Add trained NN models and clean up legacy weights

.github/scripts/compare_cpu_gpu_builds.sh

@@ -25,23 +25,55 @@ echo "=== Building (CPU-only reference) ==="
 make -j8
 
 echo ""
-echo "--- Running test_cpu_gpu_consistency (CPU-only build) ---"
-echo "This test will skip GPU-specific tests and run CPU validation only."
-./test_cpu_gpu_consistency || {
-    echo "[INFO] CPU-only build completed (GPU tests skipped as expected)"
+echo "--- Step 1: Generate CPU reference outputs ---"
+rm -rf cpu_gpu_comparison
+mkdir -p cpu_gpu_comparison
+./test_cpu_gpu_consistency --dump-prefix cpu_gpu_comparison/cpu_ref || {
+    echo "[FAIL] CPU reference generation failed!"
+    exit 1
+}
+./test_solver_cpu_gpu --dump-prefix cpu_gpu_comparison/cpu_ref || {
+    echo "[FAIL] Solver CPU reference generation failed!"
+    exit 1
+}
+./test_time_history_consistency --dump-prefix cpu_gpu_comparison/cpu_ref || {
+    echo "[FAIL] Time-history CPU reference generation failed!"
+    exit 1
 }
-
-# Now run with the GPU-offload build
-cd "$WORKDIR/build_ci_gpu_correctness"
 
 echo ""
-echo "--- Running test_cpu_gpu_consistency (GPU-offload build) ---"
-echo "This test compares CPU and GPU execution paths within the same binary."
-./test_cpu_gpu_consistency || {
-    echo "[FAIL] GPU consistency test failed!"
+echo "--- Step 2: Run GPU and compare against CPU reference ---"
+if [ ! -d "$WORKDIR/build_ci_gpu_correctness" ]; then
+    echo "[INFO] build_ci_gpu_correctness not found; creating GPU-offload build..."
+    rm -rf "$WORKDIR/build_ci_gpu_correctness"
+    mkdir -p "$WORKDIR/build_ci_gpu_correctness"
+    cd "$WORKDIR/build_ci_gpu_correctness"
+
+    echo "=== CMake Configuration (GPU-offload) ==="
+    CC=nvc CXX=nvc++ cmake .. -DCMAKE_BUILD_TYPE=Release -DUSE_GPU_OFFLOAD=ON 2>&1 | tee cmake_config.log
+    echo ""
+    echo "=== Building (GPU-offload) ==="
+    make -j8
+else
+    cd "$WORKDIR/build_ci_gpu_correctness"
+    echo "=== Building (GPU-offload incremental) ==="
+    make -j8
+fi
+
+./test_cpu_gpu_consistency --compare-prefix "$WORKDIR/build_ci_cpu_ref/cpu_gpu_comparison/cpu_ref" || {
+    echo "[FAIL] GPU vs CPU comparison failed!"
+    exit 1
+}
+./test_solver_cpu_gpu --compare-prefix "$WORKDIR/build_ci_cpu_ref/cpu_gpu_comparison/cpu_ref" || {
+    echo "[FAIL] Solver GPU vs CPU comparison failed!"
+    exit 1
+}
+./test_time_history_consistency --compare-prefix "$WORKDIR/build_ci_cpu_ref/cpu_gpu_comparison/cpu_ref" || {
+    echo "[FAIL] Time-history GPU vs CPU comparison failed!"
     exit 1
 }
 
 echo ""
 echo "[PASS] CPU-only vs GPU-offload comparison completed successfully"
+echo "      GPU results match CPU reference within tolerance"
 

.github/workflows/ci.yml

@@ -2,42 +2,24 @@ name: CI
 
 on:
   push:
-    branches: [ main, master, develop ]
-    paths-ignore:
-      - '**.md'
-      - 'docs/**'
-      - 'LICENSE'
-      - '.gitignore'
-  pull_request:
-    branches: [ main, master, develop ]
-    paths-ignore:
-      - '**.md'
-      - 'docs/**'
-      - 'LICENSE'
-      - '.gitignore'
 
 jobs:
   build-and-test:
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest, macos-latest]
+        os: [ubuntu-latest]
         build_type: [Release, Debug]
 
     steps:
     - uses: actions/checkout@v3
 
-    - name: Install dependencies (Ubuntu)
-      if: runner.os == 'Linux'
+    - name: Install dependencies
       run: |
         sudo apt-get update
         sudo apt-get install -y cmake build-essential
 
