Abstract
Freezing can assist the distribution of low-moisture Mozzarella cheese, but the impact of freezing under industrial conditions in a pallet is not well understood. Heat transfer during the freezing and thawing of 96 blocks of 10 kg cheese was slower than observed for smaller masses of cheese (0.70–0.87 °C day⁻¹ for freezing and 0.80–6.00 °C day⁻¹ for thawing). The rate of heat transfer also differed between inner and outer blocks, particularly during thawing. Block temperature was predicted with a maximum root mean square error of 3.60 °C, using heat and mass transfer simulations. While several changes in physicochemical properties were observed, the impact on cheese functionality appeared small. Large reversible salt migration was observed by simulation, causing local concentrations of up to 33% salt in free moisture in outer blocks at the end of freezing. Intact casein was 3–4% lower after thawing compared to in refrigerated control cheese but the microstructural, textural, and functional properties were similar, except for the appearance of a greater number of calcium crystal complexes in inner blocks. The microstructural, textural, and functional properties of inner and outer blocks were also similar, despite differing rates of heat transfer. Linear regression could predict the concentration of soluble nitrogen in thawed samples using data for refrigerated samples. Machine learning methods were also applied to predict non-linear behavior while minimizing the need for experimental data. A linear multi-fidelity Gaussian process model best predicted soluble nitrogen by combining historical data from refrigerated samples with limited experimental data from thawed samples. This study increases our understanding of freezing and thawing of cheese in an industrial setting and offers tools for optimizing these processes to minimize proteolysis in order to reduce the impact on product quality.
Introduction
Freezing is a vital tool for the global distribution of low-moisture (LM) Mozzarella, extending its shelf life beyond the typical three-month window. However, most research on cheese freezing happens at the "lab scale"—small blocks frozen rapidly. In the real world, cheese moves in massive pallets, creating a significant heat transfer challenge: inner blocks freeze and thaw much slower than outer blocks. This "pallet effect" creates temperature gradients that can lead to uneven proteolysis, potentially affecting the final melt and stretch characteristics of the cheese. This study bridges the gap between lab theory and industrial reality using a "digital twin" approach, combining physical experiments with mathematical and machine learning models.
Methods and Materials
The research utilized an integrated experimental and computational approach. Two full pallets, each containing 96 vacuum-packed 10 kg blocks of Mozzarella, were monitored during a 70-day freezing cycle at -18 °C and an 18-day thawing cycle at 8 °C.
Mathematical and Digital Modeling
A 2D heat and mass transfer model was developed using an enthalpy method to track temperature and the lever rule to simulate salt migration. The model accounted for the complex layers of cardboard and plastic wrapping surrounding the cheese blocks.
To predict soluble nitrogen (a marker for proteolysis) with limited experimental data, the team employed Multi-fidelity Gaussian Process (GP) models. This "data-thrifty" approach combined a large volume of historical "low-fidelity" data from refrigerated cheese (327 data points) with a small set of "high-fidelity" data from the frozen-thawed experiment (30 data points).
Results and Discussion
Thermal Dynamics and the Pallet Effect
Industrial-scale freezing is significantly slower than lab-scale processes, with cooling rates of only 0.70–0.87 °C per day. The heat transfer simulations accurately predicted block temperatures with a maximum Root Mean Square Error (RMSE) of 3.60 °C.
A key finding was the simulation of salt migration. As ice crystals form, they reject salt into the remaining liquid moisture. Simulations showed local salt concentrations spiking to 33% in the free moisture of outer blocks. However, this process appeared largely reversible, with salt redistributing evenly after approximately 30 days of thawing.
Impact on Quality and Functionality
While frozen-thawed samples exhibited 3–4% lower intact casein than refrigerated controls, indicating accelerated protein breakdown, the final functional properties remained robust. Despite the massive temperature differences between the core and the exterior of the pallet, the stretchability and meltability were surprisingly similar. One notable difference was the higher frequency of calcium crystal complexes in inner blocks, which remained in a "warm" frozen state longer than outer blocks.
Machine Learning for Quality Prediction
The linear multi-fidelity Gaussian process model outperformed other methods, predicting soluble nitrogen levels with a Mean Absolute Percentage Error (MAPE) of 11.2% using only three high-fidelity data points. This demonstrates the potential for manufacturers to predict product quality accurately without extensive and expensive laboratory testing.
Conclusion
This study demonstrates that Mozzarella cheese is remarkably resilient to the slow freezing and thawing rates inherent in industrial-scale palletized distribution. The development of accurate thermal simulations and multi-fidelity machine learning models provides a framework for "smart" inventory management. Future work could refine these models into 3D simulations and link proteolysis predictions directly to specific culinary performance metrics like oiling-off or browning.
Summary updated