Artificial intelligence-based automated generation of drink recipes

Abstract

This invention relates to the field of AI technology for beverage formulation, and discloses an automated method for generating beverage formulations based on artificial intelligence. The method includes collecting flavor characteristic data of five raw materials—raspberry, mulberry, goji berry, pomegranate, and purple grape—as well as health component data including basic nutrients, plant compounds, and the synergistic effects of their derivatives, establishing a structured raw material database. A machine learning model is used to extract features from the database data, generating flavor vectors and health component vectors for each raw material, achieving quantitative representation of the data. An optimization algorithm is employed to perform multi-objective optimization on the two types of vectors, selecting a set of candidate formulations that simultaneously meet preset flavor and health thresholds. The candidate formulations are further filtered based on user preference data, ultimately determining the Five Tiger Divine Powder formulation, which is then output to the formulation execution system. This method precisely meets users' personalized needs, providing a reliable technical path for the standardized production of Five Tiger Divine Powder.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology for beverage recipes, specifically relating to an automated method for generating beverage recipes based on artificial intelligence. Background Technology

[0002] In today's booming beverage industry, consumer demand has moved beyond the basic function of quenching thirst; they increasingly value unique flavors and the health benefits of healthy ingredients. Under this trend, compound puree beverages, combining excellent taste and high nutritional value, have gradually emerged as a market focus. Compound puree beverages, represented by "Five Tiger Divine Puree," have stood out in this wave and gained popularity thanks to their unique formula using a variety of carefully selected natural purees. However, despite the growing market demand, the current formula development of "Five Tiger Divine Puree" and similar compound puree beverages still relies heavily on traditional, experience-based methods. While this model has certain advantages in terms of practical experience, its limitations are also obvious: on the one hand, human experience cannot fully cover the complex nutritional combinations and flavor harmonies required; on the other hand, facing increasingly diverse health demands from consumers, the traditional model falls short in terms of innovation efficiency and precision. This undoubtedly presents new challenges for the further development of the industry.

[0003] In traditional formula development, researchers primarily rely on their subjective understanding of the flavors and components of different raw materials, repeatedly experimenting to adjust the proportions of each ingredient in order to achieve a balance between flavor and health. However, this process is not only time-consuming and costly in terms of raw materials, but also involves a lengthy trial-and-error cycle, resulting in low overall efficiency. Furthermore, due to individual differences in flavor perception and limited quantitative analysis capabilities regarding health components, it is difficult to accurately control the synergistic flavor effects and the rational combination of health components produced by mixing different raw materials. These limitations lead to formulas that ultimately exhibit insufficient stability, failing to meet consumers' dual demands for consistent flavor and standardized health benefits.

[0004] With increasingly segmented consumer groups, different users exhibit significant differences in their demands for beverage flavors (such as sweetness and acidity, and richness) and health attributes (such as the content of specific nutrients). However, traditional beverage development models are inadequate in addressing these personalized needs and struggle to respond quickly. Although some existing technologies have attempted to incorporate data collection and analysis into formula development, these methods are often limited to single-dimensional optimization—for example, focusing solely on flavor tuning or adjusting only health components—failing to achieve multi-objective synergistic optimization of both flavor and health. Furthermore, these technologies lack effective integration with user preference data, resulting in an inability to accurately match users' personalized needs. Ultimately, this leads to formulas with weak market adaptability, failing to meet the demands of a diversified market.

[0005] Currently, raw material data management is largely fragmented, lacking a unified database to integrate and store flavor characteristic and health component data from various raw material pulps. This situation leads to low data retrieval efficiency during R&D, hindering efficient data reuse and in-depth analysis, thus restricting the intelligent and precise progress of formula development. Therefore, overcoming the limitations of traditional manual experience and constructing an automated formula generation method that integrates raw material data, achieves multi-objective optimization of flavor and health, and incorporates user preferences has become a key requirement for promoting the development and upgrading of "Five Tiger Divine Pulp" and similar compound raw material pulp beverage formulas. Summary of the Invention

[0006] The purpose of this invention is to provide an automated method for generating beverage recipes based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an automated beverage recipe generation method based on artificial intelligence, the method comprising:

[0008] Flavor characteristic data of five raw materials—raspberry, mulberry, goji berry, pomegranate, and purple grape—as well as health component data including basic nutrients, plant compounds, and their derivatives, were collected to establish a structured raw material database.

[0009] By using machine learning models to extract features from database data, flavor vectors and health component vectors of each ingredient are generated, thereby achieving quantitative representation of the data.

[0010] An optimization algorithm is used to perform multi-objective optimization on the two types of vectors, and a set of candidate formulas that simultaneously meet the preset flavor threshold and health threshold is selected.

[0011] The candidate formulas were further screened by combining user preference data, and the Five Tiger Divine Elixir formula was finally determined and output to the formula execution system.

[0012] Preferably, the collection of flavor characteristic data and health component data of raspberry puree, mulberry puree, wolfberry puree, pomegranate puree, and purple grape puree includes:

[0013] Data on volatile flavor compounds were extracted from raspberry pulp, mulberry pulp, wolfberry pulp, pomegranate pulp, and purple grape pulp using gas chromatography-mass spectrometry.

[0014] Data on polyphenols, vitamins, and antioxidants were extracted from raspberry pulp, mulberry pulp, wolfberry pulp, pomegranate pulp, and purple grape pulp using high-performance liquid chromatography.

[0015] The data on volatile flavor compounds, polyphenols, vitamins, and antioxidants are normalized and then stored in the raw material database.

[0016] Preferably, the feature extraction of flavor characteristic data and health component data from the raw material database using a machine learning model includes:

[0017] The dimensionality reduction of the volatile flavor compound data is performed using a deep neural network to generate flavor vectors for each raw material.

[0018] The random forest algorithm is used to select features from the polyphenol, vitamin, and antioxidant data to generate a vector of the health components of each raw material;

[0019] The flavor vector and health component vector are associated with the corresponding raw material entries in the raw material database.

[0020] Preferably, the multi-objective optimization of the flavor vector and health component vector based on the optimization algorithm includes:

[0021] A multi-objective optimization function is constructed with the goal of maximizing flavor harmony and health benefits;

[0022] A genetic algorithm is used to iteratively solve the multi-objective optimization function to generate a set of candidate formulas that meet preset flavor and health thresholds;

[0023] The proportions and weights of each raw material are dynamically adjusted during the iteration process to ensure the diversity of the candidate formulation set.

[0024] Preferably, the step of filtering the candidate recipe set based on user preference data includes:

[0025] Obtain users' historical selection data and analyze their preferences for sweetness, acidity, and healthy ingredients;

[0026] The matching degree between the candidate recipe set and the user preference data is calculated using a collaborative filtering algorithm;

[0027] The candidate formula with the highest matching degree was selected as the final Five Tiger Divine Elixir formula.