-    - name: Install dependencies (macOS)
-      if: runner.os == 'macOS'
-      run: |
-        brew install cmake
 
     - name: Configure CMake
       run: |
@@ -152,4 +134,3 @@ jobs:
           exit 1
         fi
         echo "No warnings detected (unused-variable warnings excluded)"
-

.github/workflows/gpu-ci.yml

@@ -1,12 +1,7 @@
 name: GPU CI
 
 on:
-  pull_request:
-    paths-ignore:
-      - '**.md'
-      - 'docs/**'
-      - 'LICENSE'
-      - '.gitignore'
+  push:
 
 jobs:
   gpu-tests:

.gitignore

@@ -86,12 +86,12 @@ model_comparison.*
 # Logs
 *.log
 
-# Keep example models but ignore trained models
+# NN Models: commit curated models, ignore transient training runs
+# Keep: curated models (*_caseholdout), README
+# Ignore: job-id suffixed directories (e.g., tbnn_channel_2996744)
+data/models/*_[0-9][0-9][0-9][0-9][0-9][0-9][0-9]/
 data/models/test_*
-data/models/*_real
-data/models/tbnn_*
-data/models/mlp_*
-!data/models/example_*
+!data/models/*_caseholdout/
 !data/models/README.md
 
 # Temp files

CMakeLists.txt

@@ -186,13 +186,13 @@ if(BUILD_TESTS)
     target_link_libraries(test_nn_integration nn_cfd_core)
     add_test(NAME NNIntegrationTest COMMAND test_nn_integration)
 
-    add_executable(test_gpu_execution tests/test_gpu_execution.cpp)
-    target_link_libraries(test_gpu_execution nn_cfd_core)
-    add_test(NAME GPUExecutionTest COMMAND test_gpu_execution)
+    add_executable(test_backend_execution tests/test_backend_execution.cpp)
+    target_link_libraries(test_backend_execution nn_cfd_core)
+    add_test(NAME BackendExecutionTest COMMAND test_backend_execution)
 
     add_executable(test_cpu_gpu_consistency tests/test_cpu_gpu_consistency.cpp)
     target_link_libraries(test_cpu_gpu_consistency nn_cfd_core)
-    add_test(NAME CPUGPUConsistencyTest COMMAND test_cpu_gpu_consistency)
+    add_test(NAME ConsistencyTest COMMAND test_cpu_gpu_consistency)
 
     add_executable(test_solver_cpu_gpu tests/test_solver_cpu_gpu.cpp)
     target_link_libraries(test_solver_cpu_gpu nn_cfd_core)
@@ -224,6 +224,11 @@ if(BUILD_TESTS)
     add_executable(test_turbulence_guard tests/test_turbulence_guard.cpp)
     target_link_libraries(test_turbulence_guard nn_cfd_core)
     add_test(NAME NanInfGuardTest COMMAND test_turbulence_guard)
+
+    # Turbulence feature tests - analytic validation of features, invariants, and model response
+    add_executable(test_turbulence_features tests/test_turbulence_features.cpp)
+    target_link_libraries(test_turbulence_features nn_cfd_core)
+    add_test(NAME TurbulenceFeaturesTest COMMAND test_turbulence_features)
 endif()
 
 # Installation

README.md

@@ -55,8 +55,9 @@ make -j4
 # EARSM (Wallin-Johansson)
 ./channel --model earsm_wj --adaptive_dt
 
-# Neural network model
-./channel --model nn_tbnn --nn_preset test_tbnn --adaptive_dt
+# Neural network models
+./channel --model nn_mlp --nn_preset mlp_channel_caseholdout --adaptive_dt  # Fast, GPU-optimized
+./channel --model nn_tbnn --nn_preset tbnn_channel_caseholdout --adaptive_dt  # Physically consistent
 
 # Periodic hills
 ./periodic_hills --Nx 64 --Ny 96 --model baseline --adaptive_dt

data/README.md

@@ -1,105 +1,29 @@
-# Neural Network Weights Directory
+# Data Directory
 
-This directory contains neural network weights and scaling parameters for the NN-based turbulence models.
+This directory contains data assets for the CFD-NN project.
 
