harshaperla/fan-design-surrogate-physicsnemo

Fan Design Surrogate Model — PhysicsNeMo Version

A physics-informed deep learning surrogate model for axial fan aerodynamic performance prediction, built with NVIDIA PhysicsNeMo.

R² = 0.9964 | MAPE = 1.40% (32% better than pure data-driven approach)

How This Differs from the Standard PyTorch Version

Aspect	Standard PyTorch	PhysicsNeMo (this model)
Framework	Pure PyTorch nn.Module	PhysicsNeMo FullyConnected (extends nn.Module)
Loss Function	MSE on data only	MSE + physics-informed constraints
Architecture	Custom ResidualMLP	PhysicsNeMo FullyConnected with skip_connections=True
Activation	GELU	SiLU (Swish)
Parameters	634K	1.33M
Avg R²	0.9952	0.9964 (+0.12%)
Avg MAPE	2.08%	1.40% (-32%)
Extrapolation	Poor (data-only)	Better (physics constraints regularize)
Serialization	torch.save()	.mdlus native + torch.save()
Checkpoint	Manual	physicsnemo.utils.checkpoint

Key Advantage: Physics-Informed Loss

The standard model learns purely from data. The PhysicsNeMo version adds 4 physics constraint terms to the loss function that enforce thermodynamic and aerodynamic consistency:

1. Power-Pressure Consistency: P = ΔPt × Q / η (energy conservation)
2. Flow Rate Consistency: Q = Vx × A_annulus (continuity equation)
3. Efficiency Bounds: 0 < η < 1 (second law of thermodynamics)
4. Euler Work Equation: ΔPt ≈ ρ × η × U × ΔVθ (turbomachinery fundamental)

These constraints act as regularizers — even without labeled data, the model learns physically meaningful relationships. This dramatically improves:

• Extrapolation to designs outside the training distribution
• Consistency between predicted outputs (power, pressure, flow are now thermodynamically coherent)
• Sample efficiency (same accuracy with less training data)

Performance (Test Set: 2000 samples)

Output Metric	R²	MAPE
Total Pressure Rise (Pa)	0.9976	3.39%
Isentropic Efficiency	0.9804	0.69%
Power Consumption (W)	0.9987	2.70%
Flow Rate (m³/s)	0.9998	1.08%
Specific Speed	0.9975	1.48%
Degree of Reaction	0.9988	0.46%
Diffusion Factor	0.9991	0.88%
Noise Estimate (dBA)	0.9994	0.51%
Average	0.9964	1.40%

Usage

import torch
import numpy as np
import json
from physicsnemo.models.mlp.fully_connected import FullyConnected

# Load model
model = FullyConnected(
    in_features=14, out_features=8,
    num_layers=6, layer_size=512,
    activation_fn='silu', skip_connections=True,
)

# Option 1: Load from PhysicsNeMo native format
model = FullyConnected.from_checkpoint("fan_surrogate.mdlus")

# Option 2: Load from PyTorch state dict  
state = torch.load("model.pt", map_location="cpu")
model.load_state_dict(state)
model.eval()

# Load scalers
with open("scalers.json") as f:
    scalers = json.load(f)
sx_mean = np.array(scalers['scaler_X_mean'], dtype=np.float32)
sx_scale = np.array(scalers['scaler_X_scale'], dtype=np.float32)
sy_mean = np.array(scalers['scaler_y_mean'], dtype=np.float32)
sy_scale = np.array(scalers['scaler_y_scale'], dtype=np.float32)

# Predict
INPUT_COLS = ["blade_inlet_angle_deg", "blade_turning_angle_deg", "chord_length_mm",
    "blade_thickness_ratio", "stagger_angle_deg", "hub_tip_ratio",
    "tip_clearance_ratio", "num_blades", "aspect_ratio", "solidity",
    "sweep_angle_deg", "flow_coefficient", "rotational_speed_rpm", "tip_radius_mm"]
LOG_IDX = [0, 2, 3]

design = [55, 15, 100, 0.06, 45, 0.5, 0.015, 12, 2.5, 1.0, 0, 0.5, 3000, 300]
x = np.array([design], dtype=np.float32)
x_scaled = (x - sx_mean) / sx_scale

with torch.no_grad():
    y_scaled = model(torch.from_numpy(x_scaled)).numpy()

y_proc = y_scaled * sy_scale + sy_mean
for i in LOG_IDX:
    y_proc[0, i] = np.expm1(y_proc[0, i])

print(f"Pressure Rise: {y_proc[0,0]:.1f} Pa")
print(f"Efficiency: {y_proc[0,1]:.3f}")
print(f"Power: {y_proc[0,2]:.1f} W")
print(f"Flow Rate: {y_proc[0,3]:.3f} m³/s")

Installation

pip install nvidia-physicsnemo torch numpy

Architecture Details

FullyConnected(
  (layers): ModuleList(
    (0): FCLayer(Linear(14→512) + SiLU)
    (1-5): 5 × FCLayer(Linear(512→512) + SiLU + skip_connection)
  )
  (final_layer): FCLayer(Linear(512→8))
)
Total parameters: 1,325,064

When to Use PhysicsNeMo vs Standard PyTorch

Scenario	Recommendation
Abundant training data (>5000 samples)	Either works; PhysicsNeMo slightly better
Sparse data (<500 samples)	PhysicsNeMo (physics constraints compensate for missing data)
Extrapolation needed	PhysicsNeMo (constraints prevent unphysical predictions)
Pure interpolation	Standard PyTorch (simpler, faster training)
Need ONNX export / deployment	PhysicsNeMo (built-in export support)
Multi-physics coupling	PhysicsNeMo (add PDE residuals, boundary conditions)
Quick prototyping	Standard PyTorch (fewer dependencies)

Dataset

Trained on harshaperla/fan-design-dataset (8000 train / 2000 test).

References

• NVIDIA PhysicsNeMo — Framework
• C(NN)FD — CNN for turbomachinery CFD
• Standard version — Pure PyTorch baseline

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern