Method and data processing device for creating a formulation for a target aroma profile
A computer-implemented algorithm efficiently determines a recipe for a target aroma profile by analyzing sensory and instrumental data, addressing the challenges of aroma interaction complexity and incomplete analytical data, enabling automated aroma creation with desired sensory replication and specific product constraints.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- DOHLER GMBH
- Filing Date
- 2024-12-20
- Publication Date
- 2026-06-25
Smart Images
Figure EP2024087950_25062026_PF_FP_ABST
Abstract
Description
[0001] P2024, 0478 WO E December 20, 2024
[0002] Description
[0003] METHOD FOR CREATING A TARGET FLAVOR PROFILE, DATA PROCESSING DEVICE FOR EXECUTING THE METHOD AND SYSTEM INDICATING THE DATA PROCESSING DEVICE
[0004] The present invention relates to a method for creating a formulation for a target aroma profile. An aroma profile refers to the properties of a product that are perceived through sensory impressions, in particular through smell and taste. An aroma profile is also referred to as the organoleptic quality of a product.
[0005] For example, an aroma profile is determined by tasting and describing the properties such as smell, taste and mouthfeel.
[0006] An aroma profile can be characterized by olfactory (smell), gustatory (taste), and / or other sensory perceptions. For example, chemesthetic or trigeminal perceptions, such as...
[0007] Feelings of warmth, cold, pain and touch are incorporated into the aroma.
[0008] One of the most important tasks in aroma development is creating an aroma using a target aroma profile. For example, an aroma should replicate the taste and smell of an apple. The created aroma can then be used to flavor a product, such as a foodstuff. It can also be an aroma profile for a flavoring preparation. An aroma is, for example, a so-called B2B (business-to-business) product. Typically P2024, 0478 WO E December 20, 2024
[0009] Such B2B products contain 1 to approximately 200 raw materials in varying proportions.
[0010] The product to be flavored can be, for example, an edible foodstuff or a fragrance. Specifically, it can be a foodstuff, a luxury item, or a cosmetic product. It can also be another type of product, such as a consumer good. It may also be intended that a target aroma profile of a product, such as a foodstuff or a perfume, is directly replicated for an end consumer.
[0011] Flavor development often involves an instrumental or human sensory analysis of a product's target aroma profile and the development of a flavor formulation that replicates this profile as closely as possible. The formulation of the target aroma profile itself is not known. The goal of flavor development is to achieve optimal sensory matching between the product's target aroma profile and the created flavor.
[0012] A specialist, such as a flavorist or perfumer, uses their experience to determine a formula that replicates the target aroma profile. The recombined product is then sensorially evaluated, and the formula is adjusted using flavorist knowledge and experience. To achieve an optimal result, a formula is iteratively created, the aroma is tasted, and the formula is subsequently adjusted again. This iterative process is called "Targeted Creation." The best or sufficiently good matching variants of the formula, i.e., P2024, 0478 WO E 20 December 2024
[0013] 3
[0014] Those recipes that best replicate the aroma profile of the target aroma are called a "match".
[0015] The iterative process is time-consuming and does not always lead to an optimal result, as aroma development is highly complex and a complete understanding of the mechanisms of action and interactions of aromas is not expected in the coming years. Furthermore, the perception of aromas involves numerous interactions that cannot be predicted for all recipes.
[0016] Furthermore, according to current scientific consensus, the olfactory space contains well over 10,000 molecules, resulting in an immense variety of aroma formulation possibilities.
[0017] A process is known from WO 2022 / 053648 Al in which a mixture of several brewer's grains is produced using a database and specified nutrient content, flavor, and color of the mixture. WO 2023 / 227566 describes a process in which the resulting flavors are digitally quantified from a specified set of ingredients and vice versa, with the flavors consisting of the components sweetness, bitterness, umami, licorice, aftertaste, sourness, and saltiness.
[0018] The object of the present invention is to provide a method for creating a recipe for a target aroma profile with improved properties.
[0019] In a method for creating a formulation for a target aroma profile of a product, a large number of mappings of aroma profiles to formulations are provided, P2024, 0478 WO E 20 December 2024
[0020] 4
[0021] The aroma profiles are each described by the proportions of identified aroma components, and the recipes each specify the proportions of raw materials. A description of the target aroma profile is determined in the form of proportions of identified aroma components. The aroma components are determined primarily analytically, for example, through instrumental analysis and / or human sensory analysis. A computer-implemented algorithm determines a recipe for the target aroma profile based on the numerous assignments. In addition to the description of the target aroma profile, one or more properties of the recipe and / or the raw materials contained in the recipe can also be entered into the computer-implemented algorithm.
[0022] The computer-implemented algorithm can be a machine learning algorithm based on an artificial neural network. Alternatively, a large language model can be used for recipe generation. The numerous mappings of aroma profiles to recipes can be used to train and / or test the algorithm. For example, an iterative feedback mechanism can be used to improve the results by integrating evaluations of the generated aroma profile into the learning process.
[0023] The algorithm could, for example, also be a decision tree method (e.g., Random Forest, Gradient Boosting Trees). In such methods, decisions are represented in a tree-like structure, where each node represents a decision based on an attribute. P2024, 0478 WO E December 20, 2024
[0024] 5
[0025] The multitude of mappings between aroma profiles and recipes is generated by a person skilled in the art, particularly a flavorist or perfumer. This process begins with the analytical determination of the proportions of aroma components in a product, followed by the attempt to replicate the composition as closely as possible using a recipe. The person skilled in the art starts with an initial recipe, for example, for an aroma or flavoring preparation, and evaluates the aroma profile of this recipe, for instance, by tasting or smelling it, in comparison to the target aroma profile. The recipe is then iteratively corrected and re-evaluated. This iterative process is repeated until a suitable mapping is found. The resulting mappings can then be used as training data for the algorithm.
[0026] The product is, for example, an aroma, a foodstuff, a luxury item, a fragrance, a fragrance mixture such as a perfume, and / or a cosmetic product.
[0027] The multitude of possible flavor combinations can be systematically determined; for example, various fruit flavors can be recreated for given end products, such as yogurt. These combinations can also exist as experiential knowledge for a wide variety of products and flavors and can be further expanded.
