Multiphysics models can assist in geometric design of frozen microwaveable foods, but the conventional 'parametric-sweeping' strategy is computationally intensive. This study develops an online machine learning (ML)-supervised multiphysics modeling strategy to simultaneously optimize multiple geometrical parameters (e.g., top surface width, top surface length, and the ratio of top-to-bottom dimensions) for optimal microwave heating uniformity. First, one or more paired geometric dimensional parameters and heating uniformity results obtained from multiphysics modeling are used as initial training data for the ML optimization process that integrates Gaussian Process Regression (GPR) and Bayesian optimization. Then, the multiphysics modeling procedure is supervised by an ML process to generate new paired geometry-heating uniformity results to expand the training dataset. The loop of multiphysics modeling and ML optimization is conducted to effectively identify geometry designs with good heating uniformity. Results indicate that the ML-supervised optimization strategy can efficiently identify good geometric designs with low heating uniformity by using much fewer multiphysics models than the parametric sweep approach. The ML-supervised approach also exhibits great robustness, delivering good performance even with small (as little as one model) and randomly selected initial training data. This ML-supervised approach is flexible and can be modified to meet specific needs for broad industrial implementation in the development of microwaveable food.