Research · Deliverable D2
Experimental version of artificial intelligence software for the modelling and optimisation of dough-based bakery product recipes
| Work package | WP3 |
|---|---|
| Delivery date | 31.12.2025 |
| Status | Completed |
| Web link | smartbakeai.com/research/software |
D2 documents the transition of the project from the scientific analysis and the initial design of WP2 to a functional experimental technological implementation: the selected algorithmic models were integrated into a web application that accepts parameters and produces experimental predictions. This is a translation. The authoritative version of this deliverable is the Greek one.
Nature of the software. The implementation was not a finished commercial product, but experimental software for assessing the technological feasibility of the models developed. The values produced are experimental estimates and do not substitute for trial production.
1
Preparation of the data
The development of D2 was based on the experimental data produced during WP1 and initially assessed and organised during WP2. Before their introduction into the algorithmic models, procedures of cleaning, normalisation, transformation and organisation were carried out, in order to remove or manage inconsistent values, to homogenise different variables and to create data sets suitable for training and evaluation.
The process was critical, because the quality of the models’ results depends directly on the quality, the consistency and the structure of the input data.
Appropriate input features were then developed, so that the raw records of the experimental trials could be converted into a form usable by the algorithms. In parallel, the quality characteristics used as output variables or as targets for prediction and optimisation were determined.
2
Comparative investigation of algorithmic approaches
A core object of the research activity was the comparative investigation of different algorithmic approaches. No single model was adopted in advance as a given solution. On the contrary, different architectures and parameterisations were examined, in order to assess which approach could capture more effectively the complex relationships between the characteristics of the recipe, the production conditions and the final result.
Development cycle per approach
- Initial parameterisation
- Training
- Experimental evaluation
- Recording of deviations and limitations
- Retraining
- Optimisation of the parameters
The results were not considered static. Where deviations, limited accuracy or instability were observed, retraining and improvement of the selected models were carried out, as well as redefinition of parameters and input features. This iterative process allowed the rejection of approaches that did not show adequate behaviour.
3
Evaluation criteria
The effectiveness of the models was not examined against a single measurement, but against a set of criteria. The aim was to establish not only whether a model could be fitted to the existing data, but whether it showed adequate behaviour so as to form a basis for further research and technological development.
Accuracy
How close the predictions are to the actual measurements.
Stability
How consistent the results are across repeated runs.
Reliability
The behaviour of the model in different cases.
Generalisability
The performance on data not used in the initial training.
Interpretability
The extent to which it can be explained which factors led to a result.
4
The experimental web application
A functional application was developed, hosted in a suitable computing environment, through which the user enters selected parameters relating to the recipe or to the production conditions and receives corresponding experimental predictions. In this way a transition was made from the purely computational training of the models to a functional research prototype.
Input parameters
- Flour type and protein content
- Hydration (water on flour)
- Fat, sugar, salt, yeast
- Fermentation temperature and time
- Baking temperature and time
Experimental predictions
- Specific volume (mL/g)
- Crumb moisture (%)
- Crumb hardness (N)
- Crust lightness (L* index)
- Estimated shelf life (days)
Functions that document the research character
Extrapolation check
When a parameter falls outside the experimental range of the training data, the application states this explicitly and warns that reliability decreases.
Sensitivity analysis
The contribution of each parameter to the current prediction is displayed — an element of interpretability, not only of prediction.
Experimental uncertainty range
Each prediction is accompanied by a range and not by a single value, so that it is not presented as a certainty.
5
Deviations, technical limitations and unsuccessful approaches
Particular emphasis was placed on recording the deviations, the technical limitations and the unsuccessful approaches. This record constitutes a critical scientific outcome of D2, as it captures not only what worked, but also in which cases the models showed inadequate accuracy, instability or limited generalisability.
Documented limitations of the implementation
- The predictions are valid within the experimental range of the training data; outside it this is extrapolation.
- The model does not capture the quality variability of raw materials per supplier batch.
- The sensory indicators are indirect estimates and do not substitute for organoleptic evaluation.
- Generalisation to categories of dough-based bakery products outside the training set remains an open research question.
- The categories under-represented in the experimental set are flagged explicitly within the application.
MAINTENANCE NOTE — place for the comparative evaluation table of the algorithmic approaches (model, accuracy metrics, stability, generalisation, acceptance or rejection decision) and for the training charts. They are drawn from the subcontractor’s technical documentation; they are not filled in by estimation.
6
Verifiability
D2 provides the possibility of direct verification of the physical object implemented: it is possible to establish the existence and operation of the experimental application through the electronic link, to examine the ability to enter selected parameters and produce experimental predictions, and to relate the operation of the application to the technical documentation, the experimental data and the training and evaluation results of the models.
The completion and delivery of D2 on 31 December 2025 constitutes the principal technological milestone of the 2025 financial year and demonstrates the completion of WP3 and of the experimental software development foreseen for that tax year.