System and method for prescribing fermentation based on machine learning models

Abstract

The present disclosure provides a method for designing microbial consortia for fermentation to approximate a targeted organoleptic profile of a fermented outcome by training and applying a machine learning model using microbial-substrate interaction data under various fermentation parameters. The model, by taking fermentation parameters predicts metabolites produced by combinations of the candidate microbial strains with candidate substrates under various fermentation parameters and selects at least one microbial strain, substrate, and fermentation parameter for the microbial consortia. The present disclosure also provides a method for predicting an organoleptic profile of a fermented outcome based on the inputs of fermentation parameters, microbial strains, and substrates.

Description

SYSTEM AND METHOD FOR PRESCRIBING FERMENTATION BASED ON MACHINE LEARNING MODELSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Application No. 63 / 700,529, titled SYSTEM AND METHOD FOR OPTIMIZING BEVERAGE FERMENTATION PROCESSES BASED ON ARTIFICIAL INTELLIGENCE MACHINE LEARNING MODELS, filed September 27, 2024, which is hereby incorporated by reference in its entirety.FIELD OF INVENTION

[0002] The present disclosure relates to machine learning systems for prescribing fermentation, and more particularly, to a system and method for designing microbial consortia with taigeted organoleptic profiles.BACKGROUND

[0003] Fermented food and beverages have been produced for thousands of years through traditional methods that rely on naturally occurring microorganisms or established starter cultures to convert sugars and other substrates into alcohol, acids, and various flavor compounds. These can be applied to fermented beverages such as wine, beer, or spirits or to food items such as yogurt, sourdough, and other known fermented comestibles. This process begins with the selection of suitable fermentation agents, such as yeasts, bacteria, and other microorganisms that are responsible for converting sugars into alcohol and other byproducts. The choice of fermentation agent directly influences the flavor, aroma, and overall quality of the final product. The yeasts, bacteria, and other microorganisms carry out complex biochemical reactions that transform the initial ingredients into products with distinct organoleptic properties such as taste, aroma, texture, and appearance.

[0004] Traditional fermentation processes for fermentable comestibles such as production typically involve empirical approaches based on historical knowledge, trial -and-error experimentation, and sensory evaluation by human panels. These conventional methods often require extensive time periods to develop new recipes or improve existing formulations, as producers must conduct numerous fermentation trials to achieve desired flavor profiles and functional characteristics. For example, in winemaking, the selection may include non-Saccharomyces yeasts to achieve unique flavor profiles. The selected fermentation agents are introduced to a substrate, such as freshly pressed grape juice for wine or wort for beer or the like, under controlled conditions to initiate their respective fermentation processes. Parameters such as temperature, pH and nutrient availability are additionally meticulously controlled to ensure consistent and high-quality food or beverage items.

[0005] While experimentation with various fermentation agents and processes can eventually lead to a satisfactory or even excellent food or beverage, selecting appropriate fermentation agents and developing a fermentation process for a desired organoleptic profile can be very time consuming, requiring many trial and error attempts with no guarantee of success. Therefore, it can be challenging to predict the final product characteristics from initial ingredient selections.

[0006] At times, the comestible product may also need to rely on pasteurization, chemical preservatives and synthetic flavour additives in order to shelf preserve the products for consumption and to create an appealing flavour profile. However, for health -conscious consumers, alternatives limiting synthetic additives may be more desirable. Therefore, a need exists that addresses some of these and other disadvantages.SUMMARY

[0007] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0008] According to an aspect of the present disclosure, a method for designing microbial consortia for fermentation is provided. The method involves obtaining data representing microbial - substrate interactions under fermentation parameters and training a machine learning model using the microbial-substrate interaction data to predict metabolites produced by candidate microbial strains with candidate substrates under fermentation parameters. The method further involves receiving a taigeted organoleptic profile for a desired fermented comestible. The method involves applying a trained machine learning model to predict metabolites produced by combinations of the candidate microbial strains and candidate substrates under fermentation parameters. The method additionally involves selecting at least one microbial strain and at least one predicted substrate precursor for achieving the microbial consortia. Here, the at least one microbial strain is selected from the candidate microbial strains predicted to produce the metabolites that contribute to the taigeted organoleptic profile.

[0009] According to another aspect of the present disclosure, a method for designing microbial consortia for fermentation of a substrate is provided. The method includes obtaining latent embeddings of genomic data of candidate microbial strains and obtaining substrate feature embeddings derived from molecular precursors of candidate substrates. Next, the method includes generating pairwise embeddings representing microbial -sub state interactions by combining the latent embeddings with the substrate feature embeddings. It further includes training a supervised machine learning model to predict produced microbial metabolites from the pairwise embeddings. The method separately involves receiving a taigeted oiganoleptic profile for a desired fermented comestible. It furtherinvolves applying said trained supervised machine learning model to generate a predicted organoleptic profile value based on the pairwise embeddings representing microbial -substrate combinations. It subsequently involves applying residual modeling to calculate the difference between the predicted organoleptic profile value and the targeted organoleptic profile value. The method then provides for iteratively evaluating candidate microbial strain and substrate combinations that reduce the difference between the predicted organoleptic profile value and said targeted organoleptic profile value. The method results in selecting at least one microbial strain and at least one substrate for designing the microbial consortia. The at least one microbial strain and at least one microbial substrate are selected from the candidate microbial strain and substrate combinations predicted to produce the microbial metabolites that contribute to said taigeted oiganoleptic profile.

[0010] In another aspect ofthe present disclosure, amethod for generating a recipe fora fermented comestible is provided. The method includes the steps of receiving a target oiganoleptic profile for a fermented comestible and designing microbial consortia and substrate using the steps earlier provided for preparing the fermented comestible. The method further involves providing ratios of each of the microbial strains within the consortia and substrate quantities to generate metabolites contributing to the taiget organoleptic profile for the recipe for the fermented comestible.

[0011] According to another aspect of the present disclosure, a method for dynamic consortia control is provided. The method includes introducing a multi -strain microbial consortium into a fermentation substrate. The method involves then applying fermentation parameters to selectively promote or suppress individual strains. Therefore, the method then involves dynamically steering fermentation outcomes through controlled strain activity.

[0012] According to yet another aspect of the present disclosure, a system for optimizing fermentation processes is provided. The system includes a processor and a memory coupled to the processor. The system includes anon-transitory computer-readable storage medium storing instructions that, when executed by the processor, causes the system to collect fermentation data comprising ingredient compositions, fermentation agent characteristics, and oiganoleptic measurements. The instructions cause the system to generate digital twin representations of pre -fermentation organoleptic profiles and post-fermentation organoleptic profiles. The instructions cause the system to train a machine learning model using the digital twin representations to predict fermentation outcomes. The instructions cause the system to output optimized fermentation parameters for achieving fermented outcome characteristics.

[0013] In yet another aspect of the present disclosure, a computer-implemented method of optimizing a fermentation process for a fermented outcome is provided. The method includes: obtaining a pre-fermentation organoleptic profile of a fermented outcome made from a list of ingredients, and generating a digital twin representation ofthe pre-fermentation organoleptic profile. It further includesobtaining a post-fermentation oiganoleptic profile of the fermented outcome after fermentation utilizing one or more selected fermentation agents and a given fermentation process, and generating a digital twin representation of the post -fermentation oiganoleptic profile. It additionally includes training a machine learning model to analyze changes between the digital twin representation of the pre- fermentation oiganoleptic profile and the digital twin representation of the post-fermentation organoleptic profile. It finally includes developing a machine learning model to predict a postfermentation organoleptic profile given a list of ingredients, one or more selected fermentation agents, and a given fermentation process.

[0014] In another aspect of the present disclosure, a method for predicting an organoleptic profile for a fermentation outcome is provided. The method involves obtaining data representing microbial - substrate interactions of candidate microbial strains and candidate substrate. It further involves training a machine learning model using the microbial -substrate interaction data to predict metabolites produced by the candidate microbial strains with the candidate substrates under a variety of fermentation conditions. It additionally involves providing an input of at least one microbial strain, at least one substrate, and at least one fermentation parameter. It further involves applying the trained machine learning model to predict metabolites produced by combinations of the candidate microbial strains and the candidate substrates under the input fermentation parameters. Finally, it involves generating a predicted oiganoleptic profile based on the predicted metabolites produced.

[0015] The foregoing general description of the illustrative embodiments and the following detailed description thereof are merely exemplary aspects of the teachings of this disclosure and are not restrictive.BRIEF DESCRIPTION OF FIGURES

[0016] Non-limiting and non-exhaustive examples are described with reference to the following figures.

[0017] FIGURE 1 illustrates a flowchart of a method for microbial consortia design.

