A large language model (LLM) based microbial fermentation experiment intelligent design system and method

Abstract

The application discloses a resveratrol fermentation experiment intelligent design method and system based on a large language model (LLM). The method analyzes user natural language instructions through an LLM structure, calls historical data of resveratrol fermentation of Yarrowia lipolytica through retrieval enhancement technology to generate prompt words, combines real-time resource states of a laboratory, and uses an LLM to infer and output an optimal experiment design scheme. The optimized process designed by using the system can make the final concentration of resveratrol of Yarrowia lipolytica in a 60-ton fermentation tank reach 53.2 g / L. The application significantly improves the experiment design efficiency and resveratrol fermentation performance under industrial amplification conditions, is suitable for industrial fermentation amplification processes of 60-ton and above fermentation tanks, and reduces the product extraction cost by 37%.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioengineering and artificial intelligence, specifically relating to an intelligent design system and method for microbial fermentation experiments based on large language model (LLM), used for the automatic generation and execution of experimental schemes. Background Technology

[0002] Resveratrol is a non-flavonoid polyphenol compound with significant antioxidant, anti-aging, anti-tumor, and cardiovascular protective activities, making it highly valuable for commercial applications in the food, cosmetics, and pharmaceutical industries. With surging global market demand, the acquisition of resveratrol is rapidly shifting from traditional, inefficient plant extraction to green, high-yield microbial fermentation methods.

[0003] Yarrowia lipolytica, as an extremely important unconventional yeast, is widely recognized as an ideal chassis cell for the heterologous synthesis of hydrophobic secondary metabolites such as resveratrol due to its robust lipid metabolism network and abundant acetyl-CoA precursor pool. However, its actual industrialization faces significant technical bottlenecks: the optimization of fermentation processes involves highly coupled multidimensional variables, especially during the scaling-up from laboratory pilot-scale production (e.g., a 60-ton fermenter), where a significant "fermentation scale-up effect" exists. Under the hydrodynamic environment of a 60-ton fermenter, the consumption rate of complex carbon source substrates (e.g., the synergistic addition of glucose and glycerol), the non-uniform distribution of dissolved oxygen, and drastic fluctuations in ambient pH can easily prevent the effective redirection of carbon flux to the target product.

[0004] In traditional processes, directly scaling up parameters derived from laboratory trials to a 60-ton tank often results in a more than 40% decrease in the dissolved oxygen transfer coefficient (kLa) due to the significant difference in fluid shear forces between the large tank and the shake flask. Under these non-ideal fluid conditions, the carbon flux of *Yarrowia lipolytica* is readily redirected to byproducts (such as citric acid), making it difficult to replicate the high yields achieved in the laboratory during industrial scale-up.

[0005] On the other hand, although the modern bio-fermentation industry is gradually introducing digital platforms, they have not yet been widely used in structured scientific research experimental design, and there is a lack of in-depth logical mining and independent experimental design capabilities for massive historical fermentation data and multi-omics data. In recent years, LLM has shown strong capabilities in natural language understanding and complex logical reasoning, but it has not yet been effectively applied to the design of structured fermentation schemes for specific products (such as resveratrol).

[0006] Therefore, the industry urgently needs an intelligent design system for microbial fermentation. This system needs to deeply understand the natural language intent of researchers and automatically combine it with a private historical database for knowledge enhancement, thereby autonomously reasoning out standard task solutions that include optimal substrate ratios and dynamic process parameters. The goal is to overcome the challenge of industrial-scale scaling up of resveratrol and achieve a balance between high automation and ultra-high yield. Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide an intelligent design method and system for resveratrol fermentation experiments based on LLM.

[0008] Addressing the challenge of highly nonlinear multivariate coupling relationships during industrial fermentation scale-up (e.g., the nonlinear redirection effect of fermenter stirring power and dissolved oxygen level on acetyl-CoA precursor flux), this invention innovatively utilizes the attention mechanism of LLM to capture such implicit biochemical correlations. Specifically, the model can learn from massive historical data practical experience such as "in a 60-ton tank scale, when the glycerol flow acceleration rate is >5 g / (L·h), the stirring speed must be simultaneously increased to 300 rpm to maintain dissolved oxygen," thus effectively avoiding the direct transfer errors of traditional small-scale experience.