[0028] Preferably, the method further includes:

[0029] Real-time monitoring of changes in flavor characteristics and health component data of each raw material in the raw material database;

[0030] When a change in data is detected, the machine learning model is triggered to regenerate the flavor vector and the health component vector.

[0031] Multi-objective optimization and user preference screening were re-performed based on the updated flavor vector and health ingredient vector.

[0032] Preferably, the real-time monitoring of changes in the flavor characteristic data and health component data of each raw material in the raw material database includes:

[0033] The freshness and component stability indicators of each raw material are collected through IoT sensors.

[0034] If the freshness index or component stability index exceeds the preset range, the corresponding raw material data will be marked as pending update.

[0035] Feature extraction and optimization calculations are only re-performed on raw material data marked as pending updates.

[0036] Preferably, the method further includes:

[0037] After generating the final Five Tiger Divine Elixir formula, record the flavor characteristics and health component data of the Five Tiger Divine Elixir formula.

[0038] The recorded flavor profile data and health component data are fed back into the machine learning model to optimize the subsequent feature extraction process.

[0039] Preferably, the method further includes:

[0040] Establish a recipe generation log to store the input data, intermediate vectors, and optimization parameters for each generation of the Five Tiger Divine Elixir recipe;

[0041] The flavor and health contributions of each ingredient are analyzed based on the recipe generation log.

[0042] The initial weights of each ingredient in the raw material database are dynamically adjusted based on its flavor and health contributions.

[0043] Preferably, the method further includes:

[0044] After the Five Tiger Divine Elixir formula is executed in the formula execution system, sensory evaluation data and component detection data of the actual finished product are collected.

[0045] The sensory evaluation data and component detection data are compared with theoretical data to calculate the formulation execution deviation;

[0046] If the deviation from the formula exceeds the allowable range, the process of regenerating the Five Tiger Divine Elixir formula will be triggered.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] By systematically collecting flavor characteristic and health component data of raspberry pulp, mulberry pulp, goji berry pulp, pomegranate pulp, and purple grape pulp, and constructing a unified raw material database, the system achieves integrated storage and standardized management of core information from various pulps. Compared to the scattered and inconvenient access to raw material data in traditional development, the unified raw material database can centrally archive various data sets, facilitating subsequent data analysis and reuse. This enables rapid acquisition of necessary raw material information during the R&D process, avoiding R&D interruptions or information deviations caused by fragmented data, and providing a comprehensive and reliable data foundation for formula development.

[0049] By leveraging machine learning models to extract features from data in the raw material database, flavor vectors and health component vectors for each raw material are generated. This transforms previously difficult-to-quantify flavor characteristics and complex health component information into a calculable and analyzable digital form. This digital representation method overcomes the limitations of traditional development methods that rely on subjective human perception. It can more accurately capture the flavor characteristics and health component composition of different raw materials, clearly presenting the core attributes of each raw material. This provides precise quantitative evidence for subsequent formula optimization, making the raw material characteristic analysis in the formula development process more objective and scientific.

[0050] This method utilizes an optimization algorithm to perform multi-objective optimization of flavor and health component vectors, simultaneously addressing both the flavor and health requirements of the Five Tiger Divine Elixir. Traditional development often struggles to balance these two aspects, frequently resulting in either prioritizing flavor at the expense of health, or vice versa. This new method, by setting preset flavor and health thresholds, allows the optimization algorithm to automatically select a set of candidate formulations that meet these threshold requirements. This ensures that the generated candidate formulations possess both excellent flavor and meet the expected health standards, achieving synergistic optimization of flavor and health, and improving the overall quality of the formulation.

[0051] This method incorporates user preference data to filter the candidate formula set, enabling the final Five Tiger Divine Elixir formula to better meet the personalized needs of different users. Different consumer groups have varying preferences for flavors such as sweetness and acidity, and richness, as well as different levels of attention to specific nutritional components. By integrating user preference data into the screening process, the formula that best meets the needs of the target users can be accurately selected from the candidate formulas, enhancing the formula's market adaptability and allowing Five Tiger Divine Elixir to better meet the needs of segmented consumer groups, thereby improving the product's market competitiveness.

[0052] The entire formula generation process is highly automated, from raw material data collection and feature extraction to multi-objective optimization, formula screening, and finally, the output of the formula to the execution system, all without extensive manual intervention. This not only significantly shortens the formula development cycle and reduces the time and raw material cost waste caused by manual trial and error, but also reduces reliance on the subjective experience of R&D personnel, improving the stability and efficiency of formula development. At the same time, the automated process makes the formula development process highly replicable, facilitating subsequent adjustments based on changes in raw material characteristics or market demands, and quickly generating new, suitable formulas. This provides strong technical support for the standardized and large-scale production of Wuhu Shenjiang (Five Tiger Divine Noodles), driving the R&D and production of Wuhu Shenjiang and similar compound pulp beverages towards a more intelligent and efficient direction. Attached Figure Description

[0053] Figure 1 A feature analysis diagram of the automated recipe generation process;

[0054] Figure 2 A flowchart for feature extraction using a machine learning model;

[0055] Figure 3 A flowchart for monitoring and updating raw material data changes;

[0056] Figure 4 Diagram illustrating the feedback learning and continuous optimization mechanism. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1This invention provides an automated beverage formula generation method based on artificial intelligence. The method includes: First, data collection. For five core ingredients—raspberry puree, mulberry puree, goji berry puree, pomegranate puree, and purple grape puree—flavor characteristic data and health component data are systematically collected. The collected data, after standardization, is stored in a structured ingredient database, which forms the data foundation of the entire method. Next, a feature extraction stage is performed. A preset machine learning model is used to conduct in-depth analysis of the raw data in the ingredient database. The machine learning model extracts flavor vectors representing the core flavor attributes of the ingredients from the complex flavor characteristic data, and simultaneously extracts health component vectors reflecting the functional value of the ingredients from the health component data. These vectors are low-dimensional quantitative representations of the ingredient characteristics. Then, multi-objective optimization is performed. Based on an optimization algorithm, the system aims to generate a formula with harmonious flavor and maximized health benefits. The flavor vectors and health component vectors obtained in the previous steps are collaboratively calculated. Under the premise of meeting preset flavor and health thresholds, the optimization algorithm explores different combinations of ingredient ratios, thereby generating a candidate formula set containing multiple feasible solutions. Finally, personalized selection and output are performed. Based on the user preference data input from the outside, the system evaluates the matching degree of each scheme in the candidate formula set, selects the formula that best meets the user's personalized needs as the final Five Tiger Divine Nectar formula, and outputs the formula in a specific format to the downstream formula execution system to drive the automated production equipment to complete the beverage preparation.