[0018] FIGURE 2 illustrates a knowledge base structure for data acquisition for the machine learning model, in accordance with an exemplary embodiment;

[0019] FIGURE 3 illustrates an overall architecture of the machine learning system for designing microbial consortia, in accordance with an exemplary embodiment;

[0020] FIGURE 4 is a flowchart diagram illustrating conversion of raw whole genomic sequences into a feature matrix, in accordance with an exemplary embodiment;

[0021] FIGURE 5 shows the process for encoding microbial features into latent representations using variational autoencoders (VAE);

[0022] FIGURE 6 illustrates the process of generating a latent embedding for substrate features suitable for integration with microbial latent vectors in predictive modeling of fermentation outcomes, in accordance with an exemplary embodiment;

[0023] FIGURE 7 illustrates the process of generating pairwise embeddings that integrate microbial and substrate latent representations for predictive modeling of fermentation phenotypes, in accordance with an exemplary embodiment;

[0024] FIGURE 8 shows supervised predicting modeling stage of a multidimensional phenotype, in accordance with an exemplary embodiment;

[0025] FIGURE 9 depicts an inverse design workflow of designing a microbial consortia, in accordance with an exemplary embodiment;

[0026] FIGURE 10 illustrates the cascading residual selection process, in accordance with an exemplary embodiment;

[0027] FIGURE 11 illustrates multi-strain consortia modeling, in accordance with an exemplary embodiment;

[0028] FIGURE 12 illustrates fermentation controls using stressors, in accordance with an exemplary embodiment;

[0029] FIGURE 13 shows a schematic block diagram outlining the components in the feedback loop, in accordance with an embodiment;

[0030] FIGURE 14 illustrates a flowchart outlining the data generation and integration steps for the digital twin, in accordance with an exemplary embodiment; and

[0031] FIGURE 15 shows a schematic block diagram of a system architecture in accordance with an illustrative embodiment;

[0032] FIGURE 16 shows a flowchart outlining a method for generating a recipe for a fermented comestible, in accordance with an illustrative embodiment;

[0033] FIGURE 17 shows a flowchart outlining a method for forward -prediction of an organoleptic profile, in accordance with an illustrative embodiment;

[0034] FIGURE 18 shows a non-alcoholic fermentation pathway; and

[0035] FIGURE 19 shows consortia interaction and cross-feeding.DETAILED DESCRIPTION

[0036] The present disclosure relates to generally rational design and control of microbial consortia and substrate combinations. These can be applicable, for example, in industrial fermentedfood and beverage production also referred to as fermented comestible. The rational design and control is accomplished by machine learning models trained on genomic, metabolic, substrate, and phenotypic data to predict functional contributions of candidate microoiganisms and substrates under various fermentation parameters that in can turn reduce exhaustive empirical testing. Microorganisms, microbes, microbial strains, are all used interchangeably within this disclosure. The substrates identified are capable of biotransformation by the candidate microbial strains under certain fermentation conditions. The system integrates latent feature representation, supervised phenotype prediction, inverse design algorithms, and iterative validation to identify optimal strain-substrate consortia, which will be described in greater detail below.

[0037] Microbes used as precursors for fermentation can encompass yeasts, bacteria and various strains and combinations thereof. The complex interaction between such various microbe strains can result in a fermented comestible such as food and beverage items or ingestible compositions. In the present disclosure, a method for designing the microbial strains or consortia based on machine learning models is provided. The design of microbial strains or consortia can prescribe fermentation to allow for the preparation of a fermented comestible with targeted oiganoleptic profiles.

[0038] Referring now to Figure 1, a method for designing microbial consortia 100 is provided. The microbial consortia design method 100 obtains genomic data 102a of candidate microbial strains from multiple data sources and formats or a knowledge base, as shown for example in FIG. 2. The knowledge base itself provides, for example, genomic sequences, metabolic pathway data, substrate chemical properties, and, where available, phenotypic measurements. Candidate microoiganisms or microbes from which genomic data is extracted may include yeasts and bacteria that safe for ingestion. This can include strains deemed safe by various regulatory bodies such as FDA GRAS strains or EFSA- QPS strains. This can also include other genetically engineered organisms, microorganism approved or subsequently approved as non-pathogenic and suitable for human consumption in food, beverage, nutraceutical, or pharmaceutical applications. For the purposes of this application, any such microbe safe for ingestion, demonstrably or deemed, is conveniently referred to as food-safe microbes and where more appropriate, specified as food-safe bacteria. These could be, for example, Saccharomyces cerevisiae and non-Saccharomyces species, as well as food-safe bacteria including Lactobacillus species (L. brevis, L. delbrueckii, L. kunkeei, L. paracasei, L. plantarum, L. rhamnosus) and other strains as will be known to one skilled in the art. The genomic data 102a may include genomic sequences, nucleotide sequences, gene presence / absence matrices, gene annotations and other genomic information obtained from various sources. For example, these sources could include standard databases, such as SGD (Saccharomyces Genome Database), KEGG, and UniProt, as well as nonpublic and proprietary databases or any combination thereof as will be known to one with skill in the art. Further, the obtained could include gene annotations, metabolic pathways, microbes -genome scale metabolic models. In addition to publicly available databases, non-public databases or subscriptiondatabases or other private databases may also be utilized. This could include experimental genomic sequences, phenotypic measurements including metabolite concentrations, ethanol content, sensory evaluations, and other functional attributes, as well as other machine learning models and / or knowledge graphs among others.

[0039] Referring now to Figure 5, the genomic data 102a may, for example, include whole genome sequences processed through alignment and annotation procedures to generate feature matrices containing gene presence and gene expression data. Raw microbial samples may undergo whole genome sequencing to produce FASTA files containing DNA sequence data, followed by alignment and annotation steps that generate GFF3 / VCF files with gene annotations. This transforms qualitative genomic information into quantitative gene presence and expression matrices suitable for machine learning analysis.

[0040] Additionally, the genomic data 102a may further incorporate metabolic pathway information derived from genome-scale metabolic models such as YeastGEM, MetaCyc, BioCyc, and HMDB. As shown in Figure 6, flux balance analysis (FBA) may be applied to predict metabolic fluxes and pathway activities, providing insights into the metabolic potential of each candidate strain under various substrate conditions and fermentation parameters.

[0041] Referring additionally to Figures 1, 3 and 5, the genomic data 102a may be processed through unsupervised learning algorithms to derive latent embeddings 104a that capture essential microbial characteristics relevant to fermentation performance. These latent embeddings 104a serve as compact, information-rich representations that encode both the intrinsic properties of the microoiganisms and their potential interaction effects relevant to fermentation outcomes. As shown, for example, in Figure 5, variational autoencoders may be utilized to encode genomic features, including gene presence, variants, and pathway-level flux information, into lower-dimensional latent representations.

[0042] The genomic features may first be normalized and standardized to account for differences in scale across features, with standardization including centering the data to zero mean and scaling to unit variance. The variational autoencoder may be trained to minimize a combined loss function comprising reconstmction loss and regularization terms that ensure smooth latent space topology. Alternative unsupervised learning algorithms are also envisioned. These may include sparse PCA, Bayesian PCA, standard autoencoders, and graph autoencoders, each offering different approaches to dimensionality reduction and feature extraction. Still, other unsupervised learning algorithms may be used.

[0043] The latent embedding generation process transforms high -dimensional genomic and metabolic pathway features into a lower-dimensional latent space that captures essential variance relevant to fermentation performance. In the example shown in Figure 5, the variational autoencoder istrained to minimize a combined loss function comprising reconstruction loss, typically measured as mean squared error between the original and reconstructed inputs, and relative entropy that regularizes the latent distribution towards a multivariate Gaussian, thereby ensuring smoothness and disentanglement in the latent space.

[0044] In addition to genomic data 102a of microbial strains, the microbial consortia design method 100 obtains substrate data 102b from various sources as shown in Figure 3, including known public databases such as PubChem and F00DB, as well as proprietary experimental data sources. The substrate data 102b may encompass chemical descriptors, nutritional content, physicochemical properties, and functional additives of candidate substrates. In some aspects, the substrate data 102b may include molecular structures represented by standardized formats such as SMILES or InChi strings, molecular descriptors and fingerprints capturing atom connectivity, bond types, ring structures, and other structural characteristics relevant to microbial metabolism. Other databases and data content may also be included as will be known to the skilled person.

[0045] Accordingly, the substrate data collection process may involve data decomposition to molecular precursors 104b, wherein candidate substrates are broken down into their constituent molecular components that are capable of biotransformation by the candidate microbial strains. This decomposition may identify fermentable molecules and correspond them with taiget metabolites, allowing recognition ofthe same molecular precursors across different substrate sources. In some cases, the method may analyze molecular precursors present in various substrates and other forms of substrates to determine alternative sources of the same molecular components. For example, for the substrate apple juice, the fundamental molecules (molecular precursors) of fructose, malic acid etc are identified. Practically, apple juice is no longer a required substrate as the substrate data allows for the identification of molecular precursors, alternative substrates, and biotransformation. By breaking substrates into molecular precursors, the same precursors across different sources may be recognized. Therefore, by recognizing the molecular precursors in various substrates, alternative substrate sources of the same molecular components can be identified for substitution. This allows for more variability for sourcing, accounting for allergies or intolerance, or simply more options. For the substrates of apple juice, grape juice, malt, or others, the molecular precursors, or the fundamental molecules that microbes can actually biotransform are identified.

[0046] The substrate data 102b may further include predicted or experimentally validated metabolic reactions that the substrate may undergo within target microoiganisms, derived from genome-scale metabolic models such as YeastGEM. This may encode knowledge of substrate utilization pathways, potential metabolite formation, and flux distributions associated with the substrate in the context of specific microorganisms. Properties such as pH, solubility, sugar and amino acid content, viscosity, and other parameters relevant to microbial metabolism and fermentation behaviormay also be incorporated into the substrate feature set to facilitate a more comprehensive analysis of microbial-substrate interactions.

[0047] The genomic and metabolic pathway features may be processed through unsupervised learning algorithms to derive latent embeddings 104a that capture essential microbial characteristics relevant to fermentation performance. These latent embeddings 104a may serve as compact, information-rich representations that encode both the intrinsic properties of the microoiganism and substrate, as well as their potential interaction effects relevant to fermentation outcomes. In some aspects, variational autoencoders may be utilized to encode genomic features, including gene presence, variants, and pathway-level flux information, into lower-dimensional latent representations.

