Multi-Property Modeling of Aluminum Alloys Using Symbolic Regression (GP and GE)

Sammanfattning

Predicting the mechanical properties of aluminium alloys from composition and processing parameters is an essential advancement for accelerating alloy design, yet existingmachine learning approaches predominantly employ black-box models that offer limitedphysical insight. This thesis investigates whether Symbolic regression (SR) can produceinterpretable closed-form expressions to predict Yield Strength (YS) and Ultimate TensileStrength (UTS) of AA5xxx aluminium alloys from a limited experimental dataset of approximately 313 records. Two symbolic regression paradigms are explored: Genetic Programming(GP) and Grammatical Evolution (GE) A two-stage GE+Ridge and a three-stage GP+AST+Ridge framework are proposed,where GE+Ridge combines GE for symbolic structure search with Ridge regression forcoefficient optimisation, and is compared with five baseline models (OLS, Ridge, Lasso,Random Forest, XGBoost) across four experimental configurations and 30 independent runsper condition. GE+Ridge substantially outperforms plain GE, raising median test R2from 0.754 to0.852 for YS and reducing run-to-run standard deviation from 0.188 to 0.015. In theGP+AST+Ridge framework, GP evolved trees are decomposed via Abstract Syntax Treeparsing before Ridge coefficient optimisation to enhance numerical stability over pure GPand achieved Ridge R2values (up to 0.836 for YS and 0.907 for UTS). With grain size included, GE+Ridge achieves median R2 = 0.908 for YS and 0.934for UTS, approaching the XGBoost performance while providing interpretable expressions.Conservative data augmentation primarily improves generalisation reliability rather thaninflating test-set performance. The evolved expressions demonstrate strong physical consistency without explicit constraints. The Labusch c2/3solid-solution strengthening exponent emerges spontaneously forMg and Mn, and grain size recovers the Hall–Petch relationship directly from data.These results demonstrate that GE+Ridge is a viable approach to mechanical propertyprediction in AA5xxx alloys, bridging predictive accuracy and physical interpretability.