A multi-gustatory interactive perception reverse formula design method and system based on a generative adversarial network
Patent Information
- Application Number
- CN202611093706.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-22
- Publication Date
- 2026-10-27
AI Technical Summary
[0009]本发明的目的在于提供一种基于生成对抗网络的多味觉交互感知逆向配方设计方法及系统，以解决现有技术中存在的无法逆向设计、体系单一、缺乏时序动态特性和个性化能力不足的技术问题
[0035]（1）实现了从“浓度→感知”到“感知→配方”的技术突破，大幅缩短复合风味食品的研发周期，降低研发成本。
Smart Images
Figure CN122991476A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of food flavor science and artificial intelligence, specifically involving a method and system for reverse recipe design based on generative adversarial networks and multi-taste interaction perception, particularly involving taste interaction prediction and reverse recipe generation of systems such as citric acid / capsaicin, acetic acid-piperine, and lactic acid-gingerol. Background Technology
[0002] Food flavor is one of the core driving forces behind consumer food choices. In the development of compound flavored foods (such as hot and sour sauce, pickled chicken feet, and hot and sour soup base), researchers usually need to conduct numerous repeated experiments to determine the proportions of various flavor substances in order to achieve the desired sensory experience. This traditional "trial and error" formula development model is inefficient, costly, and highly dependent on the personal experience of experienced blenders.
[0003] In recent years, with the development of sensory science and computing technology, researchers have begun to attempt to establish mathematical models to describe the relationship between taste perception and stimulus concentration. Based on the Weber-Fechner law, a model of taste interaction perception changes has been established, enabling quantitative prediction of sour-spicy interaction perception. However, this model has the following technical limitations:
[0004] (1) It can only achieve forward prediction (concentration → perception) and cannot reverse design (perception → formulation). In actual product development, R&D personnel usually first determine the target flavor profile and then reverse-engineer the required formulation. Forward prediction models cannot meet this core requirement.
[0005] (2) The model is only applicable to a single system and cannot be extended to other important acid-spicy stimulant combinations, thus limiting its application scope.
[0006] (3) The model is a static model and does not take into account the dynamic changes in taste perception over time (such as onset delay, peak time, duration, etc.), and cannot guide the development of products with specific temporal flavor characteristics.
[0007] (4) The model is based on the average data of the population and does not take into account individual differences, so it cannot achieve personalized formula design.
[0008] In the field of artificial intelligence, Generative Adversarial Networks (GANs) have been successfully applied to image generation, drug molecule design, and other areas, but their application in reverse engineering of food flavor formulations is still lacking. Combining GANs with sensory science models holds promise for overcoming these technical bottlenecks. Summary of the Invention
[0009] The purpose of this invention is to provide a multi-taste interactive perception reverse recipe design method and system based on generative adversarial networks, so as to solve the technical problems of existing technologies such as inability to reverse design, single system, lack of time-series dynamic characteristics and insufficient personalization capabilities.
[0010] To achieve the above objectives, the present invention adopts the following technical solution:
[0011] A multi-taste interactive perception inverse recipe design method based on generative adversarial networks includes the following steps:
[0012] S1: Construct a multi-taste interactive perception discrimination model. This model is based on a modified Weber-Fechner law, introducing a concentration-dependent interaction correction factor α(C). sour C spicy The time-series dynamic decay function β(t) is used to predict the time series of perceived intensity of each taste in a variety of acid-spicy stimuli combination systems.
[0013] The various acid-spicy stimulant combination systems include, but are not limited to: citric acid-capsaicin system, acetic acid-piperine system, lactic acid-gingerol system, malic acid-capsaicin system, and tartaric acid-sanshool system.
[0014] The mathematical expression for the improved Weber-Fechner law is:
[0015] ;
[0016] Where S(C,t) is the perceived intensity at time t, k is the stimulus specificity coefficient, C is the stimulus concentration, and DT is the detection threshold.
[0017] The interaction correction factor α(C) sour C spicy ) is calculated using the following piecewise function: When C sour ≤ C critical When, α = a·ln(C) sour ) + b·ln(C spicy ) + c; when C sour > C critical At that time, α = d·ln(C sour ) + e·ln(C spicy ) + f;
[0018] Where C critical The critical concentration (for the acetic acid-piperine system, C) is the critical concentration. critical ≈1.0 g / L; for the lactic acid-gingerol system, C critical(≈1.5 g / L), a, b, c, d, e, and f are system-specific parameters obtained through sensory experiments.