[0059] Example 1: See Figure 2 The data acquisition process began with a systematic analysis of five core raw materials: raspberry pulp, mulberry pulp, goji berry pulp, pomegranate pulp, and purple grape pulp. The goal of this stage was to obtain comprehensive and quantifiable flavor and component information. Gas chromatography-mass spectrometry (GC-MS) was deployed for flavor characteristic data extraction. This technique efficiently separates volatile compounds in the raw material samples using a gas chromatography module. The separated components are then sequentially ionized and analyzed by a mass spectrometer. Each pulp sample underwent rigorous pretreatment, such as enriching its volatile components using headspace solid-phase microextraction (HSP). Separation was then performed under predetermined chromatographic conditions. Mass spectra were compared with standard spectral libraries to qualitatively and quantitatively identify compounds such as ketones unique to raspberries, ester aroma components in mulberries, faint aldehydes in goji berries, complex olefins in pomegranates, and abundant alcohols in purple grapes. Each analysis requires multiple technical repetitions to calculate the relative standard deviation of the peak area of ​​each flavor compound. Only data of compounds with stable retention times and relative standard deviations below a preset threshold are recorded, thereby minimizing the impact of instrument errors and operational fluctuations. Ultimately, a detailed dataset containing hundreds of volatile compounds and their relative contents is generated for each raw material.

[0060] The collection of health component data relies on high-performance liquid chromatography (HPLC), a technique particularly suitable for analyzing substances that are not easily volatile but beneficial to health. For the detection of polyphenols, an HPLC system equipped with a photodiode array detector is typically used, eluting and measuring samples at a specific wavelength (e.g., 520 nm for anthocyanin detection). For vitamins and antioxidants, a fluorescence detector may be used to achieve higher sensitivity. In practice, for vitamins, such as vitamin C or B vitamins, the system is configured with a fluorescence detector for quantitative analysis. The fluorescence detector utilizes the fluorescence characteristic of vitamins under photoexcitation by setting specific excitation and emission wavelengths, thereby achieving highly selective detection of trace components. The sample pretreatment stage must ensure the purity of the extraction solvent to avoid fluorescence quenching, and establish a linear relationship between fluorescence intensity and concentration through calibration with standards to ensure the reliability of the detection results. For antioxidants, such as some polyphenols or flavonoids, the system uses an electrochemical detector for determination. The electrochemical detector is based on the principle of redox reactions of analytes on the electrode surface. It achieves sensitive detection of electroactive substances by controlling the potential of the working electrode. During implementation, the mobile phase composition and pH value need to be optimized to enhance the electrochemical response, and a three-electrode system (working electrode, reference electrode, and counter electrode) is used to ensure potential stability. Before detection, the optimal detection potential is determined using cyclic voltammetry and quantitatively validated using a standard curve to improve accuracy and repeatability. The entire detection process needs to be coordinated with sample pretreatment steps, including accurate weighing of raw material samples, ultrasonic-assisted extraction using a suitable solvent, centrifugation to remove impurities, and filtration through a membrane to obtain a clear analyte. After the analytical method is established, systematic methodological validation is required, covering linearity testing, intra-day precision assessment, and spiked recovery experiments to ensure that the detection method can stably quantify the content of vitamins and antioxidants in different raw materials. Normalized data is stored in a raw material database to provide standardized input for subsequent feature extraction. Sample pretreatment includes precise weighing, solvent extraction, centrifugation, filtration, and dilution to ensure effective extraction of the analytes without introducing interfering substances. The establishment of the analytical method requires systematic methodological validation, including linear range, precision, and spiked recovery experiments, to confirm that the method can reliably quantify the content of carotenoids in wolfberry, ellagic acid in pomegranate, resveratrol in purple grapes, and vitamin C found in various raw materials. All detection data are recorded in standard concentration units, forming a health component profile for each raw material.

[0061] The raw data collected vary in magnitude and unit due to their different sources, and direct comparison or calculation will introduce bias; therefore, normalization is necessary. For volatile flavor compound data obtained from gas chromatography-mass spectrometry (GC-MS), area normalization is typically used to convert the peak area of ​​each compound on the total ion current chromatogram into a percentage content relative to the total peak area. This eliminates the influence of minor differences in injection volume, making the flavor intensity of different batches and raw materials comparable. For polyphenol and vitamin content data obtained from high-performance liquid chromatography (HPLC), since these are absolute concentration values, a minimum-maximum normalization method is used. This linearly transforms the concentration value of each component to the interval [0,1]. The maximum and minimum values ​​used for the transformation are derived from the historically accumulated maximum and minimum values ​​of that component among all raw materials, thus assigning a consistent data scale to the processed data. After normalization, the data on volatile flavor compounds, polyphenols, vitamins, and antioxidants are structured and stored in specific data tables in the raw material database. Each record corresponds to a complete analysis of a single raw material and includes metadata such as sample number, collection time, and analysis conditions.

[0062] After the raw material database is populated with valid data, the feature extraction process begins. This process utilizes machine learning models to extract essential features from the high-dimensional raw data. For flavor feature data, deep neural networks are chosen as the feature extraction tool. The network structure is typically designed as an autoencoder architecture containing an input layer, multiple hidden layers, and a bottleneck layer. The number of neurons in the input layer corresponds to the number of detected volatile compounds. Normalized compound content data is fed into the network as an input vector. During forward propagation, the network learns the complex interactions and combination effects between compounds through the nonlinear activation functions in the hidden layers. Finally, the bottleneck layer, with a dimension much lower than the input layer, outputs a low-dimensional, dense real-valued vector, i.e., the flavor vector. Each dimension of this vector does not directly correspond to a specific compound but represents an abstract feature learned by the network that comprehensively reflects the overall flavor profile of the raw material. Training such a network requires a large amount of raw material sample data. The backpropagation algorithm minimizes the error between the input data and the network's reconstructed output, enabling the network to master the ability to compress and express core flavor features from a detailed list of compounds.

[0063] For processing health component data, a random forest algorithm is used for feature selection to generate health component vectors. The random forest consists of a large number of decision trees, each grown using a bootstrap sample set of the training data, and splitting at each node considering only a subset of randomly selected features. After training, the algorithm calculates the importance score for each health component feature, based on the sum of the impurity reductions brought about by that feature when used to split nodes across all trees. Highly important features mean that the component contributes more to distinguishing the health attributes of different raw materials or predicting their functional value. The system sets an importance score threshold or selects the top K features to filter out a subset of key health components. For each raw material, the normalized content values ​​of its key health components are arranged in a fixed order to form the health component vector for that raw material. This method not only reduces the dimensionality of the data but also highlights the components that contribute the most to health benefits, effectively filtering out redundant components with limited information.