[0028] A formulation is a list of the raw materials to be used, including their respective proportions. These raw materials are primarily flavoring or flavor-modulating substances. Depending on the product, a matrix may also be added. The matrix is a base substance P2024, 0478 WO E 20 December 2024
[0029] A term refers to a product's matrix that does not directly contribute to the aroma profile but indirectly influences the perception of that aroma. For example, the matrix of a food, such as bread, can include starch, gluten, and water.
[0030] A matrix and / or a product type can be specified as additional information in the aroma profiles of the assignments and / or in the target aroma profile. This allows the algorithm to consider the matrix when determining the recipe, for example, by giving greater weight to training data with the corresponding matrix. It is also possible to use only assignments with a matching matrix for training the algorithm.
[0031] The proportions of the aroma components of the target aroma profile in the product can be determined using instrumental analysis. In this case, the aroma components comprise instrumentally determined, analytical aroma components, also called analytes. The analyte composition can also be referred to as the analytical aroma profile.
[0032] Ideally, the analyte composition is a list of all contained analytes with their respective proportions, such that the sum of the proportions theoretically equals 100%. In practice, however, it is either impossible or prohibitively expensive to determine the analyte composition completely. Typically, less than 95% of the analytes in a product can be reliably identified.
[0033] Additionally or alternatively, the aroma components may include descriptive aroma components determined by human sensory analysis. In particular, the following may be used to describe the target P2024, 0478 WO E 20 December 2024
[0034] Aroma profiles involve an organoleptic evaluation. This determines the proportions of different aroma qualities (descriptors) in the product. The descriptor composition can also be referred to as a quantitative descriptive or organoleptic aroma profile.
[0035] This aroma profile is captured, for example, via a QDA (Quantitative Descriptive Analysis), in which all contained aroma qualities are listed with their respective intensities, so that the sum of the intensities is normalized to 100%. In practice, however, it is either impossible or unacceptably costly to determine the complete composition of descriptors. As a rule, less than 95% of the descriptive components of a product can be reliably determined.
[0036] It is also possible to describe an aroma profile by combining instrumentally analytical and human sensory descriptive aroma components.
[0037] In both the instrumental-analytical and descriptive descriptions of the target aroma profile, it is generally not possible to fully capture all components. For example, a target aroma profile may contain aroma components for which there is no direct indication through instrumental analysis or descriptive analysis, but which nevertheless contribute to the aroma profile. By using known, flavoristically evaluated associations as the basis for the algorithm, the resulting recipe can still contain such aroma components. Thus, the algorithm also utilizes the flavorist knowledge embedded in the associations. In other words, a recipe does not necessarily achieve the best reproduction. P2024, 0478 WO E 20 December 2024
[0038] 8
[0039] A target aroma profile is not achieved if the aroma profile obtained with the formulation completely replicates the analyte composition. Rather, the best match is given by organoleptic similarity.
[0040] This method allows for the efficient determination of a suitable recipe for a target aroma profile based on known recipes for aroma profiles. Once the recipe is determined, the aroma, or another product containing the corresponding aroma, can be manufactured, and organoleptic evaluation can be used to verify whether the target aroma profile has been adequately reproduced. If not, the flavorist can make further optimizations to the recipe and use the results to improve the algorithm.
[0041] In the simplest case, the raw materials available for the recipe to be determined are specified by the raw materials included in the allocations.
[0042] It is also possible to use a list of possible raw materials as input for the algorithm. The possible raw materials are, for example, the raw materials actually available and / or desired for the recipe. The list may contain fewer raw materials than the recipes of the assignments. It is also possible for the list to contain raw materials that are not included in the underlying recipes. In this case, it may be specified, for example, which known raw materials can be replaced by a new raw material. P2024, 0478 WO E 20 December 2024
[0043] 9
[0044] The algorithm can be trained to output only recipes that contain only raw materials listed in the list.
[0045] In this process, the list of raw materials can already be used to train the algorithm. For example, only those assignments that contain only the available raw materials are used for training. Different lists can also be provided, and the learning algorithm can be trained differently with each list.
[0046] It is also possible that the list of available raw materials is only entered and / or used when determining a specific target aroma profile. In this case, for example, training is performed non-selectively with the entire available database, and the selection of recipes based on the entered list only occurs when determining the target aroma profile. It is also possible to enter different lists for training and for generating a specific recipe.
[0047] This makes it possible to limit the raw materials for the recipe to available, cost-effective, and / or other criteria-based raw materials. Generally, the target flavor profile can be generated from raw materials that do not appear in the list of available ingredients. Thus, a recipe contains a "substitute" for the unavailable raw materials.
[0048] In general, one or more properties of the formulation to be determined and / or of the raw materials contained in the formulation can be entered into the algorithm. P2024, 0478 WO E 20 December 2024
[0049] 10
[0050] An example of such an input property is the previously described input of a list of possible raw materials. Thus, the recipe only contains raw materials that are present in the list. This property can be viewed as a property of the recipe itself as well as of its individual attributes.
[0051] Another possible attribute is product characteristics such as halal, kosher, vegan, allergen-free, or GMO-free (Genetically Modified Organism). Any number of other attributes or certifications, such as Fair Trade, can also be entered, even up to individual customer requests. This attribute can be considered both a characteristic of the recipe and of the individual attributes. For example, a list of raw materials would specify the properties of each individual raw material.
[0052] The input of a desired recipe property can be considered during algorithm training, so that, for example, only recipes with that property are used for training. It's also possible to select recipes that possess the respective property when determining the recipe. Thus, for example, the property "halal" is entered for a target flavor profile, and the algorithm then only outputs recipes that meet this criterion. It's also possible for the algorithm to output other recipes as well, but the "halal" and "non-halal" recipes are clearly marked as such.
[0053] For example, it is also possible that a product from which the target aroma profile is obtained does not possess the specified property. Therefore, a formulation for the target aroma profile should be developed according to P2024, 0478 WO E, dated December 20, 2024.
[0054] 11
[0055] The goal is to obtain an aroma profile that exhibits this property, thus resulting in an alternative product with a suitable target aroma profile. For example, the actual product is not halal, while the desired recipe should be halal.
[0056] Another possible property that can be entered into the algorithm is, for example, the maximum cost of a recipe. For this purpose, the cost of a raw material can be stored in a list of raw materials. The algorithm then only generates recipes where the maximum cost of the contained raw materials is not exceeded, or it outputs the maximum cost for each recipe.