[0048] The genomic features may first be normalized and standardized to account for differences in scale across features, with standardization including centering the data to zero mean and scaling to unit variance. The variational autoencoder may be trained to minimize a combined loss function comprising reconstmction loss and regularization terms that ensure smooth latent space topology. Alternative unsupervised learning algorithms may include sparse PCA, Bayesian PCA, standard autoencoders, and graph autoencoders, each offering different approaches to dimensionality reduction and feature extraction.

[0049] Similarly, substrate features may undergo encoding processes via, for example, feedforward encoder networks to produce latent vectors. The substrate features may comprise multiple complementary components including chemical structure and molecular formula in standardized formats described on or about paragraph

[0040] , These normalized features may be input into a feature encoder to transform the high-dimensional substrate feature space into a latent vector representation that captures the essential variation and functional characteristics of the substrate.

[0050] The substrate data 102b may further include predicted or experimentally validated metabolic reactions that the substrate may undergo within target microoiganisms, derived from genome-scale metabolic models such as YeastGEM. This may encode knowledge of substrate utilization pathways, potential metabolite formation, and flux distributions associated with the substrate in the context of specific microorganisms. Properties such as pH, solubility, sugar and amino acid content, viscosity, and other parameters relevant to microbial metabolism and fermentation behavior may also be incorporated into the substrate feature set to facilitate comprehensive analysis of microbial - substrate interactions.

[0051] Referring now to Figure 1, 3, and additionally 6 and 7, the method 100 may combine the genomic data 102a with the precursor data 102b to generate microbial -substrate interaction data 106. In context, this is shown as genomic latent embeddings 104a with substrate feature embeddings 104b to generate pairwise embeddings 106 representing microbial -substrate interactions. For example, the pairwise embeddings 106 may represent the interaction between a specific Saccharomyces cerevisiaestrain and apple juice substrate, encoding how the yeast's genomic features interact with the substrate's molecular precursors to produce specific metabolites such as ethyl acetate for fruity aromas or glycerol for mouthfeel enhancement. The pairwise embeddings 106 capture potential syneigistic, antagonistic, or additive effects between candidate microorganisms and substrates through concatenation methods, element-wise operations, or learned interaction layers via neural networks. The resulting pairwise embeddings 106 then serve as inputs to train supervised learning models 108 for predicting multidimensional phenotypic outcomes in fermentation processes.

[0052] Alternative approaches for generating pairwise embeddings may include concatenation methods as shown in Figure 8, where microbial latent embeddings Zm and substrate latent embeddings Zs are combined into a single vector, Combined Interaction Embedding Zij = [Zm, Zs]. Element-wise operations may also be employed, such as element -wise multiplication, addition, or other algebraic operations to model interactions explicitly. In some embodiments, learned interaction layers via neural networks may be implemented, including fully connected layers, residual blocks, or attention-based mechanisms trained to maximize predictive accuracy for downstream phenotype modeling. These pairwise embeddings can serve as compact, information -rich representations that encode both the intrinsic properties of the microoiganism and substrate, as well as their potential interaction effects relevant to fermentation outcomes. Although shown as latent embeddings, substrate feature embeddings, and pairwise embeddings, other combinations may also be used as will be known to the skilled person.

[0053] Continuing to refer to Figure 1, the microbial consortia design method 100 then at step 108 trains a machine learning model using the microbial -substrate interaction data to predict metabolites produced by candidate microbial strains with candidate substrates. The machine learning model may employ supervised learning algorithms, including but not limited to gradient-boosted trees such as XGBoost, LightGBM, and CatBoost, Random Forest, Support Vector Machines, neural networks, logistic regression, Ridge regression, LASSO, and Elastic Net methods. In some embodiments, the supervised learning model utilizes pairwise embeddings as input to enable effective learning of complex microbial-substrate interactions. These pairwise embeddings may be obtained through various techniques, including concatenation methods where latent embeddings from genomic data are combined with substrate feature embeddings, element-wise operations such as multiplication or addition, or learned interaction layers implemented via neural networks as additionally described above. Loss functions are selected based on phenotype data types: mean squared error for continuous features, multi-label logistic loss for categorical attributes, or weighted combinations for heterogeneous targets. Model hyperparameters are optimized via k-fold cross-validation, grid search, or Bayesian optimization to improve prediction accuracy while reducing the risk of overfitting. Performance is evaluated using appropriate metrics including coefficient of determination (R2), mean absolute percentage error (MAPE), and correlation coefficients for continuous outputs.

[0054] The pairwise embedding (Figure 7) represents a unified combination where latent values from one side (rich genomic data of microbial strains) are combined with latent values from an other side (substrate data) to create a detailed interaction representation. Specifically, the pairwise embedding of microbe and substrate represents latent features that capture how combinations of genes and metabolic pathways interact under specific contexts. These embeddings encode not only direct metabolic outputs but also the potential for cross-feeding, competition, or pathway complementarity between strains.

[0055] While pairwise embeddings can optimize the method's performance by capturing syneigistic, antagonistic, or additive effects between microoiganisms and substrates, they are not strictly necessary for model training. While the disclosed supervised models are using these latent embeddings to find the relationship with a target, it could be an experimental data or in-silico data like Flux Balance Analysis or even the similarity between strains such as a suggestion of an alternative strain or strain that is competing to the same nutrient. For example, the embedding space can also incorporate similarity between strains, enabling the system to suggest alternative strains with comparable functionality or identify strains likely to compete for the same nutrient resources. By structuring the representation in this way, the system provides insight into additive effects, synergistic interactions, and competitive dynamics within microbial consortia, thereby guiding the design of strain combinations optimized for desired fermentation outcomes. The supervised learning models are trained on experimental data correlating genomic and substrate features with observed metabolite production under various fermentation conditions, enabling accurate prediction of multi-dimensional phenotypic outcomes.

[0056] The microbial consortia design method 100 as shown in Figure 1 also receives a target or taigeted oiganoleptic profile 150 fora desired fermented comestible. This taigeted organoleptic profile 150 may be specified as a multi-dimensional vector encompassing various sensory and functional attributes of the target fermented comestible. The targeted organoleptic profile 150 includes at least one of flavor, aroma, texture, mouthfeel, visual appearance, acidity, bitterness, prebiotic, probiotic, and postbiotic functionalities. These attributes may include flavor characteristics such as sweetness, bitterness, sourness, and umami levels, measured on standardized scales or through quantitative chemical analysis. For example:- Sweetness: Measured in degrees Brix (°Bx) or perceived sweetness level (scale 1 -10)- Bitterness: Quantified using International Bitterness Units (IBU) or perceived bitterness level (scale 1-10)- Sourness: Measured by pH level or perceived sourness (scale 1 -10)- Umami: Quantified by glutamate concentration or perceived umami level (scale 1-10)

[0057] Aroma components may be specified through concentrations of specific volatile organic compounds (VOCs) such as esters (e.g., ethyl acetate), terpenes (e.g., linalool), and phenolic compounds (e.g., eugenol). Aroma compounds could then have specific measurements:- Fruity Aromas: Concentration of esters like ethyl acetate (measured in mg / L)- Floral Notes: Concentration of terpenes such as linalool (measured in mg / L)- Spicy Aromas: Presence of compounds like eugenol (measured in mg / L)

[0058] Other functional attributes are also contemplated. Others will also be known with skill in the art.

[0059] Texture and mouthfeel attributes may include viscosity measurements, carbonation levels expressed in volumes of CO2, and perceived body or creaminess scores. Visual appearance specifications may encompass, for example, color parameters using colorimetric values, and clarity measurements. Functional properties may also be incorporated into the targeted profile, including probiotic capabilities through viable cell counts of beneficial microorganisms, prebiotic effects through the presence of fermentable oligosaccharides, and postbiotic functionalities via bioactive metabolites such as short-chain fatty acids, bioavailable polyphenols, and mood -enhancing compounds like GABA or serotonin precursors.

[0060] The targeted organoleptic profile 150 may also specify functional properties including probiotic capabilities through viable cell counts of beneficial microorganisms, prebiotic effects through the presence of fermentable oligosaccharides, and postbiotic functionalities via bioactive metabolites such as short-chain fatty acids, bioavailable polyphenols, vitamins, and mood -enhancing compounds like GABA, serotonin precursors, and dopamine precursors. Additionally, the profile may include natural preservation functionalities through antimicrobial compounds, natural antioxidants such as polyphenols and flavonoids, and shelf-stability metabolites including lactic acid, acetic acid, and ethanol that naturally preserve the fermented comestible. The taigeted organoleptic profile may also specify production-related parameters such as alcohol content including non-alcoholic specifications below 0.5% ABV, pH levels, acidity measurements, and shelf-stability requirements.

[0061] As shown in Figure 1, the microbial consortia design method 100 at step 108 applies the trained machine learning model to predict metabolites produced by combinations of said candidate microbial strains and said candidate substrates. The application of the trained machine learning model involves generating predictions for various combinations of candidate microbial strains using their genomic data (itself rich with data from DNA, metabolic pathway under various fermentation parameters and others) as input along with molecular precursor data for substrates. Certain known fermentation parameters 102c may also be input.

[0062] In a further refinement, the metabolite identification step may consider the metabolic pathways through which these compounds are produced. This additional refinement can consider whether the selected target metabolites are achievable through microbial fermentation processes. For example, this may involve analyzing genome-scale metabolic models to confirm that candidate strains possess the necessary enzymatic machinery to produce the desired compounds.

[0063] The application of the trained machine learning model 110 involves generating predictions for various combinations of candidate microbial strains using their genomic data as input along with predictions for substrate combinations and fermentation parameters. Finally, the microbial consortia design method 100 selects at least one microbial strain for the microbial consortia 114. This process utilizes the latent embeddings generated during the training phase, where high -dimensional genomic features have been encoded into compact, informative representations that capture essential microbial characteristics relevant to fermentation performance.