[0019] The expression for the time-series dynamic decay function β(t) is:
[0020] ;
[0021] Where λ is the attenuation coefficient, γ is the oscillation amplitude coefficient, and ω is the oscillation angular frequency. For the acetic acid-piperine system, λ≈0.015 s⁻¹, γ≈0.08, and ω≈0.05 rad / s; for the lactic acid-gingerol system, λ≈0.012 s⁻¹, γ≈0.06, and ω≈0.04 rad / s.
[0022] S2: The time curve of target taste perception intensity P target (t)=[S sour (t), S spicy [(t)] serves as the input condition vector for the generator network. The target perception contour can be drawn by researchers through a graphical interface or extracted from sensory evaluation data of existing products.
[0023] S3: The generator network is based on a conditional Transformer architecture, including an encoder layer, a cross-attention layer, and a decoder layer. The encoder layer generates the target-aware contour P. target (t) is encoded as a latent space representation; the cross-attention layer fuses prior distribution information of the recipe parameters; the decoder layer outputs a candidate recipe parameter vector R=[C sour C spicy C aux , T process pH].
[0024] The loss function of the generator network is:
[0025] ;
[0026] Where L recon To perceive contour reconstruction loss, L phys L represents the physical constraint loss of the formulation parameters. adv λ1 and λ2 are the discriminator feedback loss and the balance coefficients.
[0027] The physical constraint loss of the formulation parameters L phys Includes: Concentration nonnegativity constraint: L nonneg = Σ[max(0, -C i )]²; Concentration upper limit constraint: L upper = Σ[max(0, C i - C max)]²; Acid-base balance constraint: L pH = |pH - pH eq (C sour C spicy )|;
[0028] S4: Input the candidate formulation parameter R into the discrimination model to obtain the predicted perception intensity curve P. pred (t).
[0029] S5: Calculate the P pred (t) and P target The perceived contour difference D between (t) is: D = w1·MSE(S sour,pred , S sour,target ) + w2·MSE(S spicy,pred , S spicy,target ) + w3·|T max,pred – T max,target | +w4·|Dur pred Dur target |;
[0030] When D < ε (preset threshold, usually ε=5.0), the candidate recipe parameter R is output as the final recipe; otherwise, D is fed back to the generator network for gradient update, and the process returns to step S3 for iterative optimization.
[0031] This invention also provides a multi-taste interactive perception reverse recipe design system based on generative adversarial networks, comprising: a target perception input module, a generator module, a discriminator module, an optimization iteration module, a recipe output module, and an online learning module.
[0032] The system also includes a recipe database and a personalized adaptation module. The personalized adaptation module uses a meta-learning algorithm to quickly adapt to a basic taste interaction model on a small number of individual samples, based on users' historical preference data and genotype information (such as TAS2R38 bitter taste receptor genotype, TRPV1 variant, etc.).
[0033] Beneficial effects
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] (1) A technological breakthrough has been achieved from “concentration → perception” to “perception → formulation”, which has greatly shortened the R&D cycle of compound flavored foods and reduced R&D costs.
[0036] (2) The model is extended to a variety of acid-spicy stimulant combination systems such as acetic acid-piperine and lactic acid-gingerol, which significantly expands the scope of application.
[0037] (3) By introducing a time-series dynamic decay function, the dynamic characteristics of taste perception, such as onset delay, peak time, and duration, can be accurately predicted, guiding the development of products with specific time-series flavor characteristics.
[0038] (4) Through the individualized adaptation module, personalized formulas can be generated based on the individual characteristics of consumers such as genotype, age, and dietary habits, to meet the market demand for precision nutrition and personalized food.
[0039] (5) Through the online learning module, the system can continuously accumulate sensory verification data, realize the self-evolution of the model, and improve the prediction accuracy with the increase of usage. Attached Figure Description
[0040] Figure 1 This is an overall flowchart of the reverse formulation design method described in this invention;
[0041] Figure 2 This is a schematic diagram of the generative adversarial network architecture described in this invention;
[0042] Figure 3 The three-dimensional surface plot shows the interaction correction factor α of the acetic acid-piperine system as a function of concentration.
[0043] Figure 4 The time-series dynamic decay function β(t) curve of the lactic acid-gingerol system is shown.
[0044] Figure 5 This is a comparison diagram of the target perception contour and the predicted perception contour of the generated recipe in Example 1.
