Experimental-data-driven thermal conductivity prediction and inverse composition design for alloys

This work develops a data-driven framework for predicting the thermal conductivity of metals and multi-component alloys and for inversely proposing compositions that meet a target conductivity. We collect, to our knowledge, the largest experimental dataset containing 6259 data points spanning 49 elements and temperatures from 0 to 1400 K. Using alloy composition and temperature as inputs, we train and benchmark several regression models and obtain high predictive accuracy reaching R ² > 0.99 and RMSE of 6-9 W m^-1 K^-1. The approach remains quantitatively reliable for challenging cases including dilute-doped Mg alloys and commercial steel over broad temperature ranges. Based on the trained forward model, we propose an inverse-design workflow to efficiently search composition space and suggest candidate alloys expected to achieve a specified thermal-conductivity target at a given temperature. The inverse search can identify composition windows where near-target conductivity is maintained over a finite concentration range to improve the practical ability for experimental validation and scalable process.