Page 56 - FoodFocusThailand No.246 October 2026

SOURCE OF ENGINEER

A study by Jiang et al. (2025) comparing BO with become more reliable. This is particularly important when the
RSM for high-moisture extrusion of plant-based proteins reference model is no longer representative of the current product.
for meat analogue production reported that BO achieved The study by Pei et al. (2025), which developed a ginger-
better predictive performance with fewer experiments. The drying system combining far-infrared radiation with hot air and
study also found that tensile-strength data could further incorporated data from multiple sensors, including imaging and
improve model performance. However, these findings odor-related signals, demonstrates an interesting direction for
represent evidence from the specific system investigated using quality-related data to support decisions. Nevertheless,
and should not be interpreted as evidence that BO is soft sensors should be regarded as decision-support tools
superior to RSM for all processes. rather than replacements for verification testing, particularly for
A practical approach is to combine shop-floor moisture, which is not equivalent to verifying water activity or
knowledge and DOE to identify variables and ranges, then microbiological safety.
apply BO within defined safety and machine constraints
in advance. From Digital Twin to Predictive Control
A valuable digital twin is not just a 3D simulation or dashboard,
Hybrid Modeling: Integrating Physics but a model linked to equipment and continuously updated with
and Machine Learning for Practical, real-time data. Investors must assess actual functionality, from
Real-World Models status monitoring and prediction to sending commands back to
Food raw materials vary by variety, season, moisture, and control the process, as well as the level at which each capability
ripeness. Data-only models may perform well initially but has been tested and validated (Abdurrahman & Ferrari, 2025).
degrade under new conditions. Physics-based models With a reliable dynamic model, Model Predictive Control (MPC)
may be too computationally heavy for real-time use. forecasts process behavior, determines adjustments, implements
Hybrid Models offer an interesting approach by actions, and recalculates strategies as new data arrive. For
combining physics knowledge with machine learning, for example, MPC can be used to adjust temperature or conveyor
example, using mass and energy balances as structure, speed toward a target quality outcome while remaining within
while machine learning learns variable parameters that constraints such as equipment capacity and allowable process-
vary with conditions or residual errors that the model change rates. Though MPC has existed for decades, model and
cannot explain. Starting with understandable, validated data quality continue to evolve (Vassiliadis et al., 2020).
models is often more practical than starting with complex For existing plants, starting in Advisory Mode may be more
ones. appropriate: the system provides recommended settings to the
Physics-Informed Neural Networks (PINNs) embed machine operator and records the result before moving toward fully
physical equations and boundary conditions into training, automated control. The system should return to a validated recipe
reducing predictions that contradict known behavior. when sensors fail, data are missing, or the model encounters
However, equations alone do not guarantee compliance conditions outside its validated operating range. Meanwhile, safety
with physical laws; therefore, deviations arising from both systems and interlocks must remain independent of AI.
the equations and the data still need to be evaluated
separately (Karniadakis et al., 2021). Optimizing Robustness,
Wijerathne et al. (2025) reviewed PINNs in food drying, Not Just Laboratory-Optimal Conditions
highlighting their potential to connect data with multiscale Operating conditions based on average raw material properties
changes from cellular to product level. Limitations remain may lie too close to specification limits, making consistent
in data availability and model complexity, so PINNs are production performance difficult. When raw materials vary or the
an emerging technology rather than a ready-to-deploy model contains prediction errors, the product may fall outside the
system. required specifications.
Robust Optimization accounts for raw material variability
Soft Sensor: Turning Process Data into and model uncertainty. Meanwhile, Stochastic Optimization
Real-Time Quality Assessment incorporates probability distributions when sufficient data are
A challenge in food plants is timely quality measurement. available. The key idea is to shift the question from “What is
Temperature and process conditions are monitored the best value?” to “How should the conditions be selected to
continuously, but moisture, texture, or compounds may achieve good outcomes under realistic variability?” (Vassiliadis
only be known after production and lab analysis. et al., 2020).
Soft Sensors help bridge this gap by integrating data Advanced Process Optimization should begin with problems
from measurement instruments with models to estimate whose value can be measured in practical terms, with production
quality attributes that cannot be measured directly or data correctly connected and models evaluated using data that
are difficult to measure during processing. For example, were not used during training. This helps ensure that the results
spectral data, images, weight, and environmental reflect real-world operating conditions. Ultimately, the value of
conditions can be combined to estimate moisture content. Process Optimization does not lie in having the most sophisticated
Ashtiani and Martynenko (2025) analyzed the AI, but in making better and auditable decisions by combining the
use of soft sensors in food drying and highlighted capabilities of AI with the constraints and practical experience of
key challenges, including calibration and maintaining plant personnel, thereby improving outcomes across the entire
prediction accuracy when the product or environmental system.
conditions change. Therefore, having more frequent
data does not automatically mean that the estimates will

56 FOOD FOCUS THAILAND OCT 2026 More Information Service Info C008

22/9/2569 BE 19:16
52-56_Source of Engineer_Ravis.indd 56
52-56_Source of Engineer_Ravis.indd 56 22/9/2569 BE 19:16