[0045] Figure 6 This is a comparison chart of the perception prediction error before and after individualized adaptation in Example 2;
[0046] Figure 7 This is a block diagram of the reverse recipe design system described in this invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.
[0048] Example 1: Reverse design of acetic acid-piperine chutney formulation
[0049] (1) Constructing a discriminant model
[0050] Using acetic acid (food grade, purity ≥99%) and piperine (purity ≥98%, HPLC) as stimulants, the detection threshold DT of acetic acid in pure water was determined by the 3-AFC method. acetic =0.012 g / L, the detection threshold DT of piperine in pure water piperine=0.00008 g / L.
[0051] The justification difference (JND) for different concentration combinations was determined using the 2-AFC method, and the interaction correction factor parameters were obtained by fitting: a=-0.35, b=0.28, c=0.15, d=-0.12, e=0.45, f=0.08, C critical =1.0 g / L;
[0052] The parameters of the time-series dynamic decay function were calibrated using time-intensity (TI) experiments: λ=0.015 s⁻¹, γ=0.08, ω=0.05 rad / s.
[0053] (2) Set the target perception contour
[0054] The target product is a hot and sour sauce with a balanced sour and spicy flavor profile, characterized by a "sour initial taste followed by a spicy finish." The target perception profile is: S sour (t): Rapidly rises to intensity 50 in 0-10s, remains plateaued in 10-30s, and slowly decays to 10 in 30-60s; S spicy (t): 0-15s slowly rises to intensity 20, 15-45s rapidly rises to intensity 60, 45-120s slowly decays.
[0055] (3) Generator network training
[0056] The generator network employs a 6-layer Transformer encoder and a 6-layer Transformer decoder, with a latent space dimension of 512 and 8 attention heads. The discriminator is a pre-trained acetic acid-piperine interaction perception and discrimination model.
[0057] The training data includes: historical formula data: sensory evaluation data of 500 groups of different concentrations of acetic acid-piperine; auxiliary ingredient data: concentration range of seasonings such as salt, sugar, and monosodium glutamate; process parameters: processing temperature 25-85°C, pH 2.5-5.0.
[0058] Training was performed using the Adam optimizer with a learning rate of 1e-4, a batch size of 32, and for 200 epochs.
[0059] (4) Reverse formulation generation and verification
[0060] The target-aware contour is input into the trained generator network, and after 15 iterations of optimization, the candidate recipe is output: C acetic = 1.85 g / L, C piperine = 0.0032 g / L, C NaCl = 15.0 g / L, C sugar = 8.5 g / L, T process=45°C, pH = 3.8.
[0061] Inputting the formula into the discrimination model, the predicted difference between the perceived contour and the target contour is D=3.2 < ε=5.0, which satisfies the convergence condition.
[0062] (5) Sensory verification
[0063] Twenty trained sensory evaluators (10 males and 10 females, aged 22-28) were recruited to perform gLMS scoring and TI curve determination on the hot sauce prepared according to the above formula.
[0064] Verification results: Peak acidity intensity: measured 48.5±3.2 vs predicted 50.0, relative error 3.0%; Peak spiciness intensity: measured 58.3±4.1 vs predicted 60.0, relative error 2.8%; Acidity T max Actual time: 12.5 s vs. predicted time: 10 s, deviation: 2.5 s; Spicy T-shirt max : Actual measurement time 42.0 s vs. predicted time 45 s, deviation 3.0 s; Overall contour similarity (Pearson correlation coefficient): r=0.94.
[0065] The results show that the method of the present invention can accurately reverse engineer a formula that satisfies the target perception contour.
[0066] Example 2: Reverse engineering of a lactic acid-gingerol system for a pickled pepper flavored beverage
[0067] (1) Constructing a discriminant model
[0068] Using lactic acid (food grade, purity ≥85%) and 6-gingerol (purity ≥95%, HPLC) as stimulants, the detection threshold DT of lactic acid in pure water was determined by the 3-AFC method. lactic =0.025 g / L, the detection threshold DT of 6-gingerol in pure water gingerol =0.00015 g / L.
[0069] The fitting yielded the following interaction correction factor parameters: a = -0.42, b = 0.35, c = 0.20, d = -0.18, e = 0.52, f = 0.12, C critical =1.5 g / L;
[0070] The parameters of the time-series dynamic decay function are: λ=0.012 s⁻¹, γ=0.06, ω=0.04 rad / s.