[0064] The generated flavor and health component vectors need to establish a robust association with the original entries in the raw material database. In the database design, two extended tables can be created for each raw material, one for storing its historical flavor vector and the other for storing its health component vector. Each vector record includes a timestamp and the original data batch number used to generate the vector, and is associated with the raw material's master information table through its unique identifier. When new analytical data is entered into the database and triggers feature extraction, the newly generated vector is added as a new record to the corresponding extended table, and can be marked as the current valid version. This design allows the system to track the evolution of raw material characteristics over time and facilitates the rapid retrieval and calling of the latest vectorized information for any raw material in subsequent optimization calculations. The database management system, by establishing appropriate indexes, ensures that vector queries based on raw material identifiers remain efficient even in the context of massive amounts of data. The entire data acquisition and feature extraction process constitutes a rigorous data preprocessing pipeline, starting from basic instrumental analysis, through data cleaning and standardization, and then to intelligent feature dimensionality reduction and filtering using machine learning, ultimately transforming the complex sensory and health attributes of the five natural raw materials into structured mathematical representations that can be efficiently processed by computers.

[0065] Example 2: The multi-objective optimization process relies on the flavor vector and health component vector generated in the previous feature extraction stage. These vectors encapsulate the core attributes of each ingredient in numerical form. The primary task of the optimization algorithm is to construct a mathematical framework that can simultaneously measure the flavor performance and health benefits of the formulation. This framework needs to define a clear optimization objective, typically set as finding the optimal solution that combines flavor harmony and health benefits under given constraints. The first objective function is the flavor harmony objective function, used to quantify the degree of flavor harmony in the blended formulation. Its specific form is:

[0066]

[0067] in, Representing the The weight of each raw material in the formula according to its mass percentage should meet the requirements. , Corresponding to raspberry pulp, Corresponding to mulberry puree, Corresponding to wolfberry juice, Corresponding to pomegranate pulp, Corresponding to purple grape puree; Representing the The flavor vectors of the raw materials are generated by dimensionality reduction using a deep neural network. 3D real vector; The ideal flavor vector is represented by a pre-defined vector, which is constructed by surveying the flavor characteristics of mainstream compound beverages on the market and combining the scoring data of a professional tasting team. The score represents the flavor harmony, and the value range is [value range missing]. The closer the value is to 1, the more harmonious the flavor of the recipe is with the ideal flavor.

[0068] The second objective function is the health benefit objective function, used to evaluate the overall health value of the formula. Its specific form is as follows:

[0069]

[0070] in, Representing the The first of the raw materials The content of key health components was obtained by high performance liquid chromatography detection and normalization; Representing the The weighting coefficients for each health-promoting ingredient are set based on its nutritional value; for example, the weighting coefficients for core antioxidants such as anthocyanins and resveratrol are determined by their respective nutritional values. The value is higher than that of basic vitamins; The total number of types representing key health components; The score represents the health benefits, and its value range is [value range missing]. A higher value indicates a higher health value of the formula. The constraints for multi-objective optimization are: and ,in and These are preset flavor and health thresholds, determined by those skilled in the art based on product positioning. The first objective function revolves around flavor harmony. Its core idea is to evaluate whether the overall flavor profile formed by mixing ingredients in different proportions is harmonious and meets preset standards. The calculation requires treating each candidate formulation as a weighted sum of the flavor vectors of each ingredient, with the weight representing the proportion of that ingredient in the formulation. This synthesized overall flavor vector is then compared with a predefined ideal flavor profile. Similarity can be measured using methods such as cosine similarity or the reciprocal of Euclidean distance. The second objective function focuses on maximizing health benefits. Its calculation is based on a weighted combination of the health component vectors of each ingredient, with the weight also being the proportion. The final result is a comprehensive scalar representing the overall health value of the formulation. This scalar can be further refined into sub-objectives targeting specific health needs, such as total antioxidant capacity or the total content of a specific vitamin group. These two objectives often have an inherent conflict. For example, increasing the content of certain health components may require increasing the amount of a specific ingredient, which may have a strong flavor profile and thus affect the overall harmony. Therefore, optimization algorithms must be able to handle this trade-off and explore the so-called Pareto optimal frontier, which is the solution set that cannot improve the other objective without sacrificing one objective.

[0071] Genetic algorithms are chosen as the primary tool for solving this multi-objective optimization problem due to their powerful global search capabilities and good adaptability to non-convex and nonlinear problems. The initialization phase of the algorithm involves randomly generating a large population of hundreds of individuals, each representing a complete recipe for the Five Tiger Divine Elixir. Their chromosomes are typically encoded using real numbers, directly representing the percentage proportions of the five elixir ingredients (raspberry, mulberry, goji berry, pomegranate, and purple grape) in a five-dimensional vector, with the sum of all dimensions equal to 100%. The initial population covers the broadest possible ratio space, providing a rich diversity foundation for subsequent evolution. In the evaluation phase, the recipe encoded by each individual's chromosome is decoded and substituted into the aforementioned multi-objective function for calculation, yielding its fitness value for both flavor harmony and health benefits. The fitness value directly reflects the quality of the recipe. Selection is based on the fitness value, favoring individuals that perform better overall on both objectives. Common strategies include roulette wheel selection or tournament selection, aiming to pass on superior genes to the next generation. Crossover simulates gene recombination in biological reproduction. Two parent individuals are randomly selected from the population, and their chromosomes are partially exchanged through arithmetic crossover or simulated binary crossover, resulting in offspring with characteristics of both parents. Mutation, on the other hand, randomly alters the proportions of one or more genetic materials in an individual's chromosomes with a small probability. This operation injects new genetic material into the population, helping the algorithm escape local optima and avoid premature convergence. The iterative process repeats selection, crossover, mutation, and evaluation, evolving the population generation after generation. Each generation generates new candidate solutions and eliminates solutions with low fitness.

[0072] Maintaining diversity in the candidate formula set is a key challenge in the iterative solution process of genetic algorithms. A lack of diversity can cause the algorithm to converge prematurely to a local optimum, thus missing better global solutions. Dynamically adjusting the weights of each ingredient is one of the core methods for maintaining diversity, and this adjustment can be embedded into the operations of the genetic algorithm. For example, after crossover, the weights of offspring individuals can be slightly and randomly perturbed. Or, in mutation, not only can the weights of individual ingredients be changed, but their mutation probabilities can also be dynamically adjusted based on their frequency of occurrence in the current population. For ingredients with excessively high frequency of occurrence, their mutation intensity can be appropriately increased to encourage the exploration of new solutions containing fewer of these ingredients or combinations of other ingredients. Another strategy is to use niche techniques. This technique introduces a similarity penalty between individuals in the fitness evaluation to prevent overly similar individuals from over-proliferating in the population, thereby encouraging the population to be dispersed across different segments of the Pareto front. The iterative process continues until a preset termination condition is met, such as reaching the maximum number of iterations, or the improvement of the optimal solution set in several consecutive generations being less than a certain threshold. Ultimately, the algorithm outputs a non-dominated solution set, namely the candidate recipe set. Each recipe in the set satisfies the preset minimum flavor threshold and health threshold, and they cannot be easily judged as superior or inferior to each other. Together, they constitute the specific manifestation of the Pareto optimal frontier at different trade-off points.