[0057] Another possible property that can be input into the algorithm is, for example, the maximum number of different raw materials in the recipe. The algorithm then only generates recipes where the maximum number of different raw materials is not exceeded, or it specifies the maximum number of raw materials for each recipe.
[0058] Thus, it is also possible, for example, that the target aroma profile is generated from a product that is more expensive or contains more different raw materials than desired, and a suitable recipe for a replacement product is to be generated.
[0059] Another possible property that can be used as input is that the recipe to be determined only contains raw materials for which there is a reference in the aroma components of the target aroma profile. Alternatively, it can be specified as a property that a recipe can also be determined P2024, 0478 WO E 20 December 2024
[0060] 12
[0061] This will result in ingredients for which there is no indication in the aroma components of the target aroma profile. Therefore, the creativity of the proposed recipes can be limited.
[0062] In particular, a weighting can also be applied between the two variants, so that a measure of the conformity with the aroma components is specified.
[0063] The recipe determined using the computer-implemented algorithm can, for example, be output to a user. It is also possible for the recipe to be used to produce a flavoring or another product. For instance, the determined recipe can be transmitted directly to a dosing device for measuring the proportions of the raw materials in the recipe.
[0064] The description of the target aroma profile in the form of proportions of instrumentally determined, analytical aroma components can, for example, be determined by an analyzer and transmitted directly from the analyzer to a data processing device that executes the algorithm.
[0065] For example, the analyzer determines the analytical composition of the target aroma profile. This allows the entire process, from analyzing a product's target aroma profile and determining a suitable formula to dosing and creating a product such as an aroma, to be automated. This significantly accelerates the creation of a suitable formula.
[0066] According to a further aspect of the present invention, a data processing device is configured to carry out one, several, or all steps of the method described above. The data processing device can be a P2024, 0478 WO E 20 December 2024
[0067] 13
[0068] The device may have an input interface for the proportions of the aroma components, an input interface for the analytical composition, and / or an output interface for the determined recipe. Furthermore, the data processing device may include a processor designed to execute the algorithm.
[0069] According to a further aspect of the present invention, a system comprises the data processing device described above. The system also includes an analysis device for determining the proportions of the aroma components of the target aroma profile and / or a dosing device for dosing the proportions of the raw materials. Thus, the system can be configured to analyze a target aroma profile of a product, determine a suitable recipe, and dose the raw materials according to the recipe, so that an aroma or other product replicating the target aroma profile is produced.
[0070] The present invention comprises several aspects, in particular methods and devices. The features, properties, and characteristics described for one of the aspects are
[0071] The same principles should apply to the other aspect as well.
[0072] Furthermore, the description of the items listed here is not limited to the specific designs.
[0073] Rather, the features of the individual embodiments can be combined with each other – insofar as this is technically feasible.
[0074] The following section provides a more detailed explanation of the items described here, using schematic examples. P2024, 0478 WO E 20 December 2024
[0075] They show:
[0076] Figure 1 shows an embodiment of a method in a schematic flowchart.
[0077] Figure 2A shows a schematic example of a data set with assignments for the procedure according to Figure 1.
[0078] Figure 2B shows an example of a list of raw materials for the process according to Figure 1 in schematic representation.
[0079] Figure 3 shows an example of an aroma profile being assigned to a recipe for the process according to Figure 1.
[0080] Figure 4 shows another example of an aroma profile being assigned to a recipe for the process according to Figure 1.
[0081] Figure 5 shows another example of an aroma profile being assigned to a recipe for the process according to Figure 1.
[0082] Figure 6 shows an example of an aroma profile being assigned to a recipe for the process according to Figure 1.
[0083] Figure 7 shows a schematic representation of an embodiment of a system comprising a data processing device, an analysis device and a dosing device.
[0084] Preferably, the following figures refer to the same P2024, 0478 WO E dated December 20, 2024.
[0085] 15
[0086] Reference symbols to functionally or structurally corresponding parts of the different embodiments.
[0087] Figure 1 shows a schematic view of an embodiment of a method for creating a formulation R x
[0088] a target aroma profile for a product ("Target").
[0089] This could be, for example, an aroma profile of an aroma that can be used to flavor a food, beverage, or cosmetic product, particularly a so-called B2B product. It could also be the aroma profile of such an end product, particularly a so-called B2C (business-to-consumer) product for an end consumer. For example, it could also be a product in the perfumery sector, especially technical and fine perfumery. It could also be a food, beverage, or cosmetic product. It could also be a natural product, such as a fruit.
[0090] In step SO of the process, a data set 1 is provided. Data set 1 contains a mapping of numerous recipes to descriptions of aroma profiles. These mappings are also referred to as data pairs and, for each aroma profile description, specify one or more recipes that result in a matching aroma profile and can therefore also be described as a "match".
[0091] The aroma profile is described by the proportions of identified aroma components, where the aroma components are instrumentally analytical and / or P2024, 0478 WO E 20 December 2024
[0092] 16
[0093] It can be about human sensory descriptive aroma components.
[0094] Aroma components are determined through instrumental analysis. Human sensory-descriptive aroma components can be determined through organoleptic evaluation, for example, by a flavorist. The aroma profiles are generally not fully described by the identified aroma components. For instance, some aroma components cannot be detected instrumentally or descriptively.
[0095] The description of the aroma profile Z x Provides a breakdown of the product's components ("analytes" or "descriptors") obtained through sensory or instrumental analysis. The description of the aroma profile Z x is a list of all detected aroma components, in particular the detected analytes and / or descriptors, with their respective proportions.
[0096] Analytes are components of input materials, where an input material can be an available raw material or an unknown component of a product.
[0097] The analyte composition determined by instrumental analysis typically differs from the raw materials from which the product or the actual aroma profile is or can be produced. Firstly, a raw material often contains a multitude of analytical constituents. Secondly, the raw materials react chemically with one another, so that the aroma profile has a different analytical composition than the individual raw materials. P2024, 0478 WO E 20 December 2024
[0098] 17
[0099] It may exhibit certain properties. Furthermore, not all ingredients are necessarily detectable or identifiable in the analysis.
[0100] For descriptive aroma components, the aroma components are detected using human sensory methods, especially organoleptic methods, and are specified, for example, in the form of a QDA (QDA for Quantitative Descriptive Analysis).