[0064] For single-strain predictions generated with the application of the trained machine learning model 110, for example as shown in Figure 8, the system may input the latent embedding of each candidate strain to predict its individual metabolite production profile under specified fermentation conditions. In one embodiment, a single strain may be provided. For example, the latent embedding of a single strain (its genomic / metabolic features compressed into a vector) is processed to predict metabolites produced forthat one strain under specific conditions. In another embodiment, multi-strain consortia may be provided (see Figure 11). In a multi -strain consortia example, the system generates weighted combination embeddings that account for the relative abundance of each strain in the consortium. For example, the latent embeddings of multiple strains and their relative abundances or percentages are weighted to determine the amount / weight of strain. These weighted embeddings are summed to create a mixture-level embedding which represents the combined metabolic potential of all strains together. The mixture phenotype vectors are output that represent the metabolites produced for the consortia as a whole.

[0065] Consequently, the machine learning model can output mixture phenotype vectors representing predicted concentrations for multi -strain consortia, distinct from single-strain phenotype predictions. The latent embeddings of selected microoiganisms are weighted according to their relative abundance and summed to generate a mixture -level embedding that represents the combined metabolic potential of the consortium. The prediction process incorporates substrate data when available as described in paragraph

[0047] to generate pairwise embeddings that represent microbial -substrate interactions. These pairwise embeddings, are constructed using the methods as previously described in paragraph

[0048] , As such, it is possible to determine that a particular strain may behave differently in the presence of a different substrate, for example, when fermenting apple juice versus grape juice, thus providing different metabolites in each scenario.

[0066] This approach captures various potential synergistic, antagonistic, or additive effects between microorganisms and substrates that influence fermentation outcomes. The machine learning model may output multi-dimensional phenotype vectors representing predicted concentrations or production levels of various metabolites, including organic acids such as lactic acid and acetic acid for sourness, esters such as ethyl acetate for fruity notes, alcohols including ethanol and glycerol, phenolic compounds for spicy aromas, and other bioactive molecules. For example, the system may predict the production of sorbitol and mannitol for sweetness enhancement, short -chain fatty acids for prebiotic effects, and neurotransmitter precursors such as GABA for mood-enhancing functionalities. In still other scenarios, this can include amino acids and peptides (e.g., glutamic acid, aspartic acid, y -glutamyl peptides) for umami and savory contributors. In other embodiments, fatty acid derivatives lipids (e.g., short-chain fatty acids like butyric acid, caproic acid; medium-chain fatty acids) for contributing to astringency or mouthfeel.

[0067] These predictions are accompanied by confidence intervals or uncertainty estimates to guide decision-making in strain selection, enabling the rational design of fermentation processes that achieve targeted oiganoleptic profdes within a predetermined threshold, encompassing, for example, flavor, aroma, texture, mouthfeel, visual appearance, acidity, bitterness, prebiotic, probiotic, and / or postb iotic functionalities.

[0068] During strain selection 111, the method 100 predicts microbial strains that produces metabolites that contribute to said targeted organoleptic profile. This is accomplished, by example, cascading residual process 112 as shown more clearly in Figure 10. The strain selection process employs inverse design optimization algorithms to identify microbial strains or combinations thereof that are predicted to produce the specified metabolites at desired levels. Inverse design optimization works backward from target phenotypic outcomes to determine optimal microbial-substrate combinations. As shown in Figure 10, this selection utilizes greedy cascading residual optimization methods, where candidate strains are iteratively selected based on their marginal contribution to reducing the residual between predicted and target metabolite profiles. The evaluation at each iteration occurs for all candidate microbes and substrates not yet selected. In one embodiment, the combination of microbes and substrates that most reduces the residual error relative to the target phenotypic vector (target organoleptic profile) is used to update the residual for the next iteration. In other embodiments, this could be according to a predefined tolerance or other condition. Therefore, strains are chosen iteratively based on their ability to reduce the residual between the taiget profile and the predicted outcomes.

[0069] As shown in Figure 10, the optimization process begins by initializing a residual vector 112a representing the difference between the target metabolite profile and the predicted output of an empty combination. At each iteration, the system evaluates all candidate microorganisms not yet selected, determines the strain that most reduces the residual magnitude, and updates the residual for the next iteration 112b. Put another way, candidate strains are iteratively determined based on theirmarginal contribution to reducing the residual between predicted and target metabolite profiles. The evaluation at each iteration occurs for all candidate microbes and substrates not yet selected. The combination of microbes and substrates that most reduces residual error relative to the target phenotypic vector (target organoleptic profile 150) is used to update the residual for the next iteration. Therefore, strains are chosen iteratively based on their ability to reduce the residual between the target profile and the predicted outcome. This residual modeling 112 continues until a certain condition is met. For example, the process continues until residual falls below a predetermined tolerance, a maximum number of strains is reached, or negligible marginal improvement from selected strains. Other predetermined conditions could also be imposed, for example, a set number of iterations, a defined number of strains identified, or any other suitable condition as will be known to one with skill in the art

[0070] Alternative optimization approaches may include gradient -based searches in latent space or Bayesian optimization methods to refine candidate solutions. The selection process also incorporates constraints such as strain compatibility, fermentation conditions, and practical considerations for industrial implementation. The final selection may result in either single-strain solutions where one microoiganism is predicted to produce the desired metabolites, or multi -strain consortia where multiple microoiganisms work synergistically to achieve the target outcomes.

[0071] The microbial consortia design method 100 further evaluates and ranks potential strain combinations according to their predicted performance metrics (see, for example, Figure 9). The selected strains are ranked according to their predicted performance, allowing for backup options and iterative refinement based on experimental validation results. The selection process may also consider the functional contributions of each strain beyond metabolite production, including their roles in pH regulation, oxygen consumption, cross-feeding relationships, and overall fermentation stability. This holistic approach ensures that the selected consortia not only produce the desired metabolites but also maintain desirable fermentation performance under, for example, industrial conditions. The candidate microbial strains may include those earlier described, for example, in paragraph

[0034] ,

[0072] According to another aspect of the present disclosure, a method for dynamic consortia control is provided. The method includes introducing a multi -strain microbial consortium into a fermentation substrate. The multi-strain microbial consortium may comprise the candidate microoiganisms described in paragraph

[0034] and combinations thereof suitable for fermentation applications. The method involves then applying parameters to selectively promote or suppress individual strains. These parameters can be, for example, environmental stressors or process parameters to selectively promote or suppress individual strains.

[0073] For example, the environmental stressors may include pH adjustment, temperature modulation, nutrient addition, and oxygen control or a combination thereof. In some embodiments, pH control may involve modulating pH to suppress sensitive strains or promote acid-tolerant ones.Temperature modulation may involve activating specific species at cooler or warmer stages of fermentation. Oxygen control may comprise transitioning between aerobic and anaerobic conditions to modulate aroma compound production, thereby steering metabolism toward oxidative or fermentative pathways and altering aroma and metabolite output.

[0074] The process controls may include metabolite accumulation and substrate manipulation. Metabolite accumulation may comprise using at least one of lactic acid, ethanol, and acetaldehyde as control signals to regulate strain activity. In some embodiments, metabolite -mediated regulation may involve using lactic acid, ethanol, or acetaldehyde as control signals to stimulate or inhibit supporting strains during fermentation. This can also be interspecies metabolic exchange among constituent strains. In this scenario, metabolites secreted by one member can act as signaling molecules or inhibitors that modulate the growth, function, or metabolic output of others, as illustrated, for example in Figure 13. The interspecies metabolic exchange may involve cross-feeding relationships between consortium members resulting in naturally occurring metabolite production and consumption. This results from biological interaction within the system rather than external control. The metabolites secreted by one strain may promote the growth of another strain while simultaneously inhibiting a third strain, creating complex regulatory networks within the consortium.

[0075] In one example, selective strain promotion or suppression can be accomplished by carbon sources such as glucose, maltose, starch, lactose, complex plant polysaccharides. The concentration can also bias the consortium composition. Further, nitrogen sources such as Free amino nitrogen (FAN), peptides, ammonium salts may be used. Additionally, yeast and bacteria may compete or complement. In other embodiments, this can be micronutrients & cofactors such as vitamins (e.g., thiamine, riboflavin and others), minerals (Mg2+, Zn2+, Mn2+) that can all regulate specific strain performance. For example, the carbon to nitrogen ratio can skew growth toward fast-growing bacteria rather than slower yeasts. Nonetheless whether internal process control of environmental stressor, these parameters result in deliberate manipulation to control strain activity. These and other various parameters are envisioned by the present disclosure. Next, the instructions specify when and how to apply these stressors to dynamically steer fermentation outcomes through controlled strain activity, enabling fine-tuning of the fermentation process to achieve results without requiring genetic modification of the microorganisms. These results can be within a predetermined threshold.

[0076] Therefore, the method then involves dynamically steering fermentation outcomes through controlled strain activity. The dynamic steering may enable fine-tuning of the fermentation process to achieve results within a predetermined threshold without requiring genetic modification of the microoiganisms. These control mechanisms enable dynamic steering of consortia without genetic modification, complementing the predictive machine learning design layer. The dynamically steered fermentation outcomes may include producing target compounds selected from organic acids, esters, alcohols, and phenolic compounds or a combination thereof.

[0077] In another aspect of the present disclosure, a method 200 for generating a recipe for a fermented comestible is provided. Referring now to Figure 16, the method leverages the machine learning-based microbial consortia design system described herein. This method integrates the predictive capabilities of the trained models with practical recipe formulation to produce fermented comestibles with targeted organoleptic and functional characteristics.