-**Organization:**
-- Root directory (`data/`): Legacy example weights (kept for backward compatibility)
-- `models/` subdirectory: Organized model storage with metadata (recommended)
-  - `models/example_scalar_nut/` - Example MLP scalar eddy viscosity model
-  - `models/example_tbnn/` - Example TBNN tensor basis model
+## Neural Network Models
 
-To use a model from the `models/` directory, use `--nn_preset <model_name>` (e.g., `--nn_preset example_tbnn`).
+**All neural network turbulence model weights are stored in `data/models/`.**
 
-## File Format
+See `data/models/README.md` for:
+- Available trained models
+- How to use models with `--nn_preset`
+- Training your own models
+- Model architecture details
 
-### For `nn_mlp` model (scalar eddy viscosity)
+### Quick Start
 
-```
-layer0_W.txt    # Weight matrix for layer 0 (space-separated, row-major)
-layer0_b.txt    # Bias vector for layer 0 (one value per line)
-layer1_W.txt    # Weight matrix for layer 1
-layer1_b.txt    # Bias vector for layer 1
-...
-input_means.txt # Input feature means (one per line)
-input_stds.txt  # Input feature standard deviations (one per line)
-```
-
-### For `nn_tbnn` model (TBNN anisotropy)
-
-Same format as above, but the output dimension should match the number of tensor basis functions (typically 4 for 2D).
-
-## Exporting Weights from Python
-
-### PyTorch Example
-
-```python
-import torch
-import numpy as np
-
-# Load your trained model
-model = torch.load('model.pth')
-model.eval()
-
-# Export each layer
-for i, layer in enumerate(model.layers):
-    W = layer.weight.detach().cpu().numpy()  # Shape: (out_features, in_features)
-    b = layer.bias.detach().cpu().numpy()    # Shape: (out_features,)
-
-    # Save in space-separated format
-    np.savetxt(f'layer{i}_W.txt', W, fmt='%.16e')
-    np.savetxt(f'layer{i}_b.txt', b, fmt='%.16e')
-
-# Export input scaling (if used during training)
-np.savetxt('input_means.txt', feature_means, fmt='%.16e')
-np.savetxt('input_stds.txt', feature_stds, fmt='%.16e')
-```
-
-### TensorFlow/Keras Example
-
-```python
-import tensorflow as tf
-import numpy as np
+```bash
+# Use trained TBNN model for channel flow
+./channel --model nn_tbnn --nn_preset tbnn_channel_caseholdout
 
-# Load your trained model
-model = tf.keras.models.load_model('model.h5')
+# Use trained TBNN model for periodic hills
+./periodic_hills --model nn_tbnn --nn_preset tbnn_phll_caseholdout
 
-# Export each dense layer
-layer_idx = 0
-for layer in model.layers:
-    if isinstance(layer, tf.keras.layers.Dense):
-        W, b = layer.get_weights()  # W shape: (in_features, out_features)
-        W = W.T  # Transpose to match C++ convention (out, in)
-
-        np.savetxt(f'layer{layer_idx}_W.txt', W, fmt='%.16e')
-        np.savetxt(f'layer{layer_idx}_b.txt', b, fmt='%.16e')
-        layer_idx += 1
+# Use example/demo models (random weights - for testing only)
+./channel --model nn_tbnn --nn_preset example_tbnn
+./channel --model nn_mlp --nn_preset example_scalar_nut
 ```
 
-## Example File Contents
-
-**layer0_W.txt** (3 inputs, 2 outputs):
-```
-1.234567e-01 -2.345678e-01  3.456789e-01
-4.567890e-01  5.678901e-01 -6.789012e-01
-```
-
-**layer0_b.txt**:
-```
-7.890123e-02
--8.901234e-02
-```
-
-## Feature Ordering
-
-Features should match the order expected by the C++ code. For scalar `nu_t` model:
-
-0. Normalized strain rate magnitude
-1. Normalized rotation rate magnitude  
-2. Normalized wall distance
-3. Strain-rotation ratio
-4. Local Reynolds number
-5. Normalized velocity magnitude
-
-For TBNN model, features include tensor invariants as defined in `features.cpp`.
-
-
+**Note:** NN models now require explicit selection via `--nn_preset` or `--weights/--scaling`. There are no default/fallback weights.