Anshika Jain
Fermentation is a biological process in which microorganisms transform sugars, proteins, fats, and other food components into new compounds. These transformations help determine a food’s flavour, aroma, acidity, texture, nutritional characteristics, and keeping quality. From yoghurt and sourdough to fermented vegetables, coffee, and traditional batters, the final product depends on how microbial activity develops over time.
Yet fermentation is not a perfectly predictable process. Two batches prepared with the same ingredients can develop differently because of variations in temperature, moisture, microbial populations, raw material composition, and handling. A fermented food may become more acidic than expected, develop an unusual aroma, or take longer to reach its intended endpoint.
Traditionally, producers assess fermentation through experience, periodic sampling, sensory observations, and measurements such as pH, temperature, and sugar concentration. These methods remain valuable, but some provide only a partial picture of the biological changes taking place inside the food.
Fermentation trajectory AI offers a different approach. Instead of examining only the current state of a fermenting product, an AI model attempts to predict how the process will evolve. By analysing measurements collected throughout fermentation, it could estimate future acidity, changes in microbial activity, the formation of selected metabolites, and the likelihood of reaching a desired quality target.
Research on machine learning and intelligent fermentation monitoring is already exploring the prediction of fermentation outcomes using sensor measurements, chemical data, microbial information, and time-series models. A study published in 2023 introduced a data-driven platform for predicting microbial transformations in fermented foods, while newer research has investigated real-time prediction and control in products such as yoghurt.
The larger opportunity is to make fermentation more predictive without removing the biological complexity that gives fermented foods their distinctive character.
What Is a Fermentation Trajectory?
A fermentation trajectory is the pattern of changes a food undergoes as fermentation progresses. Rather than describing a product through one measurement, it considers how its properties change over a sequence of time points.
For example, the trajectory of a fermenting vegetable mixture might include a gradual decline in pH, changes in organic acid concentrations, shifts in microbial populations, and the development of characteristic aromas. A yoghurt trajectory might include acidification, changes in the structure of milk proteins, and the gradual formation of the desired texture.
These changes are connected. Microorganisms consume available nutrients and produce metabolites, while changes in acidity and the surrounding environment influence which microorganisms remain active. The result is a dynamic system in which the current state is shaped by the conditions and biological activity that came before it.
A trajectory-based model uses this history to estimate what may happen next. If a batch is acidifying more slowly than expected, for example, the model could estimate whether it is likely to reach its target acidity within the planned production period. If metabolite measurements show an unusual pattern, the model could flag a potential deviation for further investigation.
This is different from simply predicting the final quality of a food from its ingredients. A trajectory model incorporates the progression of the process, making it possible to update predictions as new information becomes available.
Why Fermentation Is Difficult to Predict
Microbial Communities Change Over Time
Fermentation often involves communities of microorganisms rather than a single organism acting independently. Bacteria, yeasts, and other microorganisms may compete for nutrients, produce compounds that affect one another, and respond differently as the environment changes.
In spontaneous fermentation, the starting microbial community can vary considerably between batches. Even when a starter culture is used, its activity may be influenced by storage, inoculation conditions, raw material composition, and temperature.
These interactions help explain why fermentation can follow a different trajectory even when the recipe remains unchanged. Reviews of AI-driven fermented-food research describe how machine learning can help analyse microbial communities, predict metabolic functions, and support the design of starter cultures, while also identifying limited data standardisation as a significant challenge.
Chemical Transformations Are Nonlinear
Fermentation does not always proceed at a constant rate. Microbial activity may increase during an initial growth phase, slow as nutrients become limited, and change again as acidity or other environmental conditions alter the community.
The relationship between temperature and fermentation speed is also complex. A change in temperature may accelerate some biological reactions within a suitable range, but conditions outside an organism’s preferred range can reduce activity or damage the culture. Different microorganisms may respond differently to the same conditions.
A simple linear model may struggle to represent these interactions. AI models can potentially learn nonlinear relationships from sufficiently diverse datasets, but they need careful validation to ensure that apparent patterns reflect real biological behaviour rather than accidental correlations.
The Most Important Biological Changes Are Not Always Directly Measurable
Temperature and pH sensors can provide frequent measurements, but many biologically meaningful properties are harder to measure continuously. The concentration of a particular metabolite, the activity of an enzyme, or the changing composition of a microbial community may require specialised laboratory analysis.
Researchers therefore investigate indirect measurements, including spectroscopy, volatile-compound analysis, imaging, and other sensor signals. Computational models can use these signals to estimate properties that would otherwise require slower or more expensive tests. This approach is often described as soft sensing.