[0078] The recipe generation method 200 begins by receiving a target organoleptic profile 150 for a fermented comestible, which serves as the foundation for all subsequent design decisions. This target organoleptic profile encompasses the detailed range of sensory and functional attributes as described in or about paragraph

[0051] ,

[0079] Other aspects of an organoleptic profile may capture texture and mouthfeel specifications within the target profile that include viscosity measurements, carbonation levels expressed in volumes of CO2, and perceived body or creaminess. The organoleptic profile may also further capture visual appearance which could encompass color specifications using colorimetric values such as CIE Lab coordinates, and clarity measurements through turbidity or visual clarity scales. The taiget organoleptic profile may additionally incorporate functional properties including probiotic capabilities through viable cell counts of beneficial microoiganisms, prebiotic effects through the presence of fermentable oligosaccharides, and postbiotic functionalities via bioactive metabolites such as short -chain fatty acids, bioavailable polyphenols, vitamins, and mood -enhancing compounds like GABA, serotonin precursors, and dopamine precursors.

[0080] The recipe generation method 200 then proceeds by designing microbial consortia 204a using method 100 defined above with the machine learning approaches described in the preceding claims, specifically employing the genomic data analysis, latent embedding generation, and predictive modeling techniques. As explained above, this design process utilizes the trained machine learning models to identify optimal combinations of candidate microbial strains that collectively produce the metabolites necessary to approximate the target oiganoleptic profile. The microbial consortia design can incorporate both single-strain solutions where one microorganism produces the desired metabolite profile, and multi-strain consortia where multiple microoiganisms work syneigistically to achieve the target outcomes, as described above.

[0081] Next, the recipe generation method 200 determines substrates 204b for the recipe. These substrates are selected based on predicted microbial -substrate interactions derived from the pairwise embedding analysis. The substrate determination process involves detailed analysis of molecular precursors present in candidate substrates, ensuring that these precursors are capable of biotransformation by the selected microbial strains. This analysis includes decomposition of candidate substrates into their constituent molecular components, identification of fermentable molecules, and correlation with taiget metabolites as applicable and defined above. The system recognizes molecularprecursors across different substrate sources, enabling identification of alternative sources of the same molecular components to provide flexibility in sourcing, accommodate intolerances, sensitivities or preferences, optimize cost considerations, and other benefits. From the precursor data, potential substrates or ingredients can be determined or sourced.

[0082] The substrate data, as further described in paragraph

[0040] , may be obtained from databases such as PubChem, F00DB, or other public, proprietary or nonpublic sources or a combination thereof. The molecular precursors of the substrate can be mapped onto a specified suggested substrate as an ingredient.

[0083] The recipe generation method 200 additionally provides precise ratios 206 of each microbial strain within the consortia and substrate quantities, calculated to generate the specific metabolites that contribute to the taiget organoleptic profile. The recipe generation method 200 calculates specific ratios (relative abundance) for each strain in the consortium based on their individual contributions to the desired metabolic outcomes. These ratios may be determined through the optimization algorithms, including, for example, greedy cascading residual optimization methods, where candidate strains are iteratively selected based on their marginal contribution to reducing the residual between predicted and target metabolite profiles. The optimization process accounts for strainstrain interactions and their combined metabolic potential when designing multi -strain consortia, enabling the creation of complex mixtures with predictable combined effects.

[0084] The recipe generation method 200 may also include quality control measures 208 and protocols to more have the fermented comestible more closely approximate the target organoleptic profile specifications. These may include sensory evaluation procedures, chemical analysis requirements, and microbiological testing protocols. In some embodiments, the quality control measures may encompass comprehensive monitoring of the fermentation process to ensure relatively consistent achievement of the taiget organoleptic profile . The sensory evaluation procedures may assess multiple dimensions of the oiganoleptic profile, including flavour characteristics such as sweetness, bitterness, sourness, and umami levels, measured on standardized scales or through quantitative chemical analysis. Aroma components may be evaluated through concentrations of specific volatile organic compounds such as esters, terpenes, and phenolic compounds. Texture and mouthfeel attributes may include viscosity measurements, carbonation levels expressed in volumes of CO2, and perceived body or creaminess scores.

[0085] Referring to Figure 15, these sensory evaluation procedures may be carried out, for example, by electronic sensors such as electronic nose sensors and electronic tongue sensors. The electronic nose (e-nose) systems may consist of sensor arrays designed to detect and differentiate volatile compounds, providing digital fingerprints of aromas that can be processed and analyzed using pattern recognition algorithms and machine learning techniques to classify and quantify scent profiles.Similarly, the electronic tongue (e-tongue) systems may employ sophisticated sensor arrays that respond to various taste stimuli, including potentiometric sensors for measuring ion concentrations associated with taste, amperometric sensors for detecting compounds like sugars and amino acids, conductometric sensors for measuring electrical conductivity changes, capacitive sensors for detecting taste molecule adsorption, optical sensors using fluorescence and absorbance techniques, and piezoelectric sensors for measuring mass or viscosity changes.

[0086] The signals generated by these sensor arrays are processed using advanced data analysis techniques, including pattern recognition algorithms, principal component analysis (PCA), and machine learning models to generate comprehensive taste and aroma profiles that differentiate between complex mixtures and provide valuable insights into the sensory properties of fermented comestibles.

[0087] Capturing and recording scent or aroma data involves both sensory evaluation and advanced analytical technologies. Similar to traditional sensory taste evaluation relying on human taste panels, sensory evaluation of scent or aroma data typically uses trained panelists who assess aroma based on a structured sensory lexicon, recording their perceptions through detailed descriptors and intensity ratings on standardized scales. This qualitative data, often recorded using structured questionnaries and rating scales helps inform an understanding human sensory responses.

[0088] In another embodiment, gas chromatography -mass spectrometry (GC-MS) and high- performance liquid chromatography (HPLC) may be employed to analyze and quantify the chemical compounds contributing to flavor. These analytical techniques provide detailed profiles of volatile and non-volatile compounds. Furthermore, machine learning algorithms can process and correlate sensory data with chemical analysis results to develop predictive models of taste. Consumer feedback collected through digital platforms, such as mobile apps, also offers valuable data, capturing real -time preferences and experiences. By integrating these diverse methods, a comprehensive and objective record of taste data can be established, facilitating more precise and consistent flavor optimization in fermented beverages.

[0089] Similar to the above for flavor, gas chromatography -mass spectrometry (GC-MS) may be used to analyze and identify volatile organic compounds responsible for aroma. GC-MS separates complex mixtures and provides detailed profiles of individual aroma compounds, offering precise quantitative data. So lid -phase microextraction (SPME) is often combined with GC-MS to efficiently capture volatile compounds from samples without the need for solvents.

[0090] In another embodiment, headspace analysis may be used for capturing and recording scent or aroma data. Headspace analysis involves a technique that measures volatile compounds present in the space above a sample is also employed to capture aroma data. This method, combined with gas chromatography, helps in understanding the composition of complex aroma profiles.

[0091] Integrating these methods allows for a comprehensive capture and recording of aroma data, facilitating detailed analysis and consistent reproduction of desired scent profiles in various applications, including fermented beverages. As will be appreciated, combining taste and smell data can enhance the analysis and understanding of overall oiganoleptic profiles, as both senses contribute significantly to the perception of flavor. Some methods for integrating taste and smell data may include, as follows:• Data Fusion Techniques: o Early Fusion: Combine raw data from taste sensors (e-tongues) and smell sensors (enoses) into a single dataset before feature extraction. This approach allows for the simultaneous processing of both types of data, capturing their interactions from the beginning. o Intermediate Fusion: Extract features separately from taste and smell data and then combine these features into a unified feature set. This method allows for the independent optimization of feature extraction techniques for each type of data. o Late Fusion: Process taste and smell data independently through separate models and then combine their outputs. This approach can involve averaging predictions, voting systems, or using another model to integrate the results.• Feature Extraction and Selection: o Extract relevant features from both taste and smell data using techniques such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), or autoencoders. Feature selection methods can then identify the most informative features from both datasets, reducing dimensionality and improving model performance.• Multimodal Deep Learning: o Use deep learning models designed to handle multiple types of input data. For example, Convolutional Neural Networks (CNNs) can process smell data, while Recurrent Neural Networks (RNNs) or fully connected layers handle taste data. These models can learn complex interactions between taste and smell features, improving the accuracy of flavor profile predictions.• Machine Learning Ensemble Methods: o Apply ensemble methods that combine the outputs of multiple machine learning models trained on taste and smell data. Techniques such as stacking, boosting, and bagging can leverage the strengths of individual models, providing a more robust and accurate final prediction.• Integrated Sensor Systems:o Develop integrated sensor systems that simultaneously capture taste and smell data. These systems can provide synchronized data streams, facilitating more coherent and meaningful analysis. Synchronization ensures that the data from both senses correspond to the same sample conditions and time points.• Correlation Analysis: o Perform correlation analysis to understand the relationships between taste and smell features. This analysis can identify patterns and interactions that contribute to the overall flavor perception, guiding the development of combined models.• Multivariate Statistical Methods: o Utilize multivariate statistical methods such as Canonical CorrelationAnalysis (CCA) or Partial Least Squares (PLS) regression to model the relationships between taste and smell data. These methods can uncover latent variables that influence both types of sensory data.• By integrating taste and smell data through these techniques, it is possible to aim for a more detailed and accurate understanding of flavor profiles, which can be valuable for applications in the food and beverage industry, quality control, and product development.