[0073] After obtaining the candidate recipe set, the system focuses on refining the selection based on user preference data. User preference data can be obtained from various sources, including historical user selection data accumulated over time. This data can be analyzed by examining users' past ratings, purchase records, or completion rates of different recipes to infer their potential preferences. Alternatively, it can be user-submitted questionnaires designed to cover their tolerance for basic tastes such as sweetness, acidity, and bitterness, as well as their level of interest in specific health benefits (such as sleep aid, energy boost, and vitamin supplementation). Collaborative filtering is the core algorithm for calculating the matching degree, and its working principle is based on the interaction history between users and items (in this case, recipes). This application uses a user-based collaborative filtering algorithm to calculate the matching degree between candidate recipes and target users. The specific steps are as follows: User similarity calculation and selection of target users. and other users in the user set Collect all user ratings for historical recipes. ,in Indicates user For the formula The rating (ranging from 1 to 5 points, with higher scores indicating greater liking). Calculated using the Pearson correlation coefficient. and Similarity:

[0074]

[0075] in, for and A collection of recipes that have been jointly rated. for Historical average rating For target users Historical formula The actual score, For other comparison users Historical formula The actual score, for The historical average rating. The range of values ​​is A positive value and a larger value indicate more similar user preferences. The calculation of predicted preference levels first filters out those similar to... Similarity Users as similar user groups For recipes in the candidate recipe set ,calculate Estimated level of liking for it:

[0076]

[0077] in, For the estimated rating, the value ranges from 1 to 5 points. A higher score indicates that the target users are more likely to like the recipe. Representing similar user groups Any user in For a particular candidate formulation in the candidate formulation set The actual score (score range 1-5 points), if The candidate formulation was not specified. Actual ratings were conducted, and the average rating of the candidate recipes from similar user groups was used as a complement. Matching ranking then sorted all recipes in the candidate recipe set according to... Sort the recipes in descending order and select the top-ranked one as the final Five Tiger Divine Elixir recipe. If user-based collaborative filtering is used, the algorithm seeks other user groups with similar preference histories to the current user, calculates the estimated preference level of this similar user group for each recipe in the candidate recipe set, and uses this as the match between the current user and the recipe. If item-based collaborative filtering is used, the algorithm calculates the similarity between each recipe in the candidate recipe set and those recipes the current user has historically liked. The similarity between recipes can be obtained by comparing the distance between their corresponding flavor vectors and health component vectors. Finally, the candidate recipe most similar to the recipes the user has liked is recommended. The match is usually quantified as a score between 0 and 1; a higher score indicates that the candidate recipe better meets the user's personalized needs.

[0078] Selecting the candidate formula with the highest matching degree as the final output is a decision-making process. The system sorts all formulas in the candidate formula set in descending order of their matching degree scores calculated by a collaborative filtering algorithm. The formula ranked first is determined as the final Five Tiger Divine Elixir formula. This formula not only achieves a good balance between flavor and health in the sense of multi-objective optimization, but more importantly, it selects the one that best matches the specific user's taste and health expectations among many balance points, thus realizing personalized customization of the formula. Before output, the system formats the proportioning data of the final formula, such as converting it into raw material dosages accurate to grams or milliliters, and includes necessary process parameter suggestions (such as mixing order and stirring time), generating a complete production instruction file. This instruction file is reliably transmitted to the downstream formula execution system through application programming interfaces or message queues, driving automated metering, mixing, and filling equipment to complete the actual preparation of the beverage.

[0079] Example 3: See Figure 3The core capability of the system in maintaining the accuracy of formula generation lies in its dynamic monitoring and response mechanism for the raw material database. This mechanism is designed to address data drift issues caused by natural variations in raw materials or batch changes. Real-time monitoring relies on an IoT sensor network deployed in the raw material storage area and pretreatment stages. This network consists of various types of micro-sensor nodes that continuously collect physicochemical parameters directly related to raw material quality. For assessing the freshness of raw pulps such as raspberry and mulberry, sensors primarily monitor their color spectral characteristics, viscosity, pH value, and core temperature changes. For example, near-infrared spectroscopy probes can detect color changes in the pulp non-contactly, while rotational viscometers can monitor its rheological properties online. The combined trends of these parameters effectively reflect the oxidation level and microbial activity of the raw materials. Monitoring component stability requires more precision and may require electrochemical sensors or specific ion-selective electrodes embedded in the raw material storage tank to track the concentration decay trend of easily oxidized components such as vitamin C, or to monitor the relative stability of total phenol content. All sensor nodes transmit the collected time-series data streams to the central data platform in real time via a low-power wide-area network or the ZigBee protocol.

[0080] Upon receiving streaming data, the central data platform immediately initiates a data quality assessment and status determination process. This process is based on preset safe fluctuation ranges for each raw material and each monitoring indicator. These ranges are not fixed but dynamically set through statistical analysis of historically stable batch data, typically centered on the mean and bounded by several standard deviations. The platform's built-in rule engine continuously compares real-time sensor readings with the corresponding preset ranges. For example, if the pH value of a batch of goji berry juice remains below its lower limit threshold L_pH for three consecutive sampling periods, or if the intensity decay rate of a characteristic peak in its color spectrum exceeds the threshold R_color, the rule engine generates an anomaly event. This event triggers a database update, automatically changing the batch's status from "valid" to "pending update" and recording the time of the anomaly trigger, the specific indicator, and the degree of deviation. This status labeling is a lightweight, non-destructive operation; it does not immediately erase or overwrite existing flavor and component data but instead assigns a label requiring re-verification.

[0081] Once a raw material is marked as "pending update," the system does not immediately initiate a complete re-collection and calculation of all data. Instead, it executes an efficient incremental update strategy. The system's monitoring module sends an asynchronous message to the feature extraction module, notifying it of the pending raw material data update request. This request includes the unique identifier of the raw material to be updated and the specific indicator information that triggered the update. Upon receiving the request, the feature extraction module initiates a targeted feature recalculation task. This task is highly selective; it only operates on raw material entries with the "pending update" status. The module retrieves the latest complete analysis data (gas chromatography-mass spectrometry and high-performance liquid chromatography) from the database for that raw material. However, this processing incorporates the sensor information that triggered the update. For example, if an abnormal pH value indicates increased acidity, the algorithm will assign different weights to or correct for acid-related compounds when generating the flavor vector. Deep neural networks and random forest algorithms, based on possibly fine-tuned parameters or corrected data considering freshness decay models, rerun the feature extraction process to generate an updated set of flavor vectors and health component vectors for that raw material. For raw materials whose status is still "valid", their feature vectors are directly retrieved from the database cache to avoid unnecessary computational overhead.

