Spring Seminar with Professor Ashim Datta - Cornell University

Seminar Abstract
As a mechanistic modeler of food processes, it is hard for some of us to turn toward AI. Compared to AI-based models, mechanistic models provide deep insight and far superior extrapolation capabilities. However, because they are built on multiphysics, to be realistic, mechanistic models end up being quite complex, taking longer to develop with often very long run times, making design use of the more complex models somewhat nonexistent. A surrogate model, built from the output generated by the mechanistic model, can drastically reduce the runtime (a factor of 100 has been achieved). I will first describe a universal multiphysics framework for understanding and optimizing food processes that our group has been developing for the past 35 years (approximately since my sabbatical at UMN). Here food is treated as a deformable (swellable/shrinkable) porous medium where multicomponent transport (water, vapor, oil, flavors, etc.), evaporation/condensation, and reactions take place. This framework approach has become the primary one in the field when modeling complex changes in a solid food product during processing. Runtime however, especially in presence of deformation (solid mechanics), becomes prohibitive for optimization. To overcome this, we developed reduced-order surrogates, that is the second part of the presentation. 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 spatiotemporal evolution of temperature, moisture, and deformation fields with near-perfect accuracy and several orders-of-magnitude speedup. Entire industry is coming up that builds such surrogate models to eventually use them in optimization. But, training the surrogate model requires the outputs from a full-fidelity mechanistic model, creating renewed needs for mechanistic models.