[0092] The chemical analysis requirements may involve monitoring production -related parameters such as alcohol content including non-alcoholic specifications below 0.5% ABV, pH levels, acidity measurements, and shelf-stability requirements. The analysis may also encompass functional properties including probiotic capabilities through viable cell counts of beneficial microoiganisms, prebiotic effects through the presence of fermentable oligosaccharides, and postbiotic functionalities via bioactive metabolites such as short -chain fatty acids, bioavailable polyphenols, vitamins, and moodenhancing compounds like GABA, serotonin precursors, and dopamine precursors.

[0093] The microbiological testing protocols may include procedures for monitoring the activity and relative abundance of selected microbial strains within the consortia throughout the fermentation process. These protocols may assess the maintenance of desired strain ratios and the production oftaiget metabolites including organic acids such as lactic acid and acetic acid, esters such as ethyl acetate, alcohols including ethanol and glycerol, and phenolic compounds.

[0094] The instructions may incorporate fermentation process sequencing data comprising timing of ingredient introduction and post -fermentation additions. This may include pre -fermentation and postfermentation additions, identification of fermentation inhibitors in ingredients, and optimal introduction sequences to enhance or suppress individual strain activity. The process sequencing may encompassenvironmental condition specifications such as temperature, pH, and nutrient availability that must be controlled to ensure consistent fermentation outcomes.

[0095] The quality control framework may support scalable industrial production by providing detailed process flow that can be utilized to scale the beverage production at commercial facilities, to improve consistency and efficiency across production batches while maintaining the target organoleptic profile specifications.

[0096] For multi-strain consortia, the latent embeddings of selected microoiganisms are weighted according to their relative abundance and summed to generate a mixture -level embedding that represents the combined metabolic potential of the consortium (see, for example, Figure 12). This approach captures various potential synergistic, antagonistic, or additive effects between microoiganisms and substrates that influence fermentation outcomes. The system predicts the production of specific metabolites including organic acids such as lactic acid and acetic acid for sourness, esters such as ethyl acetate for fruity notes, alcohols including ethanol and glycerol, phenolic compounds for spicy aromas, and other bioactive molecules such as sorbitol and mannitol for sweetness enhancement, short-chain fatty acids for prebiotic effects, and neurotransmitter precursors such as GABA for mood-enhancing functionalities.

[0097] Further, texture and mouthfeel attributes may be identified as specific metabolites. By way of example, these can include glycerol for smoothness and body, polysaccharides for viscosity enhancement, and CO2 for carbonation effects. Functional properties may be linked to specific bioactive compounds, including short-chain fatty acids for prebiotic effects, viable probiotic cells for gut health benefits, and neurotransmitter precursors such as GABA for mood -enhancing functionalities as mentioned above.

[0098] These can further include, for example, higher alcohols (fusel alcohols) (e.g., isoamyl alcohol with solvent / fusel aroma, 2 -phenylethanol for rose-like notes); carbonyl compounds such as aldehydes and ketones (e.g., acetaldehyde with green apple aroma, diacetyl with buttery aroma, 2,3- pentanedione with buttery / creamy notes); sulfur-containing volatiles (e.g., dimethyl sulfide with cooked com aroma, methanethiol with onion / garlic notes, thiols such as 3 -mercaptohexanol for passionfruit / tropical aroma); lactones (e.g., y-nonalactone and 5-decalactone for coconut, peach, and creamy aromas); pyrazines (e.g., 2 -isobutyl -3 -methoxypyrazine for green / bell pepper aroma, other alkylpyrazines for roasted / nutty notes) and others that will be known to the skilled person.

[0099] The final optional component of the recipe generation method involves providing comprehensive step-by-step 210 instructions for preparing the fermented comestible. These instructions encompass detailed fermentation process parameters and sequencing, including pre-fermentation and post-fermentation additions, timing of ingredient introduction, and identification of fermentation inhibitors in ingredients. The instructions specify environmental conditions such as temperature, pH,and nutrient availability that must be meticulously controlled to ensure consistent and high-quality fermented products. The process parameters include fermentation duration, agitation requirements, oxygen management, and monitoring protocols to track fermentation progress.[OlOO] The step-by-step instmctions incorporate dynamic consortia control methods, including the application of various parameters which can include environmental stressors to selectively promote or suppress individual strains during fermentation, as described in more detail above.

[0101] In yet another embodiment of the present disclosure, a system for optimizing fermentation processes is also provided. The system includes a processor and a memory coupled to the processor. The processor is configured to execute machine learning algorithms and the memory stores both training data and model parameters. A non-transitory computer-readable storage medium storing instructions is provided that, when executed by the processor, causes the system to undertake steps. In one step, the system collects fermentation data comprising ingredient compositions, fermentation agent characteristics, and organoleptic measurements, as described in paragraphs

[0032] -

[0046] , The system also generates digital twin representations of pre -fermentation oiganoleptic profiles and postfermentation organoleptic profiles, wherein the digital twin representations are comprehensive digital representations that simulate sensory attributes, organoleptic attributes, chemical attributes, functional properties, and production data, wherein the latent embedding generation process and variational autoencoder training are as described in paragraphs

[0035] -

[0044] . Next, the system trains a machine learning model using the digital twin representations to predict fermentation outcomes, wherein the supervised learning algorithms and model training are as described in paragraphs

[0049] -

[0050] , Finally, the system outputs optimized fermentation parameters for achieving target beverage characteristics, wherein the optimization process utilizing greedy cascading residual optimization methods, pairwise embeddings generation, and multi -strain consortia modeling are as described in paragraphs

[0057] -

[0061] ,

[0102] In the aforementioned optimization system, the digital twin representations comprise comprehensive digital representations that simulate sensory attributes, organoleptic attributes, chemical attributes, functional properties, and production data, wherein the sensory attributes include flavor characteristics such as sweetness, bitterness, sourness, and umami levels as described in paragraphs [005 l]-

[0053] , aroma components including concentrations of specific volatile organic compounds such as esters, terpenes, and phenolic compounds, texture and mouthfeel attributes including viscosity measurements and carbonation levels, and visual appearance parameters encompassing color specifications and clarity measurements. The oiganoleptic attributes encompass the multi-dimensional organoleptic profile described in paragraphs

[0051] -

[0053] including flavor, aroma, texture, mouthfeel, color, acidity, bitterness, prebiotic, probiotic, and postbiotic functionalities. The chemical attributes include predicted metabolites, esters, volatile organic compounds, and volatile acidity compounds generated through metabolic pathways of fermentation agents applied to input ingredient precursors asdescribed in paragraphs

[0032] -

[0046] . The functional properties include probiotic capabilities through viable cell counts of beneficial microoiganisms, prebiotic effects through fermentable oligosaccharides, postbiotic functionalities via bioactive metabolites, natural preservation functionalities through antimicrobial compounds, and shelf-stability metabolites as described in paragraphs

[0052] -

[0053] , The production data includes fermentation process sequencing comprising timing of ingredient introduction and post-fermentation additions, cost of goods sold data dynamically updated based on real-time integration of supplier data and manufacturing processes, and carbon footprint data comprising scope 1, scope 2, and scope 3 emissions data for the fermented outcome. In some embodiments, the fermented outcome is a beverage, in other it is food..

[0103] In the optimization system, the organoleptic measurements comprise data collected from electronic nose sensors and electronic tongue sensors as shown, for example, in Figure 15. The electronic nose (e-nose) systems consist of sensor arrays designed to detect and differentiate volatile compounds, providing digital fingerprints of aromas that can be processed and analyzed using pattern recognition algorithms and machine learning techniques to classify and quantify scent profiles. As described in paragraphs

[0082] -

[0089] above, electronic tongue (e-tongue) systems employ sophisticated sensor arrays that respond to various taste stimuli, including potentiometric sensors for measuring ion concentrations associated with taste, amperometric sensors for detecting compounds like sugars and amino acids, conductometric sensors for measuring electrical conductivity changes, capacitive sensors for detecting taste molecule adsorption, optical sensors using fluorescence and absorbance techniques, and piezoelectric sensors for measuring mass or viscosity changes. The signals generated by these sensor arrays are processed using advanced data analysis techniques, including pattern recognition algorithms, principal component analysis (PCA), and machine learning models to generate comprehensive taste and aroma profiles that differentiate between complex mixtures and provide valuable insights into the sensory properties of fermented comestibles.

[0104] In the optimization system, the digital twin representation simulates sensory attributes including flavor, aroma, texture, and appearance, wherein the sensory attributes are derived from the multi-dimensional phenotype vectors described in paragraphs

[0054] -

[0056] that encompass various organoleptic characteristics. The flavor simulation incorporates predicted concentrations of organic acids such as lactic acid and acetic acid for sourness profiles, esters such as ethyl acetate for fruity characteristics, and phenolic compounds for spicy aromatics as detailed in the predictive modeling framework. The aroma simulation utilizes volatile organic compound predictions generated through the supervised learning models trained on pairwise embeddings, capturing the complex interactions between microbial metabolic pathways and substrate molecular precursors. Texture and appearance attributes are simulated through predictions of glycerol production for mouthfeel enhancement, carbonation levels, viscosity measurements, and color parameters using the comprehensive digital twin methodology described in paragraphs

[0061] -

[0066] ,

[0105] In the optimization system, the instmctions further cause the system to predict resulting metabolites, esters, volatile organic compounds, and volatile acidity compounds using metabolic pathways of fermentation agents applied to input ingredient precursors, wherein the metabolic pathway analysis incorporates genome-scale metabolic models such as YeastGEM, MetaCyc, BioCyc, and HMDB as described in paragraphs

[0034] and

[0041] , The prediction process utilizes flux balance analysis (FBA) to determine metabolic fluxes and pathway activities under various substrate conditions, enabling accurate forecasting of metabolite production profiles. The system applies the trained machine learning models described in paragraphs

[0049] -

[0050] to predict specific metabolites. These specific metabolites and featured characteristics can include, for example, sorbitol and manitol for sweetness enhancement, short-chain fatty acids for prebiotic effects, and neurotransmitter precursors such as GABA for mood-enhancing functionalities. In still other examples, volatile organic compound predictions can encompass, for example, esters for fruity notes, terpenes for floral characteristics, and phenolic compounds for spicy aromas, all generated through the pairwise embedding methodology that captures microbial-substrate interactions detailed in paragraphs

[0047] -

[0048] , These are examples only, and others combinations may also be known or elicited by a skilled person.

