In this research a machine learning model incorporating uncertainty to enhance the creep-life prediction and high-throughput design of creep-resistant steel is proposed. The framework integrates key physical metallurgical parameter linked to precipitate coarsening and applies transfer learning to correlate short-time tensile properties with the creep performance, all within a Bayesian convolutional neural network. Unlike conventional machine learning models, which often lack an assessment of prediction credibility, this uncertainty-based approach offers more accurate and stable predictions while also providing a measure of prediction credibility. By combining the model with a genetic algorithm, the framework achieves a balance between creep life optimization and uncertainty, thereby supporting robust alloy design. The validation on newly developed martensitic heat-resistant steels with tolerable prediction uncertainty showed excellent alignment between predicted and experimentally determined creep life, underscoring the effectiveness of the framework. These findings highlight the critical role of uncertainty modeling in advancing machine learning applications for alloy design.