[0082] After obtaining the updated vector set, the multi-objective optimization process is triggered. However, this optimization does not start from scratch but is a hot start based on the results of the previous round of optimization calculations, which can significantly accelerate the convergence process. The initial population of the optimization algorithm (such as a genetic algorithm) will contain most of the effective candidate formulas from the previous round, while introducing some entirely new formula individuals with new vector information generated through mutation operations. This strategy allows the optimization process to quickly focus on exploring the design space region that has changed due to the update of raw material data, without having to repeatedly search in a vast, unchanged space. Similarly, the subsequent user preference screening step will also be re-executed based on the updated candidate formula set, but since the user preference data itself is relatively stable, the computational cost of this step is relatively small. Finally, the system will output a Five Tiger Divine Elixir formula based on the latest raw material state. This formula reflects the true properties of the currently available raw materials, thus ensuring the quality stability of the final product from the source. The entire dynamic adjustment mechanism is like a feedback closed-loop control system. It can sense changes in the environment (raw materials) and actively adjust the system's output (formula) to adapt to these changes, ensuring the robustness and practicality of the formula generation system in long-term operation. This on-demand, incremental update design allows the system to save computing resources to the maximum extent while ensuring timely response.

[0083] To quantify the degree of anomalies in monitoring data and aid decision-making, the system may use a simple deviation metric to help determine the urgency of status marking. This metric can be expressed as:

[0084]

[0085] in: This represents the overall deviation score calculated during the current sampling period. The larger the value, the greater the degree of deviation from the normal state. This indicates the total number of parameters monitored by IoT sensors deployed for the raw material, which may include, for example, pH, color index, viscosity, etc. Indicates the first The actual measured reading of each monitoring parameter at the current moment. and They represent the first number calculated based on historical stable data. The long-term average (mean) and standard deviation of a monitoring parameter together define the normal range of fluctuation of that parameter. This is a pre-set weighting coefficient that reflects the first... Each parameter represents a degree of importance to the overall quality of the raw materials, with higher-importance parameters receiving greater weight. The rule engine continuously calculates these parameters. When the score exceeds a preset critical threshold, the raw material status is decisively marked as "pending update," rather than relying solely on a single parameter threshold. This makes the status judgment more comprehensive and reliable. This internal indicator is a component of the system's intelligent status perception, but it is not directly output to the user; instead, it serves the internal control logic.

[0086] Example 4: The feedback learning and continuous optimization mechanism begins operation after each successful formula generation. Its primary step is to meticulously record the theoretical data for that formula. When the system determines the final Five Tiger Divine Elixir formula and sends it to the execution system, a data archiving routine is initiated. This routine encapsulates and stores the key theoretical data generated during the generation process. The recorded data mainly includes two parts: one part is the theoretical flavor characteristic data corresponding to the final formula. This is not a raw sensory description, but a synthetic flavor vector calculated based on the proportions of each ingredient and their flavor vectors. This vector numerically predicts the overall flavor profile that the finished product should possess, such as expected values ​​in dimensions like sweetness, acidity, and fruit aroma intensity. The other part is the theoretical health component data, also calculated by weighting the health component vectors of each ingredient and their proportions. It predicts the theoretical estimates of health-related indicators such as total polyphenol content and comprehensive antioxidant index in the finished product. This data, along with the formula's unique identifier and generation timestamp, is stored in a dedicated "formula results repository," forming a valuable historical record.

[0087] These recorded theoretical data are then fed into a feedback pipeline to optimize the feature extraction capabilities of subsequent machine learning models. Each successful recipe record in the "Recipe Success Library" is treated by the system as a high-quality, multi-objective optimized, labeled sample. These records are added to the training dataset during periodic model retraining tasks. For example, when a deep neural network relearns how to generate flavor vectors from raw material flavor compound data, it not only uses single-raw material data from the raw material database but also incorporates the synthetic flavor vectors of these successful recipes as part of the supervision signal. This guides the network to learn which flavor combinations are considered "excellent" by both the optimization algorithm and user preferences, thereby fine-tuning its network parameters to make future generated single-raw material flavor vectors more conducive to creating a harmonious overall flavor. Similarly, when re-evaluating the feature importance of health components, the random forest algorithm also refers to the key health components highlighted in these successful recipes, adjusting its feature selection strategy to make the generated health component vectors more relevant to those components that significantly contribute to actual optimization. Operating in parallel with the recipe success records is a recipe generation log system, which records the complete lifecycle data of each recipe generation at a higher granularity. The logs store input data including the original request parameters that triggered the generation task and the version numbers of flavor and component data used for each ingredient in the raw material database at that time. The intermediate vectors in the records encompass the specific numerical sequences of flavor and health component vectors generated for each ingredient in the feature extraction step. The optimization parameters are more detailed, including the initial population size, crossover rate, mutation rate, number of iterations, and a summary of the Pareto front solution set at final convergence used by multi-objective optimization algorithms (such as genetic algorithms). These log entries are linked to records in the "Recipe Results Library" via the generation task ID, forming a traceable and complete data chain.

[0088] Based on the accumulated recipe generation logs, the system can perform in-depth contribution analysis. The analysis program scans all ultimately adopted recipe records in the logs, counting the frequency of each ingredient appearing in these successful recipes. Furthermore, the program analyzes the relationship between the weighting of various ingredients in each adopted recipe and their contribution to the overall flavor vector and health component vector. Through certain statistical algorithms (such as linear regression or feature importance ranking algorithms), the system can calculate the average flavor contribution and health contribution of each ingredient over a historical time interval. Contribution is a relative value that characterizes the importance of the ingredient in achieving excellent flavor harmony and health benefits in historically successful recipes. See Table 1, which presents the analysis results under a hypothetical scenario to illustrate the quantification of contribution.

[0089] Table 1: Raw Material Contribution Analysis

[0090]

[0091] Note: The data in this table are illustrative and show the relative relationship of contribution.

[0092] The results of contribution analysis are directly applied to dynamically adjust the initial weights of each ingredient in the ingredient database. In the ingredient database, each ingredient entry stores its basic attributes and feature vectors, as well as a variable "initial weight" field. This weight does not directly represent the proportions, but rather serves as prior knowledge influencing the behavior of the multi-objective optimization algorithm during population initialization or computation. For example, based on the analysis results in Table 1, the system might set relatively high initial weights for the flavor and health aspects of raspberry puree, as historical data indicates its significant role in successful formulations; while the initial weight of purple grape puree might be set relatively low. When generating a new formulation, the optimization algorithm, while exploring the proportion space, will subconsciously tend to favor ingredients with high historical contribution in its search. This is not a rigid rule, but rather an optimization guide that helps the algorithm converge to a high-quality solution region more quickly, thereby improving the overall efficiency and quality of formulation generation. This dynamic adjustment mechanism allows the system to continuously learn from its successful experiences, evolving its internal decision preferences and forming a virtuous cycle of becoming increasingly intelligent with use.