[0101] Descriptors describe a sensory perception, such as "fruity," "tart," "sweet," "vanilla-like," etc. Such descriptors can be general (e.g., "sweet") or specific (e.g., "vanilla-like"). To describe an aroma profile, the descriptors are listed according to their proportions. Additionally, the intensity of each descriptor can be specified, for example, on a scale from 1 to 10. Descriptors, in turn, characterize raw materials or their combinations, where a raw material can be an available commodity or an unknown component of a product. A raw material can be chemically defined (molecule) or a complex mixture of chemically defined substances.
[0102] QDA is typically not performed instrumentally, but rather by tasters (panelists). This is done either through free associations or using reference descriptors. Robust results are obtained through repeated evaluations and multiple panelists. Usually, means are calculated from the individual data points, or a consensus profile is created.
[0103] If the description of the target aroma profile is entered in the form of descriptive aroma components, instrumental analysis can also be omitted, so that P2024, 0478 WO E 20 December 2024
[0104] 18
[0105] The aroma profile is not specified by means of an instrumentally analytical composition, but only a descriptive one. A combination of organoleptic and analytical descriptions is also possible.
[0106] Data set 1 can be stored on a computer-readable storage medium 3. This computer-readable storage medium 3 can be any type of data carrier, such as a physical storage medium like a hard drive, a disk drive, an optical disc, a USB flash drive, or even network storage or cloud storage. The data set is provided to a processor via the storage medium 3. Data set 1 can, for example, also be provided in a cloud.
[0107] Dataset 1, or a portion thereof, is used to train a machine learning algorithm. The algorithm is run on at least one processor. A further portion of Dataset 1 can be used to test the algorithm. The algorithm is stored in memory accessible to the processor, for example, on non-volatile storage media or in cloud-based storage.
[0108] Furthermore, a list of two possible raw materials can be provided in digital form. Possible raw materials include, for example, currently available or desired raw materials. List 2 can be provided separately from data set 1 or together with data set 1 in a computer-readable storage medium 3. List 2 can also specify various properties of the respective raw materials, for example, in the form of key figures. It is also possible that the assignments of recipes to the data sets are alternative or P2024, 0478 WO E 20 December 2024
[0109] 19
[0110] Additionally, specify properties of the recipe and / or the raw materials.
[0111] List 2 can be entered in step SO, as shown here, to train the algorithm. In this case, for example, only matches where the recipe contains only raw materials from List 2 are used for training. It is also possible to enter such a List 2 in a later step, when determining a recipe for the target aroma profile, so that, for example, the recipe suggestions can then be filtered based on the list. It is also possible to enter lists of possible raw materials both in step SO and in a later step.
[0112] In step S1, a description of a target aroma profile Z is created. xThe description can be provided through instrumental analysis and / or human sensory evaluation. As with the aroma profile descriptions provided in step SO, the proportions of instrumentally analytical and / or descriptive aroma components are specified.
[0113] The exact recipe of the target aroma profile, i.e., the raw materials used for its production and their respective proportions, is not known when using the method to obtain a recipe. However, when using the method to test and / or train the algorithm, for example, a neural network, the recipe may be known.
[0114] In step S2, the description of the target aroma profile Z is created. x Entered into a data processing device 4 P2024, 0478 WO E 20 December 2024
[0115] 20
[0116] and used as an input value for a computer-implemented algorithm.
[0117] Input can be provided via an input interface 5 of the data processing device 4, for example manually using an input device, in particular a keyboard. The input device can be directly connected to the data processing device 4 or the data can be transmitted to the data processing device 4 via a network, for example a local area network (LAN) or a wide area network (WAN), in particular the Internet. It is also possible for input to be provided automatically or semi-automatically from another device to the data processing device 4, for example via an analytical device used for analyzing the aroma profile of a product.
[0118] It is also possible to input further data, particularly metadata, into the data processing device 4. For example, the type of product, e.g., flavoring or food, and optionally a food matrix can be entered. A matrix refers to a basic component of a product that does not directly contribute to the flavor profile but indirectly influences the perception of the flavor. For example, the matrix of a food such as bread might include starch, gluten, and water.
[0119] Desired properties E1-E can also be specified j The recipe to be determined and / or the raw materials contained in the recipe must be specified. For example, it can be stated whether a recipe should be halal, kosher, vegan, allergen-free, or GMO-free. Another characteristic could, for example, be the price of P2024, 0478 WO E 20 December 2024
[0120] 21
[0121] This could be the raw materials included or a maximum number of different raw materials. Furthermore, it is also possible in this step to enter a list of possible or desired raw materials for the product.
[0122] The analytical method used, including a sample preparation method, and other properties such as density, refractive index or color of the product and / or a prepared sample can also be specified.
[0123] In step S3, one or more suitable recipes R are selected using the computer-implemented algorithm 7. x for the product's aroma profile. A recipe is considered "suitable" if the aroma profile of a product manufactured using that recipe is identical or very similar to the aroma profile of the product whose analytical composition Z xThe system is available. Multiple suitable recipes can also be displayed. Furthermore, a match score can be displayed, indicating how well the recipe matches the target aroma profile.
[0124] The algorithm used to determine a recipe may be a machine learning algorithm that uses an artificial neural network.
[0125] For this purpose, for example a multi-layer perceptron network (MLP), graph neural network (GNN), variational autoencoder (VAE) or another well-known machine learning model can be used.
[0126] Furthermore, a Large Language Model (LLM), e.g., LLaMa, can be used for recipe generation. In this process, the data is fed into the language model using Retrieval-Augmented Generation (RAG) P2024, 0478 WO E 20 December 2024
[0127] 22
[0128] The input data is provided either in a structured format (e.g., JSON, XML) or as a Knowledge Graph (KG). By processing and integrating this information, the LLM can identify patterns and relationships relevant for generating recipe suggestions. The model processes the inputs and generates combinations of raw materials that match the target aroma profile. An iterative feedback mechanism enables the continuous improvement of the suggestions by integrating user feedback and test results into the learning process.
[0129] The algorithm could, for example, also be a decision tree method (e.g., Random Forest, Gradient Boosting Trees). In such methods, decisions are represented in a tree-like structure, with each node representing a decision based on an attribute. For example, starting with the presence and proportion of one or more specific analytes, a decision is made regarding the presence and proportion of a raw material by comparing this information with the existing assignments. Each node of the tree represents a decision concerning the proportion of a specific raw material. The leaves of the tree represent the resulting formula derived from the decisions made.
