Research · Deliverable D1
Online documentation of the research methodology and of the initial experimental results
| Work package | WP2 |
|---|---|
| Delivery date | 30.11.2025 |
| Status | Completed |
| Web link | smartbakeai.com/research/documentation |
D1 records, in a systematic and verifiable manner, the transition from the primary experimental data produced in WP1 towards their organised scientific and computational investigation. This is a translation. The authoritative version of this deliverable is the Greek one.
Role of the deliverable. D1 is not a simple descriptive report, but the scientific and methodological foundation of the further experimental development. It documents that, before the software was developed, an organised research process was carried out to assess the real data, to determine the research variables and uncertainties, to formulate and initially examine research hypotheses and to design the algorithmic approach.
1
Receipt, recording and initial assessment of the experimental data
This documents the receipt, recording, classification and initial assessment of the experimental data and of the available measurements that had been produced during the execution of WP1. The data are linked to the different experimental recipes, the quantities and proportions of the raw materials used, the types of the individual ingredients, the baking temperatures and times, the production and storage conditions and the observed quality characteristics of the final test specimens.
At this stage it was examined whether the available records were sufficiently complete, consistent and suitable to be able to serve as a reliable basis for the next phase of algorithmic processing.
Assessment criteria
2
Quality control and initial preparation of the data
The completeness of the individual records, the consistency between the different experimental trials, the presence of incomplete, extreme or unusable values were examined, as well as the need to homogenise the format of the data before their introduction into the computational models.
The methodology of cleaning, normalisation, transformation and organisation to be followed in the subsequent stages was determined, so that the data would acquire a structure suitable for the training and the experimental evaluation of different algorithmic approaches.
The quality of the data was a critical factor: incomplete or inconsistent records could significantly affect the training and the evaluation of the experimental models.
3
Input and output variables of the research problem
A significant part of D1 concerns the determination and the categorisation of the input and output variables. The process was critical, because it determined which elements of the real production process would be used for the creation of the experimental algorithmic models.
| Category | Variables examined |
|---|---|
| Recipe composition | Dosage and percentages of the raw materials· proportions of the individual ingredients· types of ingredients used· type of flour· fats |
| Production conditions | Baking temperatures and times· other production conditions |
| Storage conditions | Conditions and duration of storage of the test specimens |
| Quality characteristics (outputs) | Texture· volume· colour· thermal behaviour· appearance of spoilage· storage life· overall quality of the product |
MAINTENANCE NOTE — a place is reserved here for the detailed table of variables with the real value ranges and units of each parameter, as recorded in WP1. To be completed by the company before publication.
4
Identification and documentation of the initial uncertainties
It was examined whether it is feasible to capture with sufficient accuracy, by means of artificial intelligence models, the complex and possibly non-linear relationships between the composition of a recipe, the production conditions and the final characteristics of the product.
It was not taken for granted in advance that the experimental data would be sufficient, nor that any particular algorithmic approach could produce reliable and generalisable results. For this reason, specific research hypotheses were formulated and the process through which these would be examined experimentally was defined.
Data sufficiency
Can the available data capture the composition – quality relationships?
Choice of approach
Which input features, which pre-processing and which algorithmic approaches give the most reliable results?
Stability and generalisation
Are the predictions stable? Do they generalise to different recipes? Are they interpretable?
5
Initial technological design of the algorithmic models
The basic logic of the computational process was determined: the conversion of the experimental records into suitable input features, the formation of data sets suitable for training and evaluation and the investigation of different artificial intelligence and machine learning models.
The selection of the final approaches was not made in advance. A comparative and iterative process was designed, so that different models could be trained, evaluated and, where required, rejected or improved on the basis of the real experimental results.
- Conversion of records into input features From the primary data of WP1 into a form usable by algorithms.
- Formation of training and evaluation sets Separation of data so that generalisability can be measured.
- Comparative investigation of models Different architectures and parameterisations, without a predetermined solution.
- Iterative evaluation, rejection or improvement On the criterion of the real experimental results.
6
First version of the algorithmic model and initial experimental evaluation
Within the framework of WP2, the training and the initial experimental evaluation of the first version of the algorithmic model began. The work had an exploratory character and did not aim at producing a final or commercially mature model, but at the first practical verification of the research hypothesis that the available variables can be used to derive computationally usable relationships between the composition of the recipes, the production conditions and the quality characteristics.
The results of the initial evaluation were used as the basis for the redesign, the further training and the comparative evaluation of different algorithmic approaches carried out in WP3.
MAINTENANCE NOTE — place for the numerical results of the initial evaluation (metrics, convergence diagrams, error tables). They are not filled in arbitrarily: they are drawn from the technical documentation of the subcontractor.
7
Initial deviations, limitations and unsuccessful approaches
The deliverable records the deviations, the limitations and the problems that were identified during the processing of the data and the first algorithmic trials.
The recording of the unsuccessful or insufficiently effective approaches constitutes an essential part of the research process, since it allows the documentation of the real scientific and technological uncertainty and distinguishes the project from a simple application of known or predetermined techniques.
These results fed directly into WP3, in which the extensive cycles of training, retraining, parameterisation and comparative evaluation of the models were carried out.
8
Verifiability of the physical scope of work
D1 functions as complete evidence of the execution of WP2, because it allows the path from the primary experimental data through to the initial computational design and the first experimental results to be established. The documentation links the data of WP1 with the activities of WP2 and presents the analysis methodology, the research hypotheses, the variables that were selected, the initial architecture of the models and the first results of the algorithmic investigation.
The completion of D1 on 30 November 2025 constituted the necessary technical milestone for the transition to WP3 and the development of the experimental version of artificial intelligence software delivered as D2 on 31.12.2025.