[0093] See Figure 4 Subplot (a) shows the results of the ingredient contribution analysis based on 150 historical successful recipe records. The bar chart clearly shows the relative importance of each ingredient in terms of both flavor and health contribution. The data shows that raspberry puree has the highest contribution in terms of flavor, indicating that it plays a key role in forming a superior flavor profile in historically successful recipes. Mulberry puree performs best in terms of health contribution, reflecting its important role in enhancing the health value of the recipe. Goji berry puree also performs well in terms of health contribution, consistent with the traditional understanding of the health value of goji berries. This contribution data provides a scientific basis for the system's dynamic weight adjustment. Subplot (b) shows the trend of the system dynamically adjusting the initial weights of ingredients based on the contribution analysis results. Over time, the system continuously learns from successful experiences and gradually adjusts the initial weights of each ingredient in the optimization algorithm. For example, the initial weights of raspberry puree and mulberry puree, which have high flavor and health contributions respectively, show a steady upward trend, meaning that the optimization algorithm pays more attention to these historically high-performing ingredients when searching the recipe space. Conversely, the weights of raw materials with relatively low contributions will be appropriately reduced, but the system will still maintain a certain degree of exploratory nature to avoid getting trapped in local optima.

[0094] This dynamic weight adjustment mechanism reflects the system's self-learning and continuous optimization capabilities. By continuously accumulating successful formulation experience, the system can increasingly intelligently guide the formulation generation process, improving efficiency and quality, forming a virtuous cycle of becoming smarter with use. This mechanism ensures that the system can adapt to changes in raw material characteristics and evolving user preferences, maintaining the advanced nature and practicality of formulation generation in the long term.

[0095] Example 5: Verification and closed-loop control of formula execution effectiveness begins at the end of the automated production line. After the formula execution system completes the physical preparation of the beverage according to the received digital formula, the actual finished product enters the quality assessment stage. Collecting sensory evaluation data of the actual finished product requires a standardized process, usually conducted by a trained evaluation team in a controlled environment. The evaluation room has suitable lighting and is free from odor interference, and each evaluator evaluates the sample independently. The evaluation uses a structured scale, with scale dimensions closely corresponding to the attributes described by the flavor characteristic data of the formula theory. For example, for the Five Tiger Divine Nectar, each dimension may include color saturation, raspberry and mulberry aroma intensity, sweet and sour balance, aftertaste from goji berries, and overall smoothness of the mouthfeel. Each dimension has a clear scoring anchor point (e.g., a 1-9 scale, from "extremely weak" to "extremely strong"). The tasters scored the products on the scale based on their actual tasting experience. After all the tasters' scores were collected, a consistency test was performed, invalid scores that significantly deviated from the group were removed, and the average score for each sensory dimension was calculated to form a quantitative dataset describing the sensory characteristics of the actual finished product.

[0096] The acquisition of component testing data relies on the physicochemical analysis of samples from the same batch of finished products. Sampling follows a random principle to ensure the representativeness of the samples. The sampled finished products are sent to the laboratory and analyzed using the same analytical methods as the raw materials, such as high-performance liquid chromatography (HPLC) to detect the actual content of key health components like anthocyanins and vitamin C in the finished product. The laboratory analysis report will provide precise concentration data for these components, representing the true level of health components in the Wuhu Shenjiang finished product. Both sensory evaluation data and component testing data must be collected following strict standard operating procedures to minimize the impact of human error and environmental fluctuations on the measurement results, ensuring the reliability and comparability of the obtained data. Next, the system initiates a comparative analysis process, comparing the actual collected data with the theoretical data from the formulation generation stage item by item. The comparison of sensory evaluation data is not a simple subtraction of scores; rather, it requires mapping and comparing the average score vector of the actual finished product with the theoretically predicted flavor vector. The theoretical flavor vector may also need to be converted to a similar sensory intensity expected score scale. For example, the theoretically predicted "mulberry aroma intensity" is 8.5 points (assuming a maximum score of 9), while the actual average score is 7.8 points, indicating a sensory bias of 0.7 points for this attribute. The comparison of component testing data is even more direct. For instance, the theoretically calculated total polyphenol content is 250 mg / L, while the laboratory measured value is 238 mg / L, resulting in an absolute deviation of 12 mg / L. The system presets an acceptable deviation range for each compared attribute and component. This range is determined based on the uncertainty of the measurement method, the natural fluctuations of raw materials, and the inherent fluctuations of the production process. Based on the deviations at all comparison points, the system calculates a comprehensive formulation execution deviation value. This calculation does not simply average all deviations but may introduce weights to differentiate the importance of different indicators. For example, the deviation weights for core flavor attributes (such as sweet-sour balance) and key health components (such as signature active substances) will be higher than those for secondary indicators. One possible calculation method is to first calculate the relative deviation or standardized deviation of each comparison point, then sum them according to their weights to obtain a normalized deviation score between 0 and 1. The closer the score is to 0, the better the conformity; the closer it is to 1, the greater the deviation. This comprehensive deviation score is a quantitative assessment of the overall effectiveness of the formulation's transformation from theory to practice.

[0097] The system has an internal threshold for the allowable range of formula execution deviation. This threshold is a key parameter set after careful evaluation. If the calculated overall deviation score does not exceed the allowable range, the formula generation and execution are considered successful, and the theoretical design and actual output are basically consistent. The system will package and archive all data from this cycle (including theoretical values, actual values, and deviation values) as a valid run record in the database. This successful data can also provide positive samples for subsequent model optimization. If the overall deviation score clearly exceeds the preset allowable range, the system's anomaly handling and closed-loop control mechanism is triggered. The activation of this mechanism means that a significant difference has occurred in the transfer process from theory to practice. There are many possible reasons for deviation exceeding the limit, including but not limited to: the actual flavor or component characteristics of the batch of raw materials used have drifted significantly from the data stored in the raw material database but have not been captured by the real-time monitoring system; there are accuracy deviations or mechanical wear in the formula execution system during metering, mixing, homogenization, etc.; and the theoretical prediction model itself has inherent prediction errors under certain complex interactions.

