Data-driven methods are becoming increasingly popular in the field of materials science. While most data-driven models are trained on simulation data as it is relatively easier to collect a large amount of data from physics-based simulations, there are many challenges in applying data-driven methods on experiments: 1) experimental data is usually not clean; and 2) it generally has a greater degree of heterogeneity. In this project, we have developed a data-driven methodology to address these challenges on an industrial magnet dataset, where the goal is to predict magnetic properties (forward models) at different stages of the experimental workflow. The data-driven methodology consists of data cleaning, data preprocessing, feature extraction, and model development using traditional machine learning and deep learning methods to accurately predict magnet properties. In particular, we have developed three different types of predictive models: 1) numerical model using only numerical data containing composition and processing information; 2) image model using image data representing structure information; and 3) combination model using both types of data together. In addition to predictive models, the analysis and comparison of results across the models provide several interesting data-driven insights. Such data-driven analytics has the potential to help guide future experiments and realize the inverse models, which could significantly reduce costs and accelerate the discovery of new magnets with superior properties. The proposed models are already deployed in Toyota Motor Corporation.