[0009] To achieve the objectives of this invention, the technical solution is as follows.

[0010] In a first aspect, the present invention provides an intelligent design method for microbial fermentation experiments based on LLM, comprising:

[0011] S1, the user's input natural language instructions are parsed into experimental objectives and constraints through LLM structured parsing;

[0012] S2, based on the parsed experimental objectives and constraints, automatically retrieves the user's historical experimental database. It adopts BERT vector embedding and multi-field weighting mechanism, and determines the preset similarity by performing a three-level weighted score on strain genotype, carbon source type and fermentation scale. It also integrates historical information that meets the preset similarity into enhancement conditions and outputs prompt words containing the enhancement conditions.

[0013] S3, obtain available resource information via API based on prompts containing enhanced conditions;

[0014] S4 encapsulates the experimental objectives and constraints parsed from S1, the prompts containing enhancement conditions output by S2, and the available resource information obtained by S3 into an integrated prompt. It then performs instruction reasoning through LLM and outputs the optimal experimental design scheme after multi-index collaborative optimization by combining the metabolic constraints of Yersinia lipophila and the industrial scaled-up fluid dynamics model.

[0015] S5 sends the optimal experimental design to the user and provides a visual interface for the user to confirm or modify the experimental design.

[0016] S6 converts the user-confirmed or modified experimental design into a task structure that the experimental system can recognize, and sends it to the experimental system to complete the execution of the experimental design.

[0017] Furthermore, the user's historical experiment database in S2 contains various information recorded in the user's historical experiments. The historical experiment data specifically refers to the experimental data on the production of resveratrol by fermentation of Yeast lipolyticis as a chassis cell, including but not limited to experimental objectives, strain information, culture medium information, process parameters, product information, SOP steps, and abnormal situations.

[0018] Furthermore, S2 employs BERT vector embedding to process historical experimental records. The application logic is as follows: given the highly domain-specific nature of bio-fermentation data, traditional keyword matching cannot understand the complex relationships between strain modification and metabolites. This invention utilizes the BERT model to transform textual descriptions into high-dimensional semantic vectors, capturing implicit biochemical associations through its deep attention mechanism, thereby achieving higher accuracy than traditional retrieval methods.

[0019] Furthermore, the three-layer weighted scoring described in S2 specifically employs a multi-dimensional nested scoring model. This includes keyword semantic matching, embedded vector semantic matching, and multi-field weighted matching. Specifically, the system extracts core entities from the user's experimental goals and constraints, uses BERT vector embedding technology to transform textual information into high-dimensional vectors, and calculates cosine similarity. Based on this, a comprehensive score is generated by combining domain-specific weighting factors.

[0020] First layer (strain genotype): assesses the genetic background consistency between historical strains and the target strain, with a preset weight of 0.4;

[0021] The second layer (carbon source type): assesses the similarity of substrate utilization pathways (such as the homology of polyphenol products), with a preset weight of 0.4;

[0022] The third layer (fermentation scale) focuses on evaluating the nonlinear coupling relationship between historical experiments and the target scale in terms of hydrodynamic characteristics (such as kLa), with a preset weight of 0.2. The system calculates the comprehensive similarity by weighting the total score of the above three layers, and sets a preset similarity threshold (such as 0.75) to filter out historical experimental records with real industrial scale-up reference value.

[0023] Furthermore, in S2, historical information that meets the preset similarity is generated and fused as an enhancement condition through multidimensional clustering statistics, parameter induction fusion, and natural language prompts.

[0024] Specifically, multidimensional clustering statistics involves clustering culture medium components or process parameters that appear frequently in the selected historical records and extracting recommended values.