The challenge is that an indirect signal may correspond to several different biological processes. A reliable model must establish how strongly the signal relates to the property being predicted and whether that relationship remains valid across different batches and conditions.
How Fermentation Trajectory AI Works
Building a Time-Based Dataset
The first step is collecting measurements throughout the fermentation process. A useful dataset might record temperature, pH, acidity, sugar concentration, moisture, and the time elapsed since fermentation began. Depending on the food, researchers could also include microbial counts, sequencing data, volatile compounds, and measurements of selected metabolites.
Each observation must be connected to the conditions under which it was collected. A pH measurement taken at the beginning of fermentation has a different meaning from the same measurement taken after several days. Similarly, a metabolite concentration cannot be interpreted properly without considering the food’s composition and previous processing history.
The dataset should include both successful and unsuccessful batches. If an AI system is trained only on ideal fermentation outcomes, it may learn to recognise the expected trajectory but fail to identify unusual behaviour. Data from different raw material batches, starter cultures, temperatures, and production conditions can help the model understand the range of variation it may encounter.
Learning From Fermentation Curves
Once the data has been collected, a machine learning model can learn patterns in how fermentation variables evolve. Traditional regression models may be suitable when the relationships are relatively simple and the dataset is limited. More complex time-series approaches, including recurrent neural networks and transformer-based architectures, can be considered when large, well-structured datasets are available.
The appropriate model depends on the food and the intended prediction. A producer may need to estimate the time required to reach a target pH, while a researcher may want to predict the concentration of a particular organic acid or identify a developing deviation from a normal fermentation curve.
For example, a model trained on yoghurt fermentation data could analyse the current pH, temperature history, and rate of acidification to estimate when the batch may reach its target acidity. Research published in 2025 describes an AI-based approach to real-time yoghurt fermentation monitoring and predictive control, illustrating how current measurements can support forecasts of future process conditions.
Such predictions are conditional rather than guaranteed. If the milk composition changes, the starter culture behaves differently, or the sensor becomes inaccurate, the model’s forecast may no longer be reliable. Continuous updating and independent validation are therefore essential.
Combining AI With Biological Knowledge
A purely data-driven model may discover useful patterns without explaining why they occur. In fermentation, this can be a serious limitation because microorganisms follow biological and chemical constraints that should not be ignored.
A more robust approach combines AI with mechanistic models describing microbial growth, nutrient consumption, metabolite production, and environmental changes. The mechanistic component represents established scientific relationships, while machine learning helps account for complex patterns that are difficult to model precisely.
Research on hybrid fermentation modelling suggests that combining knowledge-based predictions with machine learning can improve predictions in some settings, particularly when data is limited or the process is biologically complex. However, performance still depends on the quality of the data and whether the model is tested under conditions that differ from its training examples.
This combination is particularly valuable when a model needs to estimate not only what is likely to happen, but also whether that prediction is consistent with the known behaviour of the fermentation system.
What Biological Properties Could AI Predict Before Fermentation Ends?
Acidity and the Time to Completion
One of the most practical targets is the rate at which acidity develops. In many fermented foods, acidification is closely associated with microbial activity and contributes to flavour, texture, and preservation. Measuring pH over time provides a useful indicator of the process, although pH and total acidity are different measurements and should not be treated as interchangeable.
An AI model could estimate the likely time required to reach a specified pH or acidity target by analysing the current measurements and the trajectory so far. It could also flag a batch that is acidifying much more slowly or quickly than expected.
This would allow producers to plan production schedules more accurately and investigate deviations earlier. However, reaching a target pH does not automatically establish that a product is safe, fully fermented, or of the desired sensory quality. Those conclusions require product-specific criteria and validated procedures.
Metabolite Formation and Flavour Development
Fermentation produces a range of metabolites that influence aroma, taste, and texture. Depending on the food and microorganisms involved, these may include lactic acid, acetic acid, alcohols, esters, and other volatile or non-volatile compounds.
AI could potentially predict how the concentration of selected metabolites is likely to change during fermentation. This is an active area of research: the FFExplorer platform, described in a 2023 study, used microbial enzyme and sequence information to predict molecular biotransformations in fermented foods and investigate how microorganisms might transform food compounds.
Such predictions could help researchers understand why certain starter cultures produce particular flavour profiles or why a batch develops an unexpected aroma. In commercial settings, the same approach might help identify fermentation conditions associated with a more consistent product.
Nevertheless, flavour is not determined by a single compound. It emerges from mixtures of chemicals, their concentrations, interactions, and the sensory properties of the food matrix. Predicting one metabolite accurately is not the same as predicting the complete flavour experience.
