The modern food industry faces the critical challenge of developing products that are safe, nutritious, and sustainably produced while maintaining consistent quality under diverse processing conditions. Achieving this balance requires not only empirical knowledge but also a mechanistic understanding of the underlying transport phenomena and their influence on safety and quality. This thesis advances the emerging field of computer-aided food engineering by combining detailed multiphysics modeling with advanced deep learning surrogate modeling and interactive computational tools that together enable predictive, interpretable, and efficient process design. In this work a comprehensive multiphase--multiphysics model for the cooking of muscle foods is formulated. Departing from conventional saturated-media assumptions, the meat domain is modeled as an unsaturated porous medium composed of coexisting liquid water, water vapor, air, and a deforming solid matrix. This generalized framework explicitly resolves the nonlinear coupling among heat transfer, mass transport, phase change, and deformation, that provides deeper insight about internal evaporation and the mode of moisture loss at the boundary that govern crust formation. The model not only explains the commonly observed development of a dried crust near the surface but also provides mechanistic insight into the evolution of juiciness, microbial lethality, and chemical safety based on HCA formation across different cooking regimes. While such high-fidelity models provide invaluable physical understanding, their computational cost—arising from solving strongly coupled nonlinear partial differential equations—limits their routine use for real-time analysis or process optimization. To overcome this, we developed a rigorous framework for constructing reduced-order surrogates that retain the physical richness of the mechanistic models while operating at dramatically lower cost. Through systematic evaluation of dimensionality-reduction techniques and sequence-learning architectures, Transformer-based networks emerged as the most effective, demonstrating the capability to reproduce the full spatio-temporal evolution of temperature, moisture, and deformation fields with near-perfect accuracy and several orders-of-magnitude speedup. This work also provides a unified methodology for optimizing surrogate design, establishing best practices for preprocessing, latent-space compression, and temporal learning in multiphysics systems. Finally, these advances are translated into a deployable predictive application designed for real-time estimation of food safety and quality during drying. Built on a validated multiphase model and accelerated by the trained Transformer surrogate, the application allows users to input process and product parameters and instantly visualize spatio-temporal fields, microbial kinetics, and derived quality indices through an intuitive, multi-platform interface. The tool, built using a mechanistic model that is validated extensively against a mango drying experiment in a commercial dehydrator, demonstrates how deep-learning surrogates of mechanistic models can provide fast, scientifically grounded predictions suitable for industrial decision-making. Together, the methodologies and tools presented in this thesis bridge fundamental food physics and applied process engineering, establishing a pathway toward next-generation digital twins for food manufacturing that unify mechanistic insight, computational efficiency, and practical usability.