Benchmarked linear models, XGBoost, and a neural network for predicting material stiffness from microstructure. The tuned ANN hit a test RMSE of 2.32, versus ~8 for linear models.

Developed and compared machine learning models to predict material stiffness from microstructure features, using 9,000 observations described by the first 15 principal components of the microstructure data. Materials with nonlinear behavior are expensive to analyze traditionally, which is what makes a fast predictive model valuable.

What I did

• Models: linear regression, ridge and lasso, an artificial neural network (ANN), and XGBoost.
• Data pipeline: validation and z-score standardization, then a full benchmark suite.
• Tuning: random search on the ANN and XGBoost to maximize generalization.
• Evaluation: compared RMSE across models; simple linear models badly underfit (RMSE ~8.00).

Result

The tuned ANN performed best, with a test RMSE of 2.32 (train 1.91). It clearly beat the linear baseline and overfit less than XGBoost (test RMSE 2.65).