machine-learning-engineering

name	machine-learning-engineering
description	Machine learning for engineering — regression/classification for property prediction, neural networks, physics-informed ML (PINN), surrogate models, Gaussian process, materials informatics, process optimization.
metadata	{"priority":7,"promptSignals":{"phrases":["machine learning engineering","ML engineering","surrogate model","neural network engineering","physics informed neural","PINN","Gaussian process engineering","materials informatics"],"minScore":3}}

Machine Learning for Engineering — Complete Skill

Problem Types in Engineering ML

Regression (Continuous Output)

Predict: material property, simulation output, process parameter, fatigue life Input features: composition, process conditions, geometry, loading Output: scalar (S_u, k, ε_f) or vector (S-N curve, stress distribution)

Classification (Discrete Output)

Predict: failure mode, defect type, material phase, weld quality Input: sensor signals, microstructure images, process data Output: class labels (defect/no-defect, mode I/II/III fracture)

Anomaly Detection

Detect: bearing wear, cavitation onset, tool wear Compare current signature to baseline → novelty score → alert

Feature Engineering for Engineering Problems

Composition-based features (materials): Elemental properties: atomic radius, electronegativity, melting point, valence electron count Statistics: mean, std, min, max, range across all elements Magpie (Materials Agnostic Platform for Informatics and Exploration): generates 145 features Matminer library: composition → feature vector

Geometry features: L/D ratio, t/r ratio, stress concentration factors, moment of inertia/section modulus

Signal features (vibration, acoustic): Time domain: RMS, kurtosis, peak, crest factor Frequency domain: spectral centroid, power in bands, gear mesh frequency amplitude Time-frequency: wavelet energy, STFT features

Neural Networks for Engineering

Feedforward (MLP)

Input → hidden layers (relu activation) → output y = W_n × ... relu(W₂ × relu(W₁ × x + b₁) + b₂) ... + b_n

Training: Adam optimizer; batch gradient descent; dropout for regularization Architecture guidance: 2–5 layers; 64–512 neurons per layer; start simple

Weight initialization: He for relu; Xavier for sigmoid/tanh Batch normalization: stabilizes training; allows higher learning rates

Convolutional Neural Networks (CNN) — Microstructure

For image inputs (SEM, TEM, CT scans): Convolutional layers extract spatial features → flatten → FC layers Applications: grain size measurement, defect classification, phase segmentation

Recurrent Networks (LSTM, GRU) — Time Series

For sequential data: vibration waveform, sensor history, cycle-by-cycle fatigue LSTM cell: input, forget, output gates → handles long-term dependencies Application: remaining useful life prediction from vibration/temperature time series

Physics-Informed Neural Networks (PINN)

Concept: neural network satisfies both data AND governing differential equations

Loss function: L = L_data + λ_physics × L_physics

L_data = Σ (y_predicted - y_measured)² L_physics = Σ (PDE_residual at collocation points)²

For heat equation ∂T/∂t = α ∇²T: L_physics = |∂T_NN/∂t - α (∂²T_NN/∂x² + ∂²T_NN/∂y²)|² evaluated at random collocation points

Automatic differentiation: compute PDE residuals using backpropagation Frameworks: DeepXDE, NeuroDiffEq, JAX, PyTorch (autograd)

Applications: solving PDEs in complex domains; inverse problems (identify material properties from field data); data-sparse regimes

Gaussian Process Regression (GPR)

Bayesian non-parametric: provides prediction + uncertainty

f(x) ~ GP(μ(x), k(x,x')) [mean function + kernel/covariance function]

Prediction: posterior mean and variance given training data μ_* = K_K^{-1}y (predictive mean) σ²_ = k_** - K_K^{-1}K_^T (predictive variance → uncertainty)

Kernels: RBF (squared exponential): k(x,x') = σ² exp(-|x-x'|²/2l²) Matern 5/2: good for non-smooth functions; often better for engineering Periodic: for cyclic processes

Advantages: exact uncertainty; works with small data (10s–100s); interpretable Limitation: O(n³) training complexity; poor for > 1000 training points

Surrogate Models (Metamodels)

Replace expensive simulations (CFD, FEA) with fast approximation

Workflow:

1. Design of Experiments (DoE): Latin Hypercube Sampling (LHS), space-filling
2. Run high-fidelity simulation at each design point (parallel HPC)
3. Train surrogate: GPR, kriging, neural network, RBF interpolation
4. Use surrogate for optimization, sensitivity analysis, uncertainty quantification

Validation: leave-one-out cross-validation; RMSE, R²; add more points if error too large

Active Learning (Bayesian Optimization)

Sequential selection of experiments/simulations to maximize information

Acquisition functions: Expected Improvement: EI(x) = E[max(f(x) - f*, 0)] Upper Confidence Bound: UCB(x) = μ(x) + κσ(x)

Applications: Material discovery: select next alloy composition to test Process optimization: tune process parameters to maximize yield

Materials Informatics Pipeline

1. Data curation: literature mining (Elsevier, SpringerMaterials), experiments, AFLOW, Materials Project, ICSD
2. Featurization: Magpie, DScribe, matminer, RDKit (for organic)
3. Model: GP, RF, GBM, or neural network
4. Validation: stratified cross-validation on composition splits
5. Prediction + uncertainty: flag high-uncertainty predictions for experimental validation

Materials Project API: 150,000+ DFT-calculated compounds; E_hull, bandgap, elastic constants

Uncertainty Quantification in ML

Ensemble methods: train N models with different random seeds → variance = epistemic uncertainty MC Dropout: sample predictions with dropout active → uncertainty from variability Bayesian NN: full posterior over weights → exact uncertainty; computationally expensive

Calibration: predicted uncertainty should match true error → check with calibration plots

Output

Provide: ML model architecture/type, feature set, training data size, validation metrics (RMSE, R², MAPE), uncertainty estimate, physical consistency check (does ML prediction respect known physics limits), computational cost vs. high-fidelity simulation, recommendation for active learning next points.