Microbial Succession and Community Behaviour
In mixed-culture and spontaneous fermentations, the composition of the microbial community can change throughout the process. Some organisms may dominate early, while others become more important as nutrients and environmental conditions change.
A model could combine microbial sequencing data, temperature, pH, and metabolite measurements to estimate how the community may develop. It might identify patterns associated with a successful fermentation or detect an unusual shift that warrants further testing.
AI-driven research on fermented foods is increasingly exploring the relationship between microbial communities, metabolic pathways, starter culture selection, and flavour formation. Yet microbial abundance alone does not always reveal what an organism is doing. The presence of a microorganism does not necessarily prove that it is actively producing a particular compound or causing a specific outcome.
For this reason, microbial trajectory predictions are most useful when supported by complementary chemical and functional measurements.
Nutritional Characteristics and Potential Bioactive Compounds
Fermentation can alter the chemical composition of food, including the availability of selected nutrients and the concentration of certain bioactive compounds. Depending on the ingredients and microorganisms, fermentation may reduce some antinutritional factors, transform plant compounds, or change the digestibility of particular food components.
AI could help estimate these changes by linking the initial composition of the ingredients with microbial activity and measured metabolite trajectories. For example, a model might investigate whether a particular fermentation process is associated with a reduction in phytate or a change in the profile of selected phenolic compounds.
However, the direction and magnitude of these effects depend on the specific food and fermentation conditions. A model cannot assume that longer fermentation automatically makes a food more nutritious or that the formation of a compound establishes a health benefit. Nutritional claims require direct analytical evidence and, where relevant, appropriate biological studies.
From Prediction to Real-Time Fermentation Management
The greatest practical advantage of trajectory AI is the possibility of acting on a forecast before the process reaches its endpoint. A conventional monitoring system reports the current state of a batch. A predictive system attempts to estimate what that state is likely to become if conditions remain unchanged.
For example, if a yoghurt batch is predicted to reach its target acidity earlier than expected, a validated control system could alert an operator to prepare for cooling. If a fermentation trajectory deviates from the expected pattern, the system could recommend checking the temperature, starter culture, raw materials, or sensor performance.
A more advanced arrangement could connect the predictive model to a digital twin, a computational representation of the physical fermentation process. The digital twin would receive updated measurements, simulate possible future trajectories, and help compare alternative operating conditions before a change is made. Research into digital twins and intelligent fermentation control identifies this integration of sensing, prediction, and process optimisation as a promising direction, although industrial reliability and transferability remain important challenges.
The system should not automatically change production conditions simply because a prediction suggests that a batch is developing unusually. Any intervention must be appropriate for the food, validated for its process, and consistent with food-safety requirements. AI is most useful when its forecasts support established quality-control procedures rather than replace them.
Limitations: Why Predicting Fermentation Is Still Difficult
The first limitation is biological variability. Natural ingredients differ in their composition and microbial populations, while even commercial starter cultures may perform differently under changing conditions. A model trained on one facility or product may not generalise well to another.
The second limitation is measurement quality. Sensors can drift, become fouled, or provide indirect signals that are difficult to interpret. Missing measurements and inconsistent sampling schedules can also weaken a model’s predictions. Research on intelligent fermentation monitoring highlights sensor reliability, data integration, and model generalisation as continuing challenges.
A third challenge is predicting rare but important events. An AI system may perform well on ordinary batches but fail to recognise an unusual contamination event or a previously unseen failure mode. Predicting fermentation speed or flavour development is not equivalent to detecting every food-safety hazard.
Finally, prediction accuracy must be tested beyond the data used to develop the model. Researchers need to assess performance on independent batches, different ingredient sources, and realistic operating conditions. They should also report uncertainty so that operators can recognise when a forecast is unreliable.
Conclusion: Understanding Fermentation Before the Final Product Exists
Fermentation trajectory AI offers a promising way to study food as a changing biological system rather than a finished product with fixed characteristics. By learning from time-based measurements, microbial data, chemical analyses, and processing conditions, AI models could estimate future acidity, metabolite formation, microbial succession, and selected quality attributes before fermentation is complete.
The approach could help producers improve batch consistency, identify deviations earlier, and plan production more effectively. It could also support research into starter cultures and the biological transformations that give fermented foods their distinctive characteristics.
However, these possibilities depend on high-quality datasets, scientifically informed modelling, and validation across different products and production environments. A prediction of future acidity cannot establish food safety, and a forecast of metabolite formation cannot fully capture flavour or nutritional value.
The most useful future systems will combine AI with food microbiology, chemical analysis, real-time sensing, and carefully validated process controls. Their purpose will not be to eliminate the uncertainty inherent in fermentation, but to understand it better and make informed decisions while the process is still unfolding.