Tutorial 4 Report — ME3101
Data Analytics and Machine Learning for Engineers
Due: 09/09/2025
Name: __________________
Roll No.: _______________
Part 1 – Atmospheric Pressure vs Altitude (15 marks)
1.1 Simulated Data
Generated synthetic data for low-noise (σ = 2) and high-noise (σ = 10) scenarios.
Figure 1: Plot of low-noise dataset.
Figure 2: Plot of high-noise dataset.
1.2 Regression Fits
Fitted the model P(h) = a exp(-k h) using:
• Linear regression on log-transformed data
• Nonlinear regression (curve_fit)
Figure 3: Fits for low-noise data.
Figure 4: Fits for high-noise data.
1.3 Parameter Estimates and Errors
Case Method a k MSE
Low noise Linear 102.11 0.1213 4.07
Low noise Nonlinear 102.51 0.1215 4.05
High noise Linear 104.87 0.1520 81.79
High noise Nonlinear 93.63 0.1252 64.62
1.4 Discussion
Linear regression performs poorly because the noise model is not Gaussian in the log-domain.
Nonlinear regression minimizes residuals in the correct domain and performs better, especially
under high noise.
Part 2 – Beam Optimization (35 marks)
2.1 Unconstrained Optimization
Stress σ(w) = C w with C = 3.9×10^10 Pa/m. Function is monotonically increasing in w.
Unconstrained minimum at w → 0, h → ∞, which is impractical.
2.2 Width Constraint (w ≥ 0.02 m)
Penalty term: P1(w) = (max[0, 0.02 – w])^2. Optimal w* = 0.02 m, h* = 0.1 m, σ = 780 MPa.
Figure 5: Stress vs width with width constraint.
2.3 Width + Stress Constraints (σ ≤ 12 MPa)
Penalty term: P2(w) = (max[0, Cw – σ_max])^2. No feasible solution because required w ≤ 0.0003
m conflicts with manufacturing constraint. Solver returns w = 0.02 m, but σ = 780 MPa, infeasible.
Figure 6: Stress vs width with stress constraint (infeasible region highlighted).
2.4 Weaker Material (σ ≤ 2 MPa)
Even stricter constraint. Required w ≤ 0.00005 m. Conflict with manufacturing bound → infeasible.
Solver again returns w = 0.02 m, σ = 780 MPa. Design not possible under given parameters.
Figure 7: Stress vs width with σ ≤ 2 MPa constraint (infeasible).
Conclusion
In Q1, nonlinear regression clearly outperformed linear regression, especially under high noise. In
Q2, unconstrained optimization yielded impractical results. Adding width constraints gave a
mathematically feasible but unrealistic stress level. Stress constraints revealed the problem was
infeasible under the given design parameters, suggesting that area or load must be modified for a
realistic design.