You can deploy this model and start earning money today!

harshaperla/
fan-design-surrogate

Name: fan-design-surrogate
Availability: InStock

ml-intern

2306.05889

2302.14740

Views

No views yet

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

python

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)

Limitations

Trained on synthetic data from simplified correlations (not full 3D CFD)
Single operating point per design (no off-design/stall prediction)
Mid-span only (no radial profile prediction)
No unsteady effects (flutter, rotating stall)
Axial fan only (not centrifugal)

References

C(NN)FD: Deep learning for turbomachinery CFD (arxiv:2306.05889)
Lieblein, S. (1965). Loss and stall analysis of compressor cascades
Denton, J.D. (1993). Loss mechanisms in turbomachines
Ainley & Mathieson (1951). Secondary flow losses
Carter, A.D.S. (1950). Blade deviation angles
Fusion of ML with numerical simulation for propeller design (arxiv:2302.14740)

Training Details

Dataset: 8000 train / 2000 test (LHS sampling)
Optimizer: AdamW (lr=3e-3, weight_decay=1e-4)
Schedule: Cosine Annealing (200 epochs)
Loss: MSE on standardized outputs
Preprocessing: StandardScaler (inputs) + log1p + StandardScaler (outputs)
Training time: ~73s on CPU

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

harshaperla/fan-design-surrogate