[0071] (2) Set the target perception contour
[0072] The target product is a pickled pepper flavored beverage that is "tangy and refreshing, with a mild and lasting ginger spiciness." The target perception profile is: Ssour (t): Rapidly rises to intensity 40 in 0-5 s, maintains it in 5-20 s, and decays exponentially in 20-90 s; S spicy (t): Slowly rises to intensity 35 from 0-20 s, maintains a plateau from 20-60 s, and slowly decays from 60-180 s.
[0073] (3) Individualized adaptation
[0074] For consumers who are "sensitive to high acidity and sensitive to low spiciness" (screened based on TAS2R38 genotype and historical rating data), the basic model is fine-tuned individually using a meta-learning algorithm.
[0075] Fine-tuning of model parameters: DT lactic,individual =DT lactic × 0.75 (25% increase in acid sensitivity), DT gingerol,individual =DT gingerol × 1.30 (30% reduction in spiciness sensitivity).
[0076] (4) Reverse formulation generation
[0077] After iterative optimization of the generator network, the personalized recipe is output: C lactic = 1.20 g / L (basic model recommends 1.80 g / L, individualized reduction by 33%), C gingerol = 0.0045 g / L (the base model recommends 0.0030 g / L, individualization improves by 50%), C fructose = 25.0 g / L, C citric,acid = 0.5 g / L (auxiliary acidity adjustment), T process = 25°C, pH = 3.5.
[0078] (5) Sensory verification and comparison
[0079] Recruit 15 target consumers to conduct sensory evaluations and compare the perceived differences between the personalized formula and the basic model formula.
[0080] Results: The acidity intensity of the individualized formula was 38.2±2.8 (target 40), with a matching degree of 95.5%; the acidity intensity of the basic model formula was 52.5±3.5 (target 40), with a matching degree of 76.3%; the spiciness intensity of the individualized formula was 33.8±3.1 (target 35), with a matching degree of 96.6%; the spiciness intensity of the basic model formula was 24.2±2.5 (target 35), with a matching degree of 69.1%.
[0081] The individualized formula significantly improved the perceived matching degree of the target consumers, verifying the effectiveness of the personalized design capability of this invention.
[0082] Example 3: Cross-system transfer learning – rapid modeling of the malic acid-capsaicin system
[0083] (1) Transfer learning settings
[0084] Objective: To rapidly establish a taste interaction discrimination model for the malic acid-capsaicin system, thereby reducing the workload of sensory experiments for the new system.
[0085] Source domain model: a pre-trained discriminant model for the citric acid-capsaicin system.
[0086] (2) Fine-tuning with a small number of samples
[0087] Sensory data of the malic acid-capsaicin system at three key concentration combinations (low-low, medium-medium, and high-high) are collected, totaling 60 sets of TI curve data.
[0088] Through transfer learning, the underlying features of the source domain model (Weber-Fechner basic relationships, time-series dynamic patterns) are transferred to the target domain, with only the parameters of the top-level interaction correction factor being fine-tuned.
[0089] (3) Model performance
[0090] The fine-tuned model's prediction performance on the independent test set (5 concentration combinations not used in training): Sourness intensity prediction RMSE: 4.8 (compared to RMSE = 3.5 requiring 500 sets of data for training from scratch); Spicyness intensity prediction RMSE: 5.2; Overall contour similarity: r = 0.91
[0091] With only 12% of the experimental data, the prediction accuracy of the de novo model can be achieved to over 90%, significantly reducing the cost and time of building the new system model.
[0092] (4) Validation of reverse formulation design
[0093] Based on the malic acid-capsaicin system discrimination model established by transfer learning, a pickled pepper flavor seasoning formula that meets the target perception profile was successfully designed in reverse, with a sensory verification relative error of <8%.
Claims
1. A multi-taste interactive perception inverse recipe design method based on generative adversarial networks, characterized in that, Includes the following steps: S1: Construct a multi-taste interactive perception discrimination model, which is based on the improved Weber-Fechner law and introduces a concentration-dependent interaction correction factor α(C). sour C spicy The time-series dynamic decay function β(t) is used to predict the time series of perceived intensity of each taste in a variety of acid-spicy stimuli combination systems. S2: The time curve of target taste perception intensity P target (t)=[S sour (t), S spicy [(t)] serves as the input condition vector for the generator network; S3: The generator network is based on the Transformer architecture and outputs a candidate recipe parameter vector R=[C sour C spicy C aux , T process [, pH], where C sour C represents the concentration of acidic substances. spicy C represents the concentration of pungent substances. aux For auxiliary component concentration, T process The processing temperature is given by [reference], and pH is the acidity or alkalinity of the system. S4: Input the candidate formulation parameter R into the discrimination model to obtain the predicted perception intensity curve P. pred (t); S5: Calculate the P pred (t) and P target The perceived contour difference D between (t) is used to output the candidate formula parameter R as the final formula when D is less than the preset threshold ε. Otherwise, feed D back to the generator network for gradient updates and return to step S3 for iterative optimization.
