,9 +45,6 @@ data: # Input directory for training and validation data
     min: [-1.1, -1.2 , -0.32]
     max: [4.5 , 1.2  , 1.2]
 
-# The directory to search for checkpoints to continue training.
-resume_dir: ${output}/models
-
 variables:
   surface:
     solution:
@@ -55,40 +61,45 @@ variables:
 model:
   model_type: combined # train which model? surface, volume, combined
   loss_function: "mse" # mse or rmse
-  interp_res: [128, 64, 48] # resolution of latent space
+  interp_res: [128, 128, 128] # resolution of latent space 128, 64, 48
   use_sdf_in_basis_func: true # SDF in basis function network
   positional_encoding: false # calculate positional encoding?
   volume_points_sample: 8192 # Number of points to sample in volume per epoch
   surface_points_sample: 8192 # Number of points to sample on surface per epoch
-  geom_points_sample: 200_000 # Number of points to sample on STL per epoch
+  geom_points_sample: 300_000 # Number of points to sample on STL per epoch
   surface_neighbors: true # Pre-compute surface neighborhood from input data
   num_surface_neighbors: 7 # How many neighbors?
   use_surface_normals: true # Use surface normals and surface areas for surface computation?
-  use_only_normals: true # Use only surface normals and not surface area
-  integral_loss_scaling_factor: 0 # Scale integral loss by this factor
+  use_surface_area: true # Use only surface normals and not surface area
+  integral_loss_scaling_factor: 100 # Scale integral loss by this factor
   normalization: min_max_scaling # or mean_std_scaling
-  encode_parameters: true # encode inlet velocity and air density in the model
+  encode_parameters: false # encode inlet velocity and air density in the model
+  surf_loss_scaling: 5.0 # scale surface loss with this factor in combined mode
+  vol_loss_scaling: 1.0 # scale volume loss with this factor in combined mode
   geometry_rep: # Hyperparameters for geometry representation network
-    base_filters: 16
     geo_conv:
       base_neurons: 32 # 256 or 64
       base_neurons_out: 1
-      radius_short: 0.1
-      radius_long: 0.5 # 1.0, 1.5
+      volume_radii: [0.1, 0.5]
+      surface_radii: [0.05] # 0.05
       hops: 1
     geo_processor:
       base_filters: 8
     geo_processor_sdf:
       base_filters: 8
   nn_basis_functions: # Hyperparameters for basis function network
     base_layer: 512
+    fourier_features: false
+    num_modes: 5
   aggregation_model: # Hyperparameters for aggregation network
     base_layer: 512
   position_encoder: # Hyperparameters for position encoding network
     base_neurons: 512
   geometry_local: # Hyperparameters for local geometry extraction
-    neighbors_in_radius: 64
-    radius: 0.05 # 0.2 in expt 7
+    volume_neighbors_in_radius: [128, 128] # [64, 128]
+    surface_neighbors_in_radius: [128] # [64]
+    volume_radii: [0.05, 0.1] # [0.05. 0.1]
+    surface_radii: [0.05] # [0.05]
     base_layer: 512
   parameter_model:
     base_layer: 512
@@ -117,9 +128,5 @@ eval: # Testing configurable parameters
   test_path: /lustre/rranade/benchmarking/drivaer_aws_surface_test_new/
   save_path: /lustre/rranade/domino/mesh_predictions_surf_final1/
   checkpoint_name: DoMINO.0.50.pt
-
-data_processor: # Data processor configurable parameters
-  kind: drivaer_aws # must be either drivesim or drivaer_aws
-  output_dir: /lustre/rranade/modulus_dev/data/volume_data/
-  input_dir: /lustre/datasets/drivaer_aws/drivaer_data_full/
-  num_processors: 12
+  scaling_param_path: /lustre/rranade/modulus_dev/modulus_forked/modulus/examples/cfd/external_aerodynamics/domino/outputs/AWS_Dataset/
+  refine_stl: False # Automatically refine STL during inference