[0098] Once a recipe needs to be regenerated, the system doesn't simply repeat the previous process; instead, it initiates an enhanced generation loop. The system checks the latest monitoring data in the raw material database, especially for ingredients associated with significant deviations in sensory or component attributes, to ensure the data is up-to-date. Sometimes, the regeneration process mandates supplementary testing of key ingredients in the current batch to obtain the most accurate data to update their feature vectors. Using the updated data, the system re-executes the entire process from feature extraction to user preference filtering. During optimization, the algorithm may consider the direction of the previous deviation; for example, if the actual sweetness is lower than the theoretical value, the new optimization round might slightly favor ingredients or proportions that contribute more sweetness as a compensatory search strategy. The newly generated recipe is sent back to the recipe execution system for trial production, and the verification process described above is repeated, forming a closed loop of "generation-execution-verification-feedback-regeneration" until the overall deviation of the final product falls within an acceptable range. This closed-loop control mechanism greatly enhances the system's adaptability and robustness, enabling the AI ​​formula generation system not only to perform theoretical optimization design, but also to ensure that its design results can be reliably reproduced in complex actual production environments. This tightly integrates intelligent algorithms with actual production, forming a continuously self-calibrating intelligent production system.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating beverage recipes based on artificial intelligence, characterized in that, Includes the following steps: (1) Construct a multidimensional raw material database: Collect flavor characteristic data and health component data of five raw materials: raspberry, mulberry, wolfberry, pomegranate and purple grape, and establish a structured raw material database; (2) Feature extraction and vector generation: The database data in step (1) is used to extract features through a machine learning model to generate flavor vectors and health component vectors for each raw material; (3) Multi-objective optimization screening: The flavor vector and health component vector in step (2) are optimized by multi-objective optimization algorithm to screen the candidate formula set that simultaneously meets the preset flavor threshold and health threshold. (4) Personalized formula determination: The candidate formula set in step (3) is screened a second time based on user preference data to generate the final Five Tiger Divine Elixir formula, and the Five Tiger Divine Elixir formula is output to the formula execution system. Step (3) involves using an optimization algorithm to perform multi-objective optimization on the flavor vector and health component vector from step (2), including: A multi-objective optimization function is constructed with the goal of maximizing flavor harmony and health benefits; A genetic algorithm is used to iteratively solve the multi-objective optimization function to generate a set of candidate formulas that meet preset flavor and health thresholds; The proportions and weights of each raw material are dynamically adjusted during the iteration process to ensure the diversity of the candidate formulation set; The first objective function is the flavor harmony objective function, which is used to quantify the degree of flavor harmony in the blended formulation. Its specific form is as follows: ； in, Representing the The weight of each raw material in the formula according to its mass percentage should meet the requirements. , Corresponding to raspberry pulp, Corresponding to mulberry puree, Corresponding to wolfberry juice, Corresponding to pomegranate pulp, Corresponding to purple grape puree; Representing the The flavor vectors of the raw materials are generated by dimensionality reduction using a deep neural network. 3D real vector; The ideal flavor vector is represented by a pre-defined vector, which is constructed by surveying the flavor characteristics of mainstream compound beverages on the market and combining the scoring data of a professional tasting team. The score represents the flavor harmony, and the value range is [value range missing]. The closer the value is to 1, the more harmonious the flavor of the recipe is with the ideal flavor. The second objective function is the health benefit objective function, used to evaluate the overall health value of the formula. Its specific form is as follows: ； in, Representing the The first of the raw materials The content of key health components was obtained by high performance liquid chromatography detection and normalization; Representing the The weighting coefficients for each health component are set based on its nutritional value. The total number of types representing key health components; The score represents the health benefits, and its value range is [value range missing]. The higher the value, the higher the health value of the formula; Step (4) involves a secondary screening of the candidate recipe set from step (3) using user preference data, which includes: Obtain users' historical selection data and analyze their preferences for sweetness, acidity, and healthy ingredients; The matching degree between the candidate recipe set and the user preference data is calculated using a collaborative filtering algorithm; The candidate formula with the highest matching degree was selected as the final Five Tiger Divine Elixir formula.

2. The method for automatically generating beverage recipes based on artificial intelligence according to claim 1, characterized in that, Step (1) involves collecting flavor characteristic data and health component data for five types of raw pulp: raspberry, mulberry, wolfberry, pomegranate, and purple grape. Data on volatile flavor compounds, including sweetness, acidity, and aroma intensity, were extracted from raspberry pulp, mulberry pulp, wolfberry pulp, pomegranate pulp, and purple grape pulp using gas chromatography-mass spectrometry. The polyphenol content, vitamin levels, and antioxidant activity of raspberry pulp, mulberry pulp, wolfberry pulp, pomegranate pulp, and purple grape pulp were extracted using high performance liquid chromatography. The data on volatile flavor compounds, polyphenol content, vitamin levels, and antioxidant activity are normalized and stored in the raw material database to ensure the accuracy and comparability of the data.

3. The method for automatically generating beverage recipes based on artificial intelligence according to claim 2, characterized in that, Step (2) involves extracting features from the database data in step (1) using a machine learning model, including: The dimensionality reduction of the volatile flavor compound data is performed using a deep neural network to generate flavor vectors for each raw material. The random forest algorithm was used to select features from the polyphenol content, vitamin level, and antioxidant activity to generate a vector of the health components of each raw material. The flavor vector and health component vector are associated with the corresponding raw material entries in the raw material database.

4. The method for automatically generating beverage recipes based on artificial intelligence according to claim 3, characterized in that, It also includes a dynamic data update process: Real-time monitoring of changes in flavor characteristics and health component data of each raw material in the raw material database; When a change in data is detected, if the change exceeds the preset fluctuation thresholds of ±5% for flavor data and ±3% for health data, the machine learning model is triggered to regenerate the flavor vector and health component vector. Based on the updated flavor vector and health ingredient vector, step (3) multi-objective optimization screening is re-executed to step (4) personalized formulation determination.

5. The method for automatically generating beverage recipes based on artificial intelligence according to claim 4, characterized in that, The changes in flavor characteristic data and health component data of each raw material in the real-time monitoring raw material database include: The freshness and component stability indicators of each raw material are collected through IoT sensors. If the freshness index or component stability index exceeds the preset range, the corresponding raw material data will be marked as pending update. Feature extraction and optimization calculations are only re-performed on raw material data marked as pending updates.

6. The method according to claim 5, characterized in that, It also includes a feedback optimization mechanism: After generating the final Five Tiger Divine Elixir formula, record the flavor characteristics and health component data of the final Five Tiger Divine Elixir formula. The recorded flavor profile data and health component data are fed back into the machine learning model to optimize the subsequent feature extraction process.

7. The method according to claim 6, characterized in that, It also includes contribution analysis: Establish a recipe generation log to store the input data, intermediate vectors, compensation parameters, and output recipe for each generation of the Five Tiger Divine Elixir recipe; The flavor and health contributions of each ingredient are analyzed based on the recipe generation log. The initial weights of each ingredient in the raw material database are dynamically adjusted based on its flavor and health contributions.

8. The method according to claim 7, characterized in that, It also includes deviation control flow: After the Five Tiger Divine Elixir formula is executed in the formula execution system, sensory evaluation data and component detection data of the actual finished product are collected. The sensory evaluation data and component detection data are compared with theoretical data to calculate the formulation execution deviation; If the deviation from the formula exceeds the allowable range, or if it is caused by raw material fluctuations, the process will switch to the dynamic data update process. If it is caused by equipment errors, the execution system will be calibrated. If it is caused by algorithm defects, the process will switch to the feedback optimization mechanism, triggering the process of regenerating the Five Tiger Divine Elixir formula.

Images