2. The method according to claim 1, characterized in that, The multiple acid-spicy stimulant combination system mentioned in step S1 includes at least one of the following: citric acid-capsaicin system, acetic acid-piperine system, lactic acid-gingerol system, malic acid-capsaicin system, and tartaric acid-sanshool system.
3. The method according to claim 1, characterized in that, The mathematical expression of the improved Weber-Fechner law described in step S1 is: ； Where S(C,t) is the perceived intensity at time t, k is the stimulus specificity coefficient, C is the stimulus concentration, DT is the detection threshold, and α(C sour C spicy ) is a concentration-dependent interaction correction factor, and β(t) is a time-series dynamic decay function; the interaction correction factor α(C) is a concentration-dependent interaction correction factor. sour C spicy Calculated using the following formula: When C sour ≤ C critical , α = a·ln(C sour ) + b·ln(C spicy ) + c; When C sour > C critical , α = d·ln(C sour ) + e·ln(C spicy ) + f; Where C critical denoted as the critical concentration, and a, b, c, d, e, and f are system-specific parameters obtained through sensory experiments.
4. The method according to claim 1, characterized in that, The expression for the time-series dynamic decay function β(t) mentioned in step S1 is: ； Wherein, λ is the attenuation coefficient, γ is the oscillation amplitude coefficient, and ω is the oscillation angular frequency. The parameters λ, γ, and ω are independently calibrated for different stimulus systems through time-intensity sensory experiments.
5. The method according to claim 1, characterized in that, The generator network described in step S3 adopts a conditional Transformer architecture, including: an encoder layer for processing the target-aware contour P. target (t) is encoded as a latent space representation; a cross-attention layer is used to fuse prior distribution information of the recipe parameters; a decoder layer is used to output the candidate recipe parameter vector R; the loss function of the generator network includes: perceptual contour reconstruction loss L recon Physical constraint loss of formulation parameters L phys And discriminator feedback loss L adv .
6. The method according to claim 5, characterized in that, The physical constraint loss of the formulation parameters L phys include: Concentration nonnegativity constraint: L nonneg = Σ[max(0, -C i )]²; Concentration upper limit constraint: L upper = Σ[max(0, C i - C max )]²; Acid-base balance constraint: L pH = |pH - pH eq (C sour C spicy )|; Where C i For the concentration of each component, C max The maximum allowable concentration of each component, pH eq To balance the pH of the system.
7. The method according to claim 1, characterized in that, The formula for calculating the perceived contour difference degree D in step S5 is as follows: ； Where MSE is the mean squared error, T_max is the time to reach the maximum intensity, Dur is the duration of perception, and w1, w2, w3, and w4 are weighting coefficients.
8. The method according to claim 1, characterized in that, The method further includes: after outputting the final formula parameter R in step S5, conducting a small-batch sensory verification experiment; comparing the actual perceived data obtained from the verification experiment with the predicted data and calculating the relative error; when the relative error exceeds a preset threshold, adding the verification data to the training set and fine-tuning and updating the discrimination model online.
9. A multi-taste interactive perception reverse recipe design system based on generative adversarial networks, characterized in that, include: The target perception input module is used to receive the target taste perception intensity time curve input by the user. The generator module, based on the conditional Transformer architecture, is used to generate candidate recipe parameters; The discriminator module, based on a multi-taste interactive perception discrimination model built on an improved Weber-Fechner law, is used to verify the perceptual prediction of candidate recipes; An optimization iteration module is used to calculate the perceived contour difference and drive the gradient update of the generator network. The recipe output module is used to output the final recipe parameters that satisfy the target perception contour; The online learning module is used to fine-tune and update the discrimination model based on sensory verification experiment data.
10. The system according to claim 9, characterized in that, The system also includes: a recipe database that stores historical recipe parameters and their corresponding sensory verification data; an individualized adaptation module that fine-tunes the output of the generator network based on the user's historical preference data and genotype information; the individualized adaptation module uses a meta-learning algorithm to quickly adapt to the basic taste interaction model on a small number of individual samples.