[0130] In step S4, the recipe is output via an output interface 6. Output interface 6 can be configured for output to a human user and / or to another device. This other device could, for example, be a device for dosing the raw materials according to the determined recipe. Thus, P2024, 0478 WO E 20 December 2024
[0131] 23
[0132] For example, an aroma or flavoring preparation can be produced directly for a product.
[0133] The described method thus generates an optimal recipe for a target aroma profile without elucidating the mechanisms of aroma formation. It is possible to efficiently refine a target aroma profile. From the determined recipes, an aroma, a flavoring preparation, or the product itself can be produced and evaluated by a flavorist. The flavorist can select the most suitable recipe and / or perform further refinements. Once the recipe is suitable, it can be used to manufacture the product and also to train the algorithm.
[0134] Figure 2A shows a schematic example of a data set 1 with Mi-M assignments. c , which are provided in the procedure according to Figure 1 and can be used in particular as a training data set or as a test data set for a machine learning algorithm.
[0135] Dataset 1 contains a large number of Mi-M mappings. c of descriptions of aroma profiles Zi-Z a Regarding recipes Ri-Rb for flavorings, flavoring preparations and / or products. List 2 contains the possible raw materials Si-Sd.
[0136] The Mi-M assignments shown here c are unique, so that each analytical composition Zi is assigned exactly one formulation Ri and vice versa. It is also possible that one analytical composition Zi is assigned several formulations Ri, Rj and / or one formulation Ri is assigned several compositions Zi, Zj. The formulations, P2024, 0478 WO E 20 December 2024
[0137] While ingredients may be assigned the same flavor profile, they can differ in the properties of the raw materials used. Such properties can include, for example, different purity levels, sensory qualities of the raw materials, or characteristics such as "halal," "kosher," etc. It is also possible for recipes with entirely different raw materials to be assigned the same flavor profile, provided they meet the requirements of a "match."
[0138] The Mi-M assignments c are based on expert
[0139] (Flavoristic) evaluation competence or alternatively, methods of sensory evaluation such as triangle tests or difference-from-control tests. In very few cases does this correspond to an analytical composition Z. x assigned formula R x the actual recipe of a product from which the analytical composition Z is derived xwas obtained. For example, it is possible that not all ingredients can be analytically determined, that a corresponding raw material cannot be identified or is not available for all ingredients, or that it is known from a flavorist perspective that an ingredient does not contribute to the aroma profile. However, the aroma profile, which is evaluated organoleptically, is one obtained according to the recipe R. x The received product is identical or similar to the target aroma profile.
[0140] Figure 2B shows a schematic example of a list 2 with available raw materials S1-S3 that can be provided in step S0 and / or step S1 of the process according to Figure 1. P2024, 0478 WO E 20 December 2024
[0141] 25
[0142] The available raw materials Si-Sd can be any raw materials that are known practically or theoretically, e.g. from literature, and are available or potentially available for the manufacture of a product.
[0143] It is also possible that the raw materials Si-Sd have properties K1-K i These characteristics are specified, for example, in the form of key figures. For instance, a characteristic might indicate suitability for a particular product. For example, a characteristic could be halal, kosher, vegan, allergen-free, and / or GMO-free. Any other characteristics can also be specified, such as Fair Trade. It could also refer to the availability and / or price of a raw material.
[0144] If a corresponding property of a raw material and / or the recipe is specified in the input for the process, the algorithm will, for example, only output recipes in which the recipe and / or the raw materials it contains exhibit the specified property. The price of all raw materials contained in a recipe can also be used as input for the algorithm. If a price is specified for each individual raw material, then only recipes with matching raw material prices can be output.
[0145] In principle, it is possible to take desired properties into account during the training of the algorithm.
[0146] For example, only the assignments where the properties match are used for training.
[0147] Alternatively, the properties can be taken into account when determining a recipe for a target aroma profile. P2024, 0478 WO E 20 December 2024
[0148] - 26 -
[0149] Figure 3 shows an example of an actual raw material composition P x of a product, the description of the target aroma profile Z determined by instrumental analysis x , which can serve as input for the procedure according to Figure 1, and the recipe R output by the algorithm x .
[0150] The product in question is, for example, a flavoring or a flavoring preparation. A flavoring preparation contains, in addition to the flavoring and technologically necessary components (see, for example, Regulation (EC) No. 1334 / 2008, the "Flavorings Regulation," for Europe), other functional components, such as coloring, clouding, nutritional, or stabilizing ingredients.
[0151] The product has a composition P x from the raw materials Hl, H2, H3, H4 with the respective proportions of 35%, 15%, 40%, and 10%. The composition P given herex The product's condition is not usually known in the process.
[0152] It is also possible that the composition P x The composition is already known, but the algorithm is to find a suitable recipe with a different composition. This may be desirable if raw materials are unavailable or too expensive, or if the raw materials have properties that should not be present in the product, for example, if a halal, kosher, or allergen-free product is desired. It is also possible that the composition is known if it is a training or test dataset. P2024, 0478 WO E 20 December 2024
[0153] 27
[0154] Therefore, the algorithm is used to create a suitable recipe based on available or desired raw materials. The goal is to replicate a target aroma profile as closely as possible through a recipe.
[0155] The description of the target aroma profile Z x The present statement is instrumentally analytical, meaning it specifies the determined composition of the analytically detected aroma components (= analytes) obtained through instrumental analysis of the product. The composition includes analytes a1-a8 in the specified proportions and, where applicable, other analytes. The sum of the proportions of analytes a1-a8 can be 100%. A raw material can be characterized by several different analytes. Furthermore, different raw materials of the target aroma can contain proportions of identical analytes. Additionally, chemical interactions can occur between the components of the raw materials, meaning that an analytically detected component is not necessarily present in every raw material.
[0156] In the present case, H1-H applies to all raw materials. e Analytical clues. However, this is not necessarily the case in all instances.
[0157] Using the computer-implemented algorithm, a formula R is generated for the analytical composition. x , for example, based on available raw materials Si-Sd. The T1, T2, and T3 are determined with the specified proportions.
[0158] proposed.
[0159] Raw materials T1, T2, and T3 may overlap with raw materials H1-H4. However, they may also be entirely different raw materials. In particular, it is possible that one or more of raw materials H1-H4 are used in recipe R. x not occur. It is also possible that in P2024, 0478 WO E on December 20, 2024
[0160] 28
[0161] the formula R x Raw materials versus the actual raw material composition P x Furthermore, it is also possible that the analytical composition of an aroma mixed according to the recipe differs from the analytical composition of the original aroma.