[0025] Specifically, parameter summarization and fusion involves: extracting reference ranges for continuous parameters (such as temperature and pH) through interval analysis, and ranking candidate schemes in conjunction with historical yield indicators;

[0026] Specifically, the natural language prompt generation involves transforming the above-mentioned fused structured knowledge into enhanced prompt words in natural language form (e.g., "Based on historical high-yield experiments, it is recommended to use a glucose and glycerol complex substrate") as background knowledge for LLM inference.

[0027] Furthermore, the available resource information mentioned in S3 includes, but is not limited to, equipment information, microbial information, and material information. For example, a list of currently available fermenters (model, capacity, status), available strain inventory (number, purity, past yield), and culture medium inventory (components, batch number, cost).

[0028] Furthermore, if there are multiple indicators in the objective in S4, the optimal experimental design scheme is output through the Pareto optimal solution set.

[0029] The Pareto optimal solution set is calculated using the Pareto optimality rule, which is as follows: For each candidate experimental scheme, multiple optimization metrics (such as yield, time, cost, purity, etc.) are constructed into a multi-dimensional vector. Any two schemes are compared one by one. If scheme A is superior to scheme B in at least one metric and not inferior to scheme B in any other metric, then A dominates B. The dominated scheme is eliminated, and the remaining solutions constitute the Pareto optimal solution set. If multiple solutions exist within the set, the system can normalize and weight each metric according to user-defined or preset metric weights, and combine this with additional constraints such as resource availability and execution risk to select the final recommended optimal experimental design scheme from the set.

[0030] Furthermore, the experimental design scheme includes, but is not limited to, material selection, process parameters, equipment selection, and schedule arrangement.

[0031] Secondly, this invention provides an intelligent design system for resveratrol fermentation experiments based on LLM, comprising:

[0032] The intent parsing module is configured to execute step S1, which is used to parse the natural language instructions input by the user, including the product, strain and scale, into experimental objectives and constraints through LLM structured parsing.

[0033] The knowledge enhancement module is configured to execute step S2, which is used to automatically call the user's historical experimental database for retrieval based on the parsed experimental objectives and constraints. It adopts BERT vector embedding and multi-field weighting mechanism, and determines the preset similarity by performing a three-level weighted score on strain genotype, carbon source type and fermentation scale. It also integrates historical information that meets the preset similarity as enhancement conditions and outputs enhancement prompts containing substrate preferences and process parameter recommendations.

[0034] The resource scheduling module is configured to execute step S3, which is used to obtain available resource information via API based on prompts containing enhanced conditions, including but not limited to idle 60-ton industrial fermenters and their sensor configurations, microbial inventory and material information.

[0035] The experimental design module is configured to execute step S4, which encapsulates the parsed experimental objectives and constraints, prompts containing enhancement conditions, and available resource information into an integrated prompt, performs instruction reasoning through LLM, and outputs the optimal experimental design scheme after multi-index collaborative optimization by combining the metabolic constraints of Yeast lipolyticis and the industrial scale-up fluid dynamics model.

[0036] The human-machine collaboration module is configured to execute step S5, which is used to send the optimal experimental design scheme to the user and provide a visual interface for the user to confirm or modify the experimental design scheme, so as to ensure the safety and controllability of the scheme.

[0037] The task execution module is configured to execute step S6, which is used to convert the experimental design scheme confirmed or modified by the user into a task structure that can be recognized by the experimental system and send it to the automated fermentation execution system to complete the closed-loop execution of the experimental design scheme.

[0038] Thirdly, the present invention provides a computer-readable storage medium including a stored program, wherein, when the program is running, it controls the device where the computer-readable storage medium is located to execute the LLM-based intelligent design method for resveratrol fermentation experiments described in the first aspect.