[0106] In the optimization system, the digital twin representation includes fermentation process sequencing data comprising timing of ingredient introduction and post -fermentation additions, wherein the sequencing data incorporates the comprehensive step-by-step instructions for fermentation process parameters described in the recipe generation methodology. The timing specifications include pre- fermentation and post-fermentation additions, identification of fermentation inhibitors in ingredients, and optimal introduction sequences to enhance or suppress individual strain activity as detailed in the dynamic consortia control methods. The process sequencing encompasses environmental condition specifications such as temperature, pH, and nutrient availability that must be controlled to ensure consistent fermentation outcomes, incorporating the environmental stressor applications described for selective promotion or suppression of individual strains or the parameter application. The fermentation duration, agitation requirements, oxygen management, and monitoring protocols are integrated into the digital twin to provide complete production guidance, enabling the manufacturing process flow that can be utilized for commercial-scale production as referenced in the comprehensive fermentation parameter optimization framework.

[0107] Also included in the optimization system digital twin representation includes cost ofgoods sold data that is dynamically updated based on real-time integration of supplier data, manufacturing processes, and market conditions, wherein the cost analysis utilizes a Building Information Modeling (BIM)-like technology platform for sophisticated integration of data from various sources as described in the digital twin comprehensive representation. The dynamic cost management system provides realtime integration of data from suppliers, manufacturing processes, and market conditions to deliver up- to-date cost estimates, incorporating detailed breakdown of material costs, labor costs, and other directcosts associated with the production process. The system employs advanced simulation tools that model the cost impact of different production scenarios, allowing for optimization of cost efficiency through continuous updating of cost estimates based on changes in raw material prices, production efficiencies, and other variables. This ensures accurate financial planning and budgeting capabilities that support the commercial scalability of the fermentation recipe development process, complementing the comprehensive digital twin functionality that encompasses sensory attributes, chemical attributes, functional properties, and production data as detailed in the system architecture.

[0108] Additionally, the optimization system as described has a digital twin representation that includes carbon footprint data comprising scope 1, scope 2, and scope 3 emissions data for the fermented beverage. Finally, the system as further described, includes a feedback loop for iterative model refinement based on experimental validation results. Predicted consortia may be validated through controlled fermentation experiments and in silico validation using computational models as shown in, for example, Figure 13.

[0109] According to another aspect of the present disclosure, a computer-implemented method of optimizing a fermentation process for a fermented outcome is provided. The method involves obtaining a pre-fermentation organoleptic profile of a fermented outcome made from a list of ingredients, and generating a digital twin representation of the pre-fermentation organoleptic profile. The digital twin representation includes a detailed digital representation that simulates the sensory attributes of the beverage, including flavor, aroma, texture, and appearance as described above. ML platform achieves this by predicting the resulting metabolites, esters, volatile organic compounds (VOCs), and volatile acidity compounds (VACs) using the metabolic pathways of the biotech or fermentation agents applied to the input ingredient precursors data. The combination of these two elements — biotech and ingredient precursors — ensures that using the same biotech on different ingredient inputs results in distinct digital twin profiles.[OHO] The method further comprises obtaining a post-fermentation organoleptic profile of the fermented outcome after fermentation utilizing one or more selected fermentation agents and a given fermentation process, and generating a digital twin representation of the post -fermentation oiganoleptic profile. The post-fermentation digital twin captures the transformed sensory characteristics resulting from the fermentation process, including the production of metabolites, esters, VOCs, and VACs based on the metabolic pathways of the selected live culture strains when applied to the specific input ingredient precursors.

[0111] The method includes training a machine learning model to analyze changes between the digital twin representation of the pre-fermentation organoleptic profile and the digital twin representation of the post-fermentation organoleptic profile. Upon creating a sufficiently large data set of oiganoleptic profiles before and after fermentation, the data set is used to train the ML model so thatthe model leams which ingredients and processing steps result in which desirable organoleptic properties. The model is able to predict how certain fermentation agents alter the organoleptic profiles through their metabolic pathways and biochemical transformations.

[0112] The method further comprises applying the trained machine learning model to predict a post-fermentation oiganoleptic profile given a list of ingredients, one or more selected fermentation agents, and a given fermentation process. The trained ML model is used to generate a list of ingredients and instructions for a recipe to create a post-fermentation beverage with the desired organoleptic profiles. The model predicts the production of metabolites, esters, VOCs, and VACs based on the metabolic pathways of the selected live culture strains when applied to the specific input ingredient precursors, ensuring accurate simulation of the beverage's sensory profile for various ingredient combinations.

[0113] In some embodiments, all data generated from users' fermentation iterations will be collected and analyzed to continuously train and improve the ML model. Each iteration, whether a success or failure, provides valuable insights into the fermentation process, ingredient interactions, and resulting oiganoleptic profiles. By aggregating and anonymizing this data, the platform leverages a vast repository of real-world fermentation scenarios to refine its predictive accuracy and expand its knowledge base.

[0114] The method may optionally include providing an intuitive User Interface (UI) that allows adjustment of parameters in real-time. For examples, ingredient ratios, selection of fermentation agents, and modification of process variables such as temperature, pH, and fermentation duration are all potential adjustments. The UI provides visual feedback on the predicted changes to the beverage's organoleptic profile, enabling users to see how alterations impact flavor, aroma, texture, and appearance.

[0115] In some embodiments, the digital twin stores detailed information on the sequencing of the fermentation process and the timing of the introduction of fermentables. This may include introduction timing data on when to introduce specific ingredients to either enhance, slow down, or inhibit the activity of the cultures, post -fermentation additions information on ingredients that can be introduced post-fermentation without disrupting the final product quality, and manufacturing process flow that can be utilized to scale the beverage production at commercial facilities, ensuring consistency and efficiency.

[0116] For example, the fermentation outcome may be a beverage having specific conditions such as gut-friendly, alcoholic, non-alcoholic, without added sugars or synthetic flavours while still achieving a certain taste, or without chemical stabilizers. It may also be produced as having improved shelf stability without the use of chemical stabilizers. This could, for example, be created using naturalprebiotics, probiotics, and post biotics. The creation of recipe data leads to a prediction of flavour profile upon fermentation as described above.

[0117] According to yet another aspect of the present disclosure, a method is provided for predicting an organoleptic profile 300 for a fermentation outcome. In this forward-prediction method 300 where an organoleptic profile is predicted, the profile prediction method 300 involves obtaining data representing microbial-substrate interactions 102a of candidate microbial strains and candidate substrates. The microbial-substrate interaction data may be a combination of genomic data and metabolic pathway of candidate microbial strains under fermentation parameters and molecular precursor data of candidate substrates as described in the method 100. The profile prediction method 300 further involves training a machine learning model 108 using said microbial -substrate interaction data to predict metabolites produced by said candidate microbial strains with said candidate substrates under a variety of fermentation conditions also as in method 100.

[0118] Next, the profile prediction method 300 additionally involves providing an input of at least one microbial strain 302a, at least one substrate 302b, and at least one fermentation parameter 302c. As described above, the fermentation parameters may be selected from temperature, pH, oxygen availability, and nutrient parameters, or a combination thereof. The application of the trained machine learning model 308 involves generating predictions for various combinations of candidate microbial strains using their genomic data as input along with molecular precursor data for substrates and fermentation parameters.

[0119] Single and multi-strain predictions are contemplated as also described above. As in method100, the method here further involves applying the trained machine learning model to predict metabolites produced by combinations of the candidate microbial strains and the candidate substrates under the fermentation parameters.

[0120] The machine learning model outputs multi-dimensional phenotype vectors 120 representing predicted concentrations or production levels of various metabolites as earlier described (e.g., organic acids such as lactic acid and acetic acid for sourness, esters such as ethyl acetate for fruity notes, alcohols including ethanol and glycerol, phenolic compounds for spicy aromas, and other bioactive molecules and others).

[0121] Finally, the method involves generating a predicted organoleptic profile 150 based on the predicted metabolites produced. The predicted organoleptic profile encompasses the sensory and functional attributes described in or about paragraph

[0051] , Other features may also contribute to the organoloeptic profile as will be known to the skilled person. These predictions may also accompanied by confidence intervals or uncertainty estimates to guide decision -making in strain selection, enabling the rational design of fermentation processes that achieve targeted organoleptic profiles encompassingflavor, aroma, texture, mouthfeel, visual appearance, acidity, bitterness, prebiotic, probiotic, and postb iotic functionalities.

[0122] The above description has forth exemplary aspects of the present disclosure. It should be recognized, however, that such description is not intended as a limitation on the scope of the present disclosure. Rather, the description also encompasses combinations and modifications to those exemplary aspects described herein.