[0162] Nevertheless, the aroma profiles can be identical or nearly identical.
[0163] The target aroma profile can be described alternatively or additionally by specifying the proportions of descriptors di-dh. For example, dl: "fruity" - 40%, d2: "vanilla" - 60%. Here too, the total proportion can be 100%. An example with descriptors is shown in Figure 6.
[0164] Figure 4 shows another example of a raw material composition P x of a product, the analytical description of the target aroma profile Z determined by instrumental analysis x , which can be used as input for the procedure according to Figure 1 and the output recipe R x .
[0165] In contrast to Figure 3, the target aroma contains a raw material H5, which is not found in the analyte composition Z. xno analytical evidence is given (visible from the missing connecting lines from H5 to P) x to analytes from Z x Although the analyte composition provides no indication of raw material H5, the algorithm can be used to infer its presence based on existing expert knowledge data. Unlike Figure 3, the resulting formulation contains the additional raw material T4 at a rate of 2%. Raw material T4 may be the same as raw material H5 or different. P2024, 0478 WO E 20 December 2024
[0166] - 29 -
[0167] The compositions shown in Figures 3 and 4 may be Z x to obtain identical compositions of instrumentally analytically detected aroma components for different product compositions P x trade. Thus, for example, for an analyte composition Z x two different recipes R xwill be issued.
[0168] It is also possible that the starting materials H1-H5, analytes a1-a8 and proposed raw materials T1-T4 from Fig. 4 differ from those in Fig. 3. Therefore, they could also be different products.
[0169] Figure 5 shows another example of a raw material composition P x of a product, the analytical composition Z determined by instrumental analysis x , which can be used as input for the procedure according to Figure 1, and the output recipe R x .
[0170] In this example, the product contains a matrix M. The product could be, for example, a foodstuff. Analyte compositions are often analyzed from within a matrix. A matrix can be, for example, a simple sugar mass (hard sugar candy) or a complex system, such as an emulsion (mayonnaise). A matrix is not present in all products; for example, a flavoring may contain only flavoring components, as in the examples in Figures 3 and 4.
[0171] The matrix does not essentially contribute directly to the aroma profile and the analyte composition. x It is present, but can indirectly influence the aroma profile. Furthermore, matrix M can also affect the analysis result. Information on matrix M P2024, 0478 WO E 20 December 2024
[0172] 30
[0173] It can be included as additional input in the algorithm, for example in the form of metadata. Information about the matrix can also be included in the training data formulas.
[0174] The analyte composition Z is present. x identical to Figure 3, where the product does not contain a matrix. However, it is also possible that a different analyte composition Z may be obtained depending on the chosen analytical method. x won.
[0175] The presence of a matrix M can be taken into account in the learning algorithm. In particular, training data derived from a similar matrix can be weighted more heavily.
[0176] In the example of Figure 5, an analyte composition identical to that of Figure 3 is used. x by entering the matrix M, a slightly different recipe R xobtained. The proportions of the available raw materials T1, T2, T3 are slightly different here than in Figure 3. It is also possible that other raw materials are used in the recipe R. x be proposed.
[0177] Alternatively or additionally, the analytical processing method can also be used as metadata. The analytical processing method can depend on the matrix, as different sample processing methods or analytical procedures are applied depending on the matrix from which the analyte composition is determined.
[0178] Matrix M also has an influence on the description of the aroma profile using descriptors (P2024, 0478 WO E, December 20, 2024).
[0179] The determined descriptor composition is important because the matrix can influence the release of aroma compounds. Aroma compounds that are retained contribute less to the aroma profile and thus to the descriptor composition than aroma compounds that are released more readily.
[0180] The following are several specific examples of target flavors P. x , Analyte compositions Z x and recipes R x described. This can involve determined recipes, but also training data sets or test data sets of the algorithm.
[0181] In a first example, the product (target flavor) is 100% commercially available vanillin. Analysis of the target flavor reveals 99.999% vanillin and 0.001% guaiac. The recipe for recreating the target flavor consists of 100% vanillin of a specific grade, namely vanillin ex guaiac. The target flavor and the recipe have an identical flavor profile. The algorithm thus infers from the analysis result an available raw material, in this case vanillin ex guaiac, and a concentration, in this case 100%. After receiving the recipe, it can be verified that the flavor profiles of the target flavor and the recipe are identical.
[0182] According to a second example, a target flavor consists of 99% commercially available vanillin and 1% commercially available diacetyl. Analysis of the target flavor reveals 97.999% vanillin, 0.001% guaiac, and 2% diacetyl. The resulting formulation for recreating the target flavor consists of 99% vanillin of a special quality, namely vanillin ex guaiac, and 1% diacetyl. This shows that the analytical composition (2% diacetyl) does not necessarily correspond to the composition in P2024, 0478 WO E dated December 20, 2024.
[0183] must comply with the proposed formula (1% diacetyl) in order to replicate the aroma profile.
[0184] According to a third example, a formulation consists of 98% vanillin, 1% diacetyl, and 1% Bourbon vanilla extract. Analysis reveals 97% vanillin, 1.5% diacetyl, and a total of 1.5% other analytically detectable components. These comprise 50 analytes, including, for example, furfural, guaiac, methyl guaiac, vanillic acid, p-hydroxy benzoic acid, p-hydroxybenzaldehyde, and some unidentified molecules. Some of the analytically detected components indicate the presence of vanilla extract. The formulation for recreating the target flavor consists of 98% vanillin, 1% diacetyl, and 1% of an available vanilla extract.
[0185] According to a fourth example, a target flavor formulation consists of 97.9% vanillin, 1% diacetyl, 1% Bourbon vanilla extract, and 0.1% lemon oil. Analysis reveals 97% vanillin, 1.5% diacetyl, and a total of 1.5% other analytically detectable components. These comprise 50 analytes, including, for example, furfural, guaiac, methylguaiac, vanillic acid, p-hydroxybenzoic acid, p-hydroxybenzaldehyde, and some unidentified molecules. Some of the analytically detected components indicate the presence of vanilla extract. However, no analyte indicates the presence of lemon oil in the formulation. The formulation used to recreate the target flavor consists of 98% vanillin, 0.9% diacetyl, 1% of an available vanilla extract, and 0.1% of an available lime oil. Therefore, in this example, lime oil is suggested as a formulation ingredient, even though there is no direct indication of it in the analyte composition. P2024, 0478 WO E 20 December 2024
[0186] According to a fifth example, a target flavor formulation consists of over 100 individual ingredients. Analysis of the target flavor reveals more than 200 identified and unidentified analytes with their determined concentrations. The formulation for recreating the target flavor consists of 20 ingredients, of which 17 have analytical data and 3 do not.
