Interpretable machine learning for prediction and inverse design of electrical resistivity multicomponent aluminium alloys

Abstract

Aluminium is the standard interconnect metal in CMOS technology and can be deposited at low temperature, yet its intrinsically low electrical resistivity makes it inefficient as a Joule heating element, forcing microheater designs toward platinum or polysilicon at the cost of process complexity. This study reframes the problem as one of composition engineering rather than material substitution and develops a machine learning framework to predict and inversely design the electrical resistivity of aluminium alloys. A curated dataset of 220 commercial alloys, described by nine compositional features (Si, Fe, Cu, Mn, Mg, Cr, Zn, Ti, Zr) and spanning 2.80–6.40 µΩ·cm, was assembled from a public materials database. Four regression models were benchmarked: Linear Regression (R² = 0.635), Support Vector Regression (R² = 0.821), Random Forest (R² = 0.837), and Extreme Gradient Boosting, which performed best on an independent hold-out test partition (test R² = 0.861; test MAPE = 3.9%) and remained stable under k-fold cross-validation carried out within the training partition (CV-R² = 0.839; CV-RMSE = 0.396 µΩ·cm). The pronounced gap between the linear baseline and the non-linear learners confirms that resistivity in multicomponent aluminium alloys is governed by interacting solute effects. Feature importance analysis identified magnesium (0.554), copper (0.215), and zinc (0.153) as the dominant contributors, accounting for approximately 92% of the model's explanatory power, consistent with solid-solution and impurity scattering mechanisms. Surrogate-based inverse screening produced heavily alloyed Al–Zn–Mg–Cu candidates with predicted resistivity of 9.26–9.52 µΩ·cm. These values lie outside the multivariate domain of the training data: the candidates carry 24.2–32.3 wt.% total solute, leaving an aluminium balance of only 67.7–75.8 wt.%, and they exceed the training maximum for Cu, Zn and Zr in every case. They are therefore reported as an extrapolative indication of direction in composition space rather than as validated alloy candidates, and require aluminium-balance constraints, thermodynamic screening and experimental validation before any device-level claim can be made.