[0039] Fourthly, the present invention provides an electronic device, including a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the electronic device is triggered to execute the LLM-based intelligent design method for resveratrol fermentation experiments described in the first aspect.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0041] 1. High automation: The entire process from intent to experiment execution is automated, greatly improving the efficiency of experiment design;

[0042] 2. Knowledge Enhancement: Combining RAG technology with private historical data ensures that the design solutions are based on solid evidence;

[0043] 3. Resource awareness: By combining real-time resource status, the feasibility of the solution and scheduling efficiency are improved;

[0044] 4. Structured output: It can generate standardized experimental task structures, which facilitates system integration;

[0045] 5. Human-machine collaboration mechanism: Retain the expert optimization interface to ensure the safety and controllability of experiments.

[0046] In summary, this invention provides an intelligent design system for resveratrol fermentation experiments. It can automatically retrieve historical databases (or private knowledge bases) and query experimental resource status based on the experimental objectives input by the user. It generates experimental plans through LLM reasoning and synchronizes them to the digital system with a task structure that the experimental system can recognize to support task confirmation, optimization, and execution. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the system module connection of the present invention, in which the dashed arrows indicate the data enhancement flow of historical experimental data and enhanced prompt words, and the solid arrows indicate the task execution flow of experimental design scheme generation and execution.

[0050] Figure 2 This is a flowchart illustrating an intelligent design method for resveratrol fermentation experiments based on LLM, provided in an embodiment of the present invention.

[0051] Figure 3 This is a time history curve comparing the resveratrol fermentation concentration predicted by the LLM intelligent design system with the actual fermentation concentration in a 60-ton fermenter, as shown in Embodiment 2 of the present invention.

[0052] Figure 4 This is a comparison curve of resveratrol fermentation concentration changes under three different process design strategies in an embodiment of the present invention. Detailed Implementation

[0053] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0055] The preferred embodiments of the present invention will now be described in detail with reference to specific examples. It should be understood that the following examples are given for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various modifications and substitutions to the present invention without departing from its spirit and essence.

[0056] Example 1

[0057] See Figure 1 As shown, this embodiment provides an intelligent design system for resveratrol fermentation experiments based on LLM, including:

[0058] The intent parsing module is configured to execute step S1, which is used to parse the natural language instructions input by the user, including the product, strain and scale, into experimental objectives and constraints through LLM structured parsing.

[0059] The knowledge enhancement module is configured to execute step S2, which is used to automatically call the user's historical experimental database for retrieval based on the parsed experimental objectives and constraints. It adopts BERT vector embedding and multi-field weighting mechanism, and determines the preset similarity by performing a three-level weighted score on strain genotype, carbon source type and fermentation scale. It also integrates historical information that meets the preset similarity as enhancement conditions and outputs enhancement prompts containing substrate preferences and process parameter recommendations.

[0060] The resource scheduling module is configured to execute step S3, which is used to obtain available resource information via API based on prompts containing enhanced conditions, including but not limited to idle 60-ton industrial fermenters and their sensor configurations, microbial inventory and material information.

[0061] The experimental design module is configured to execute step S4, which encapsulates the parsed experimental objectives and constraints, prompts containing enhancement conditions, and available resource information into an integrated prompt, performs instruction reasoning through LLM, and outputs the optimal experimental design scheme after multi-index collaborative optimization by combining the metabolic constraints of Yeast lipolyticis and the industrial scale-up fluid dynamics model.

[0062] The human-machine collaboration module is configured to execute step S5, which is used to send the optimal experimental design scheme to the user and provide a visual interface for the user to confirm or modify the experimental design scheme, so as to ensure the safety and controllability of the scheme.

[0063] The task execution module is configured to execute step S6, which is used to convert the experimental design scheme confirmed or modified by the user into a task structure that can be recognized by the experimental system and send it to the automated fermentation execution system to complete the closed-loop execution of the experimental design scheme.

[0064] Example 2

[0065] This embodiment provides an intelligent design method for resveratrol fermentation experiments based on LLM, such as... Figure 2 As shown, the specific steps are as follows.

[0066] S1 parses the natural language instructions input by the user into experimental objectives and constraints using LLM structured parsing.