[0187] It is possible to specify a maximum number of raw materials in the recipe as an additional input to the algorithm, ensuring that the suggested recipes meet this criterion. For example, such recipes will then be weighted more heavily in the algorithm. This allows, for instance, more complex raw materials to be suggested instead of combinations of pure substances. It is also possible, depending on the similarity of the sensory effects of raw materials, to suggest only one representative for a particular effect instead of many similar raw materials.
[0188] Overall, the process does not represent a re-mixing of the recipes based on analytical results, but rather aims to replicate the aroma profile. Therefore, unexpected recipe components and / or unexpected proportions may be present in the recreated recipe.
[0189] Figure 6 shows another example of a formula R x , which are for a target aroma profile Z xis determined. In contrast to Figure 3, the description of the target aroma profile Z is x Determined by human sensory perception, it is described by the proportions of descriptive aroma components d1-d8. Here too, the sum of all proportions can be 100%. As shown in Figure P2024, 0478 WO E 20 December 2024
[0190] 34
[0191] 3 can describe the target aroma profile Z x serve as input for the process according to Figure 1, so that a recipe R x is determined. The training data for the algorithm can therefore also contain a description using descriptive aroma components.
[0192] For example, the descriptors could be as follows: dl: fruity; d2: rosy; d3: jasmine-like; d4: vanilla-like; d5: musky; d6: woody; d7: floral; d8: citrusy.
[0193] It is also possible to use a combination of instrumentally analytical and descriptive aroma components.
[0194] The examples in Figures 4 and 5 also apply accordingly to descriptions using descriptive aroma components.
[0195] Figure 7 shows an embodiment of a system 10 comprising a data processing device 4, an analysis device 5 and a dosing device 9. The data processing device 4 can be the data processing device 4 described above, which is configured to perform the method according to Figure 1.
[0196] The analyzer 5 is designed to describe the target aroma profile Z. x , in particular to determine the instrumental analytical composition of the aroma profile of a product. The analyzer 5 is communicatively connected to the data processing device 4, for example by a direct physical connection or by a network, and is configured to describe the target aroma profile Z. xto be transmitted to data processing device 4. P2024, 0478 WO E 20 December 2024
[0197] The data processing device 4 is configured to perform the procedure according to Figure 1 and to generate a recipe R x to determine the aroma profile.
[0198] The data processing device 4 is communicatively connected to the dosing device 9, for example by a direct physical connection or via a network. The data processing device 4 is configured to process the determined formula R. x to transmit to the dosing device 9.
[0199] The dosing device 9 is designed to dispense the amount specified in the formula R. x The raw materials are dosed and mixed according to the recipe in their respective proportions. This allows for the creation of an aroma or flavoring preparation. It is also possible to add an aroma directly to a desired matrix.
[0200] After the product has been manufactured, the aroma profile can be evaluated and it can be verified whether the target aroma profile has been accurately replicated. This evaluation can be performed by a specialist using organoleptic analysis.
[0201] The following are further concrete examples of how to obtain possible training data, analysis methods and evaluations.
[0202] The target product is, for example, a product from the natural product category (e.g., a fruit), a B2B product (e.g., a flavoring), or a B2C product (e.g., a baked good, an oral care product, or a cosmetic product). The targets for obtaining the training data are selected to cover the broadest possible range of differently defined targets in all categories. It is also possible to refer to P2024, 0478 WO E 20 December 2024
[0203] 36
[0204] to focus on a specific target product or product category. Targets can also be created for the purpose of gathering training or test data.
[0205] The analysis of the target can be performed, for example, instrumentally, organoleptically, or through a combination of methods.
[0206] Instrumental analysis can generally be divided into three steps: sample preparation, analyte separation, and analyte detection, i.e., identification and quantification. Depending on the analytical method, the analyte separation step may be omitted, e.g., in NMR analysis.
[0207] Sample preparation of the target product is possible for both instrumental and human sensory analysis. Depending on the target, sample preparation may also be omitted. During sample preparation, the matrix can be removed, and the aroma compounds extracted and concentrated. It is also possible to chemically derivatize the aroma compounds before or during analysis. Examples of sample preparation methods include dissolution or dilution, liquid-liquid extraction, liquid-solid extraction, solid-phase extraction, water / steam destination, microwave-assisted extraction, derivatization for non-volatile components, and thermal extraction. A number of other possible sample preparation methods are known.
[0208] Chromatographic methods (GC, LC) can be used to separate the aroma components. Detection is carried out using various detectors such as MS and FID for P2024, 0478 WO E 20 December 2024
[0209] 37
[0210] Identification and quantification. Examples of separation and detection methods include gas chromatography, liquid chromatography, mass spectrometry, IR, UV, Vis spectroscopy, and NMR spectroscopy.
[0211] In organoleptic analysis, suitable descriptive descriptors are defined and their perceived intensity is determined. Temporal intensity profiles during perception, e.g., during consumption, are also possible. Examples of descriptive organoleptic analysis methods include QDA, a simple descriptive test according to DIN 10964: 2014-11, conventional profiling according to DIN 10967-1: 1999-10, and consensus profiling according to DIN 10967-2: 2000-10. In addition, a number of other standardized methods for descriptive analysis are known.
[0212] To create training and / or test data, a specialist (flavorist, perfumer, etc.) can develop a recipe for a match based on the analysis results (analyte or descriptor composition).
[0213] The formulation can, in a first step, essentially be a recombination of the analyte composition using suitable raw materials. For example, the analyte ethyl butyrate is represented by the use of ethyl butyrate as a raw material, the analyte vanillin by the use of pure vanillin or vanilla extract, or the analytes octanal and decanal by the use of orange peel oil.
