Fan Design Surrogate Model

A deep learning surrogate model for predicting axial fan aerodynamic performance from geometric design parameters. Replaces expensive CFD simulations with real-time predictions (R² = 0.995, MAPE = 2.08%).

Model Description

This is a Residual MLP neural network trained to predict 8 aerodynamic performance metrics from 14 fan blade design parameters. It serves as a fast surrogate for Computational Fluid Dynamics (CFD) simulations in fan design optimization workflows.

Architecture

Type: Deep Residual MLP with LayerNorm
Parameters: 634,504
Input: 14 geometric/operating parameters → standardized
Output: 8 performance metrics (with log-transform for power-law outputs)
Backbone: 4 residual blocks (256 hidden dim)
Activation: GELU
Training: AdamW + Cosine Annealing LR, 200 epochs

Performance (Test Set, 2000 samples)

Output Metric	R²	MAPE
Total Pressure Rise (Pa)	0.9981	3.21%
Isentropic Efficiency	0.9795	0.70%
Power Consumption (W)	0.9960	5.33%
Flow Rate (m³/s)	0.9984	2.78%
Specific Speed	0.9954	1.74%
Degree of Reaction	0.9979	0.61%
Diffusion Factor	0.9983	1.29%
Noise Estimate (dBA)	0.9980	1.01%
Average	0.9952	2.08%

Input Parameters

Parameter	Range	Unit	Description
blade_inlet_angle_deg	40–70	°	Relative inlet flow angle from axial
blade_turning_angle_deg	5–30	°	Flow turning (β1 - β2)
chord_length_mm	50–150	mm	Blade chord length
blade_thickness_ratio	0.04–0.12	-	Maximum thickness / chord
stagger_angle_deg	30–60	°	Blade stagger angle
hub_tip_ratio	0.3–0.7	-	Hub-to-tip radius ratio
tip_clearance_ratio	0.005–0.03	-	Tip gap / blade height
num_blades	6–20	-	Number of blades
aspect_ratio	1.0–4.0	-	Blade height / chord
solidity	0.5–1.5	-	Chord / pitch
sweep_angle_deg	-15–15	°	Blade sweep angle
flow_coefficient	0.3–0.7	-	Vx / U_tip
rotational_speed_rpm	1000–5000	rpm	Rotational speed
tip_radius_mm	150–500	mm	Fan tip radius

Output Parameters

Output	Unit	Description
total_pressure_rise_Pa	Pa	Total-to-total pressure rise
isentropic_efficiency	-	Isentropic efficiency (0–1)
power_consumption_W	W	Shaft power consumption
flow_rate_m3s	m³/s	Volumetric flow rate
specific_speed	-	Dimensionless specific speed
degree_of_reaction	-	Degree of reaction (0–1)
diffusion_factor	-	Lieblein diffusion factor
noise_estimate_dBA	dBA	Sound pressure level estimate

Usage

import torch
import numpy as np
import json

# Load model
from model import FanSurrogate, predict_fan_performance

# Single prediction
design = {
    'blade_inlet_angle_deg': 55.0,
    'blade_turning_angle_deg': 15.0,
    'chord_length_mm': 100.0,
    'blade_thickness_ratio': 0.06,
    'stagger_angle_deg': 45.0,
    'hub_tip_ratio': 0.5,
    'tip_clearance_ratio': 0.015,
    'num_blades': 12,
    'aspect_ratio': 2.5,
    'solidity': 1.0,
    'sweep_angle_deg': 0.0,
    'flow_coefficient': 0.5,
    'rotational_speed_rpm': 3000,
    'tip_radius_mm': 300.0,
}

results = predict_fan_performance(design)
print(f"Pressure Rise: {results['total_pressure_rise_Pa']:.1f} Pa")
print(f"Efficiency: {results['isentropic_efficiency']:.3f}")
print(f"Power: {results['power_consumption_W']:.1f} W")
print(f"Flow Rate: {results['flow_rate_m3s']:.3f} m³/s")
print(f"Noise: {results['noise_estimate_dBA']:.1f} dBA")

Physics Background

The training data was generated using established turbomachinery correlations:

Euler's equation: Work input = U × ΔVθ
Lieblein's correlations: Profile loss from diffusion factor
Denton's model: Tip clearance loss
Ainley-Mathieson: Secondary/endwall losses
Carter's rule: Flow deviation at blade trailing edge
de Haller criterion: Diffusion limit (W2/W1 > 0.72)
Madison's formula: Fan noise with tip speed correction

Intended Use

Primary: Rapid fan performance screening in design optimization loops
Secondary: Design space exploration, sensitivity analysis, preliminary design
Not for: Final design verification (use full CFD for that)