Page 55 - FoodFocusThailand No.246 October 2026

SOURCE
SOURCE OF ENGINEER OF ENGINEER

OPTIMIZING FOOD PRODUCTION
PROCESSES USING MODELING AND

MACHINE LEARNING

In food manufacturing, the “best” machine settings do not always yield the “best” overall outcomes.
A dryer may consume less energy per hour yet increase costs per kilogram because drying takes longer
and more products are downgraded. Similarly, increasing the speed of an extruder may boost production
output at the machine level, but the subsequent cooling and packaging stages may be unable to handle
the increased volume.

Numbers that look favorable at one point in the process lies in combining these principles with continuously changing
may not reflect the overall efficiency of the system and may data, enabling manufacturers to select operating conditions,
conceal losses occurring at other stages. Thus, process design experiments, and make decisions at the right time
optimization should not ask, “What settings should (Erdoğdu, 2009; Vassiliadis et al., 2020).
machines operate at?” but rather, “What outcomes must the
factory achieve, and under what constraints?” From DOE and RSM
to Bayesian Optimization (BO)
From Machine Optimization to System-wide Plants that already use Design of Experiments (DOE) and
Optimization Response Surface Methodology (RSM) have a foundation for
Process optimization begins by clearly defining three advanced optimization. These methods reveal interactions
factors: decision variables, objective function, and system among variables; for example, temperature effects may
constraints. The objective is not simply reducing energy or depend on raw material moisture, rather than acting as an
maximizing production rate but maximizing incremental profit independent effect of each variable. However, a limitation is
per production hour. This requires revenue from products that experiments in plants are costly, involving raw materials,
meeting quality requirements, minus costs of raw materials, machine downtime, and waiting time for analysis results.
energy, waste, cleaning, and other expenses, while meeting Predetermined plans can be expensive.
safety requirements. Bayesian Optimization (BO) offers an alternative
Therefore, reducing electricity costs alone may not approach. BO builds a surrogate model predicting outcomes
create value if it results in increased waste. Likewise, and quantifying uncertainty, then selects the next experiment
producing more products only creates real value when there by balancing promising regions with unexplored ones. As
is sufficient market demand to support them. This concept new data become available, the model is updated and the
shifts optimization from focusing on the efficiency of individual next experimental point is selected again. This allows the
machines to evaluating the overall process outcome—from experimental process to learn iteratively rather than requiring
raw materials to finished products that meet quality standards all experimental points to be fixed in advance.
and deliver real value. The advancement of technology today

OCT 2026 FOOD FOCUS THAILAND 55

22/9/2569 BE 19:16
52-56_Source of Engineer_Ravis.indd 55
52-56_Source of Engineer_Ravis.indd 55 22/9/2569 BE 19:16