[0214] The flavorist begins with an initial recipe and evaluates the aroma profile of this recipe by tasting or P2024, 0478 WO E 20 December 2024
[0215] 38
[0216] The taster compares the aroma to the target. The recipe is then adjusted based on these varying impressions and experience regarding the sensory properties of the ingredients, and the revised recipe is re-evaluated. This iterative trial-and-error process is repeated until a suitable recipe is found. Alternatively, a tasting panel can be used for evaluation and validation. Validation can also be achieved through appropriate analytical methods, such as physicochemical or biological-chemical sensor-based or receptor-based artificial noses or tongues. Standardized discrimination tests can also be used to determine the match of aroma profiles.
[0217] After validation, an assignment is made from the description of the aroma components Z. x (instrumental analytical or human sensory descriptive) and prescription Rx found that can be used to train or test the algorithm.
[0218] The data Z x -R x are stored in an electronic database. The characteristics that define the aroma profile take on the role of features in the training process. These features represent the independent variables that are fed into the model.
[0219] The raw materials that make up the recipe represent the target variables. The target variable is the dependent variable that the model is intended to predict or explain.
[0220] Additional features that are associated with the analytical profile as metadata (such as the processing method P2024, 0478 WO E 20 December 2024)
[0221] - 39 -
[0222] or a matrix) or related to the formulation (e.g., legality) can also be used as additional input for the model. Specifically, these are added as further features to the existing data points and thus serve as extended input variables. P2024, 0478 WO E December 20, 2024
[0223] - 40 -
[0224] Reference sign
[0225] 1 data record
[0226] 2 list
[0227] 3 Storage medium
[0228] 4 Data processing device
[0229] 5 Input interface
[0230] 6 Output interface
[0231] 7 Algorithm
[0232] 8 Analyzer
[0233] 9 Dosing device
[0234] 10 System
[0235] Z1-Z a Description Aroma profile
[0236] R1-R b Recipes
[0237] M1-M c Assignments
[0238] Z x Description of target aroma profile
[0239] R x determined formula
[0240] P x Product composition
[0241] S1-S d Raw materials in list
[0242] H1-H e Raw materials of a product
[0243] T1-T f proposed raw materials
[0244] a1-a g Aroma components analytically (analytes)
[0245] d1-d h Descriptive aroma components (descriptors) K1-K i Raw material property
[0246] E1-E j Product properties
[0247] M Matrix
[0248] a, b, c, d, e, f, g, h, i, j: natural numbers
Claims
P2024, 0478 WO E December 20, 2024 Patent claims 1. Procedure for creating a prescription (R x ) for a target aroma profile (Z x ) of a product, comprising the following steps: a) Providing a large number of assignments (M1-M c ) of descriptions of aroma profiles (Z1-Z a ) to formulations (R1-R b ), where the aroma profiles (Z1-Z a ) each determined by proportions of aroma components (a1-a g , d1-d h ) are described and the recipes (R1-R n ) each share of raw materials (T1— T f ) specify, c) Determination of proportions of aroma components (a1-a g , d1-d h ) to describe the target aroma profile (Z x ), d) Determination of at least one formulation (R x ) for the target aroma profile (Z x) using the computer-implemented algorithm (7 ) based on the multitude of assignments (M1-M c ).
2. Method according to claim 1, wherein the aroma components (a1-a g , d1-d h ) instrumentally determined, analytical aroma components (a1-a g ) include.
3. Method according to any one of the preceding claims, wherein the aroma components (a1-a g , d1-d h ) human sensory-determined, descriptive aroma components (d1-d h ) include.
4. Method according to one of the preceding claims, in which a list (2) of available raw materials (S1-S d ) is entered into the algorithm (7 ), the algorithm being trained to generate recipes (R x ) to spend, which only use the available raw materials (S1-S m ) contain. P2024, 0478 WO E December 20, 2024 42 5. Method according to any one of the preceding claims, wherein in the list (2) of raw materials (S1-S d ) Properties (K1-K i ) the respective raw materials (S1-S d ) are to be specified.
6. Method according to any of the preceding claims, wherein one or more properties (E1-E j ; K1-K i ) the formulation to be determined (R x ) and / or the one in the recipe (R x ) contained raw materials are entered into the algorithm (7 ).
7. Method according to one of claims 5 or 6, where the properties (E1-E j ; K1-K i ) selected from one or more of the following: halal, kosher, vegan, allergen-free and / or GMO-free.
8. Method according to any one of claims 5 to 7, where a product from which the target aroma profile (Z) is derived x ) is obtained, the entered property (E1-E j ; K1-K i ) does not exhibit.
9. Method according to any one of claims 5 to 8, where the properties (E1-E j ; K1-K i ) are selected from one or more of: a maximum number of different raw materials (K1-K i ) in the determined formula (R x ) and / or the maximum cost of the determined formula (R x ).
10. Method according to any one of the preceding claims, wherein the properties (E1-E j ; K1-K i ) selected from: Determining a recipe (R x ) containing only raw materials (T1— T f ), which are found in the aroma components (a1-a g , d1-d h ) of the target aroma profile (Z x ) gives a clue, and also determines a recipe (R x ) containing raw materials (T1-T f ), on the P2024, 0478 WO E December 20, 2024 it in the aroma components (ai-a g , di-dh) of the target aroma profile (Z x ) gives no indication.
11. Method according to any of the preceding claims, wherein the computer-implemented algorithm (7) is a machine learning algorithm, wherein the assignments (M1-M c ) to train and / or test the algorithm (7 ).
12. Method according to one of the preceding claims, wherein the determined formula (R x ) is issued to a user and / or used to produce a product, whereby the determined recipe (R x ) directly to a dosing device ( 9) for dosing the proportions of the raw materials (T1-T f ) is transmitted.
13. Method according to any of the preceding claims, wherein the description of the target aroma profile (Z) x ) is determined by an analyzer and transmitted directly from the analyzer to a data processing device that executes the algorithm (7 ).
14. Data processing device (4) configured to perform one or more steps of the method according to any of the preceding claims, having an input interface (5) for inputting the proportions of the determined aroma components (a1-a g , d1-d h ), an output interface ( 6) for outputting the determined recipe (R x ) and a processor designed to perform the algorithm (7 ).
15. System ( 10) comprising the data processing device (4 ) according to claim 14 and comprising an analysis device ( 8 ) for determining the proportions of the aroma components (a1-a g , d1-d h ) of P2024, 0478 WO E December 20, 2024 - 44 - Target aroma profile (Z) x ) and / or a dosing device ( 9) for dosing the proportions of the raw materials (T1-T f ).