Deep Learning-Based Prediction of Dough Rheological Properties Under Repeated Freeze–Thaw Conditions

Abstract

Repeated freeze–thaw conditions impose cumulative structural and rheological changes on dough, altering its viscoelastic behavior, water distribution, gluten network integrity, and fermentation performance. Predicting these changes is difficult because dough is a heterogeneous, history-dependent material whose behavior emerges from coupled biochemical, colloidal, and mechanical processes. This paper examines deep learning-based prediction of dough rheological properties under repeated freeze–thaw cycles from a system-level perspective. It synthesizes literature on frozen dough science, rheological measurement, machine learning architectures, deployment infrastructure, governance, and sustainability. The discussion emphasizes that predictive performance depends not only on model selection but also on sensing quality, experimental design, metadata discipline, uncertainty quantification, and operational integration. Deep learning can combine rheometric time series, imaging, spectroscopy, thermal data, and formulation records, yet such multimodal systems introduce trade-offs in cost, latency, interpretability, and maintainability. The paper argues for hybrid architectures that embed physical priors while retaining representation learning capacity, supported by calibrated uncertainty estimates and transparent documentation. It further considers fairness across flour varieties, regional supply chains, and small-scale bakeries, as well as policy implications for food safety, labeling, data sharing, and energy use. Case illustrations from industrial freezing, artisanal production, cloud kitchens, and humanitarian food systems demonstrate that deployment contexts shape both model requirements and social consequences. The conclusion proposes a research agenda centered on robust benchmarks, causal inference, edge-to-cloud orchestration, and equitable governance for intelligent frozen dough systems.