When extracting essential oils, does longer heating or more intensive treatment always produce better results? Once the properties of the recovered oil are considered alongside its quantity, choosing conditions is not straightforward.
A study published by Parvej Hasan Jon and colleagues in 2026 used two mathematical models to search for conditions for a method in which leaves are pretreated with ultrasound and then distilled with water. One model was an artificial neural network, a technique used in AI. The aim was to determine whether essential-oil yield and properties assessed by chemical tests could be adjusted together.
Varying water quantity, ultrasound, and distillation time
The raw material was Indian bay leaf, Cinnamomum tamala, purchased at a market in Bangladesh. These aromatic culinary leaves come from a different plant from the bay laurel commonly known as bay leaf.
Water was added to dried, chopped leaves, which were exposed to ultrasound before the oil was recovered by hydrodistillation. Four conditions were varied: the amount of water relative to the leaves, ultrasound power, sonication time, and distillation time. The experimental design comprised 30 runs, including repeated reference conditions. The original paper states that each experiment was performed three times.
The researchers measured oil yield relative to dry leaf weight, an indicator of total phenolic content, and response to a reagent called DPPH. The latter two assess chemical properties in vitro. They are not measurements of antioxidant effects within the human body or of health benefits.
Learning relationships between conditions and outcomes
The first method compared was response surface methodology, which represents the relationship between conditions and outcomes as a surface. It approximates that relationship mathematically, including, for example, how the effect of increasing ultrasound power depends on treatment duration.
The second, an artificial neural network, learns complex relationships between conditions and outcomes from experimental data. In this study it was combined with a genetic algorithm, which searches for conditions while retaining promising candidates. Both approaches select favorable conditions across all three measured indicators rather than maximizing yield alone.
Experimental yields ranged from about 0.54% to 1.76% depending on the conditions. Model analysis did not show a consistently improving relationship as water quantity or treatment intensity increased. The authors consider both easier release of compounds and possible losses under excessive treatment. However, the experiments did not directly track leaf-cell disruption or reactions that degrade compounds.
Candidate conditions produced yields close to predictions
The neural-network approach selected a combination of about 9.9 milliliters of water per gram of dry leaves, ultrasound at about 142 watts for about 27 minutes, and subsequent distillation for about 124 minutes. Here, distillation time was counted from the appearance of the first drop in the condenser. It does not mean the entire process, including pretreatment and preliminary heating, ended in 124 minutes.
Under these conditions, the predicted yield was 1.68%, while the experimental mean was 1.67%. The total-phenolic-content indicator and DPPH response were also close to the predictions. Conditions selected by response surface methodology likewise gave a measured yield of 1.64%, close to the predicted 1.66%.
The authors report that the neural-network approach had higher predictive accuracy for this dataset. However, the difference between 1.67% and 1.64% does not establish that AI invariably produces more oil. The selected yield was also below the experimental maximum of 1.76% because conditions were chosen using the other two indicators as well as quantity.
GC–MS analysis identified linalool, eugenol, and cinnamaldehyde as major components of the recovered oil. The study did not, however, use human smell comparisons to determine how well the original aroma was preserved or whether its pleasantness improved.
Understanding why conditions were selected was another goal
Rather than simply asking a model for an answer, the study also analyzed how water quantity and ultrasound conditions contributed to its predictions. By examining predictions as individual conditions changed and assessing input importance, the researchers sought to explain the meaning of the selected conditions. The authors emphasize combining predictive ability with clues for understanding the process.
However, the data came from limited raw material and laboratory-scale equipment. Model validation was also based on the same experimental design; equivalent accuracy was not established for leaves from other origins or for factory equipment. The design did not include a direct comparison with distillation without ultrasound using the same raw material and apparatus.
Summary
The study demonstrates a way to search for extraction conditions from experimental data while considering essential-oil quantity and chemical indicators together. The authors conclude that a neural network can support this search and is worth combining with analyses that explain how conditions influence outcomes.
Application to fragrance or food-ingredient manufacturing requires validation with raw materials from different origins and with larger equipment. Assessing energy efficiency, costs, and aroma quality as well would help bring favorable laboratory conditions closer to practical manufacturing conditions.
References
Jon, P. H., Shourove, J. H., Ali, M. K., Rahman, O., Thakur, M. U., & Islam, G. M. R. (2026). Optimizing sonication-assisted hydrodistillation of Cinnamomum tamala essential oil using response surface methodology and artificial neural network modeling. Scientific Reports, 16, 14107. DOI: 10.1038/s41598-026-42869-2 / Full text.
The authors declared that they received no funding for the research, writing, or publication, and reported no competing interests. Original article: © 2026 The Author(s), CC BY 4.0.