Claims

CLAIMS1 . A method for designing microbial consortia for fermentation, the method comprising:-obtaining data representing microbial -substrate interactions under fermentation parameters;-training a machine learning model using said microbial -substrate interaction data to predict metabolites produced by candidate microbial strains with candidate substrates under fermentation parameters;-receiving ataigeted organoleptic profile for a desired fermented comestible;-applying said trained machine learning model to predict metabolites produced by combinations of said candidate microbial strains and said candidate substrates under fermentation parameters; and-selecting at least one predicted microbial strain and at least one predicted substrate precursor for the microbial consortia, wherein said at least one predicted microbial strain with said at least one predicted substrate precursor is predicted to produce metabolites that contribute to said targeted oiganoleptic profile.

2. The method of claim 1, further comprising receiving designated-fermentation parameters prior to applying said trained machine learning model.

3. The method of claim 1, further comprising selecting at least one specified fermentation parameter for the microbial consortia for fermentation with a substrate.

4. The method of claim 1, further comprising evaluating the at least one predicted microbial strain and at least one predicted substrate precursor until a predetermined threshold is met to approximate an organoleptic profile.

5. The method of claim 1, wherein said microbial -substrate interaction data is a combination of genomic data and metabolic pathway of candidate microbial strains under fermentation parameters and molecular precursor data of candidate substrates.

6. The method of claim 5, wherein said genomic data of candidate microbial strains comprises latent embeddings derived from genomic sequences and metabolic pathway data and wherein said substrate data comprises substrate feature embeddings derived from molecular precursor data of said candidate substrates under fermentation parameters.

7. The method of claim 5, wherein said molecular precursor of candidate substrates comprises decomposing candidate substrates into molecular precursors capable of biotransformation by said candidate microbial strains.

8. The method of claim 5, wherein said obtaining data representing microbial -substrate interactions comprises generating pairwise embeddings representing microbial -substrate interactions by combining said latent embeddings with said substrate feature embeddings.

9. The method of claim 1, further comprising assessing feasibility of producing predicted metabolites through microbial fermentation.

10. A method for designing microbial consortia for fermentation of a substrate, the method comprising:- obtaining latent embeddings of genomic data of candidate microbial strains;- obtaining substrate feature embeddings derived from molecular precursors of candidate substrates;- generating pairwise embeddings representing microbial -sub state interactions by combining said latent embeddings with said substrate feature embeddings;- training a supervised machine learning model to predict produced microbial metabolites from said pairwise embeddings;- receiving a targeted organoleptic profile for a desired fermented comestible;- applying said trained supervised machine learning model to generate a predicted organoleptic profile value based on said pairwise embeddings representing microbial -substrate combinations;- applying residual modeling to calculate the difference between said predicted organoleptic profile value and said targeted organoleptic profile value;- iteratively evaluating candidate microbial strain and substrate combinations that reduce the difference between said predicted organoleptic profile value and said targeted organoleptic profile value; and- selecting at least one microbial strain and at least one substrate for designing the microbial consortia, wherein said at least one microbial strain and said at least one substrate are selected from said candidate microbial strain and substrate combinations predicted to produce said microbial metabolites that contribute to said taigeted organoleptic profile.

11. The method of claim 10, wherein said latent embeddings are obtained by unsupervised learning.

12. The method of claim 11, wherein said unsupervised learning is selected from sparse PCA, Bayesian PCA, Autoencoders, Variational Autoencoders (VAE), and Graph Autoencoders.

13. The method of claim 10, wherein supervised learning algorithms are selected from Gradient Boosted Trees, Random Forest, Support Vector Machines, Neural Networks, Logistic Regression, Ridge, LASSO, and Elastic Net.

14. The method of claim 11, wherein said residual modeling continues until a predetermined condition is met.

15. The method of claim 12, wherein said predetermined condition is selected from: residual magnitude below a specified threshold, maximum strain count reached, and negligible marginal improvement from additional strains.

16. The method of any one of claim 1 or claim 10, wherein said desired oiganoleptic profile comprises at least one of flavor, aroma, texture, mouthfeel, visual appearance, acidity, bitterness, prebiotic, probiotic, and postbiotic functionalities, or any combination thereof.

17. The method of any one of claim 5 or claim 10, wherein generating pairwise embeddings comprises at least one of concatenation, element -wise multiplication, and learned interaction layers via neural networks.

18. The method of claim 10, wherein residual modeling comprises applying greedy cascading residual optimization.

19. A method for generating a recipe for a fermented comestible, the method comprising:- receiving a target oiganoleptic profile for a fermented comestible;- designing microbial consortia for fermentation with a substrate using the method of claim 1 or claim 5; and- providing ratios of each of microbial strains within said consortia and substrate quantities to generate metabolites contributing to target organoleptic profile for recipe generation.

20. The method of claim 19, further comprising providing step-by-step instructions for preparing a fermented comestible and quality control framework for improving batch consistency in large-scale production.

21. A method for dynamic consortia control, the method comprising:-introducing a multi-strain microbial consortium into a fermentation substrate;-applying fermentation parameters to selectively promote or suppress individual strains; and-dynamically steering fermentation outcomes through controlled strain activity.

22. The method of claim 21, wherein said parameters are selected from environmental stressors and process controls or a combination thereof.

23. The method of claim 21, wherein said environmental stressors are selected from pH adjustment, temperature modulation, nutrient addition, and oxygen control or a combination thereof.

24. The method of claim 22, wherein said process controls are selected from metabolite accumulation and substrate.

25. The method of claim 24, wherein metabolite accumulation comprises using at least one of lactic acid, ethanol, and acetaldehyde as control signals to regulate strain activity.

26. The method of claim 23, wherein oxygen control comprises transitioning between aerobic and anaerobic conditions to modulate aroma compound production.

27. The method of claim 21 , wherein dynamically steering fermentation outcomes comprises producing taiget compounds selected from organic acids, esters, alcohols, and phenolic compounds or a combination thereof.

28. The method of claim 21, further comprising a step of monitoring fermentation progress.

29. The method of claim 28, wherein the step of monitoring fermentation progress is performed using at least one modality selected from the group consisting of: human sensory evaluation panels, electronic sensors, chemical analysis or chromatographic analysis, and combinations thereof.

30. The method of claim 29, wherein electronic sensors are selected from electronic nose and electric tongue sensors or a combination thereof.

31. The method of claim 29, wherein chemical or chromatographic analysis is selected from gas chromatography -mass spectrometry (GC-MS) and high-perfommace liquid chromatography (HPLC)32. A system for optimizing fermentation processes, the system comprising:-a processor and a memory coupled to the processor;-a non-transitory computer-readable storage medium storing instructions that, when executed by the processor, causes the system to:-collect fermentation data comprising ingredient compositions, fermentation agent characteristics, and organoleptic measurements;-generate digital twin representations of pre -fermentation organoleptic profiles and post-fermentation organoleptic profiles;-train a machine learning model using the digital twin representations to predict fermentation outcomes; and-output optimized fermentation parameters for achieving taiget fermented outcome characteristics.

33. The system of claim 32, wherein the digital twin representations comprise comprehensive digital representations that simulate sensory attributes, organoleptic attributes, chemical attributes, functional properties, and production data.

34. The system of claim 32, wherein the organoleptic measurements comprise data collected from electronic nose sensors and electronic tongue sensors.

35. The system of claim 32, wherein the digital twin representation simulates sensory attributes including flavor, aroma, texture, and appearance.

36. The system of claim 32, wherein the instructions further cause the system to predict resulting metabolites, esters, volatile organic compounds, and volatile acidity compounds using metabolic pathways of fermentation agents applied to input ingredient precursors.

37. The system of claim 32, wherein the digital twin representation includes fermentation process sequencing data comprising timing of ingredient introduction and post -fermentation additions.

38. The system of claim 32, wherein the digital twin representation includes cost of goods sold data that is dynamically updated based on real-time integration of supplier data, manufacturing processes, and market conditions.

39. The system of claim 32, wherein the digital twin representation includes carbon footprint data comprising scope 1, scope 2, and scope 3 emissions data for the fermented outcome.

40. The system of claim 32, further comprising a feedback loop for iterative model refinement based on experimental validation results.

41. A computer-implemented method of optimizing a fermentation process for a fermented outcome, comprising:- obtaining a pre-fermentation organoleptic profile of a fermented outcome made from a list of ingredients, and generating a digital twin representation of the pre-fermentation organoleptic profile;- obtaining a post-fermentation organoleptic profile of the fermented outcome after fermentation utilizing one or more selected fermentation agents and a given fermentation process, and generating a digital twin representation of the post -fermentation organoleptic profile;- training a machine learning model to analyze changes between the digital twin representation of the pre-fermentation organoleptic profile and the digital twin representation of the post -fermentation organoleptic profile; and- applying a machine learning model to predict a post -fermentation organoleptic profile given a list of ingredients, one or more selected fermentation agents, and a given fermentation process.

42. The computer-implemented method of claim 41, wherein said fermented outcome is abeverage.

43. A method for predicting an oiganoleptic profile for a fermentation outcome, the method comprising :- obtaining data representing microbial-substrate interactions of candidate microbial strains and candidate substrate;- training a machine learning model using said microbial -substrate interaction data to predict metabolites produced by said candidate microbial strains with said candidate substrates under a variety of fermentation conditions;- providing an input of at least one microbial strain, at least one substrate, and at least one fermentation parameter;-applying said trained machine learning model to predict metabolites produced by combinations of said candidate microbial strains and said candidate substrates under said fermentation parameters; and-generating a predicted oiganoleptic profile based on said predicted metabolites produced.

44. The method of claim 43, wherein fermentation parameters are selected from temperature, pH, oxygen availability, and nutrient parameters, or a combination thereof.

Images