On the basis of a large number of experimental data, it is a challenge to establish a data-driven non-linear law between mixing characteristics and mechanical properties for the proportioning and process design of new alloy compositions. This paper proposes a performance-oriented “composition-process-property” design strategy for Al-Si-Mg alloys based on a machine learning approach, aiming to adopt multimodal experimental data on the composition, melting and heat treatment processes of divergent grades of the same system as features, and a random forest algorithm is used to find the non-linear pattern between the features and the tensile strength. Afterward, this paper sets the composition and process parameters of some of the alloys in the dataset as the target null values and uses the chain equation multiple interpolation algorithms to predict the interpolation of the target missing data. The errors of both experimental and predicted values of tensile strength of the alloys predicted or guided by this strategy are kept within ±5%; The composition ratio of Al-6.8Si-0.6Mg-0.05Sr and the heat treatment scheme of 540 ℃×10 h+170 ℃×10 h are experimentally confirmed to have a quality index Q_DJR of 517.3 for comprehensive tensile properties, which is higher than that of similar alloys below a Q_DJR value of 500. The result indicates that this strategy helps to enhance the long cycle time, high cost, and low efficiency of the traditional design method for Al-Si-Mg system alloys.