[0067] Specifically, users can input their experimental intentions in natural language through a graphical or command-line interface, such as "I want to use Yersinia lipolytica to produce resveratrol through fermentation, with the highest possible yield in a 60-ton fermenter, and the time not exceeding 144 hours."

[0068] The system uses the connected LLM to perform structured understanding of the input statement and extract:

[0069] Experimental objective: To produce resveratrol

[0070] Strain preference: Yersinia lipophila

[0071] Size constraint: 60-ton fermentation tank

[0072] Optimization metric: Maximizing output

[0073] Time constraint: No more than 144 hours

[0074] The final output is structured content for downstream modules to use:

[0075]

[0076] S2 automatically retrieves the user's historical experimental database based on the parsed experimental objectives and constraints. It employs BERT vector embedding and a multi-field weighting mechanism, and determines the preset similarity by applying a three-level weighted score based on strain genotype, carbon source type, and fermentation scale. It then integrates historical information that meets the preset similarity as enhancement conditions and outputs prompts containing the enhancement conditions.

[0077] Specifically:

[0078] S21, Database Retrieval: Based on the experimental objective (resveratrol) and constraints (60-ton scale, Yersinia lipophila) parsed from S1, the system automatically calls the user's historical experimental database records on the local or cloud platform for a full search.

[0079] S22, Weighted similarity scoring: BERT vector embedding and multi-field weighting mechanism are used to score the similarity of the historical experimental database;

[0080] The specific execution process follows the semantic retrieval logic described in the aforementioned "Invention Content," utilizing the BERT model to transform the current experimental objective and constraints into high-dimensional semantic vectors. In this embodiment, the system employs a multi-dimensional nested scoring model, with the following weight allocation for each dimension: product homology (polyphenols) weight set to 0.4, strain genotypic consistency weight set to 0.4, and fermentation scale weight set to 0.2.

[0081] Specifically, for the scoring of the fermentation scale field, the system combines the actual decay characteristics of kLa at a scale of 60 tons with the nonlinear coupling effect of the large tank fluid dynamics environment on metabolic flux redirection for calculation correction, so as to ensure that the selected historical data has real reference value for industrial-scale production.

[0082] S23, Record Filtering: The system automatically calculates the weighted comprehensive similarity score of each historical record and sets a preset similarity threshold of 0.75 to filter out high-quality historical experimental records that are higher than this threshold.

[0083] S24, Information fusion generates enhanced conditions: The system performs multidimensional clustering statistics and parameter summarization on the selected historical records. In this embodiment, the system extracts the pattern that "glucose and glycerol complex carbon source" can significantly increase the flux of acetyl-CoA precursor, and summarizes the recommended process range (e.g., temperature 28℃-30℃, pH 5.5-6.5) and the corresponding dissolved oxygen control logic as enhanced conditions;

[0084] S25, Enhanced Prompt Output: The system converts the above enhancement conditions into prompts in natural language for subsequent modules to use. For example: "Based on historical high-yield experiments, it is recommended to use the high-yield engineered strain of Yersinia lipolytica, cultured in the range of 28℃-30℃, using a glucose and glycerol composite substrate, and preferably with a final resveratrol fermentation concentration target of 50 g / L or higher."

[0085] S3 retrieves available resource information via API based on prompts containing enhanced conditions.

[0086] Specifically:

[0087] Based on the prompts containing enhanced conditions output by S2, the system calls the experimental resource platform (LIMS / MES) via API to retrieve the following resource statuses and obtain available resource information:

[0088] Currently available strain inventory (number, batch, purity);

[0089] Optional culture medium components and costs;

[0090] List of available fermenters (model, capacity, current status);

[0091] Available sensors and control devices can be configured;

[0092] Output format as follows:

[0093]

[0094] S4 encapsulates the experimental objectives and constraints parsed from S1, the prompts containing enhanced conditions output from S2, and the available resource information obtained from S3 into an integrated prompt. It then performs instruction reasoning through LLM, combines the metabolic constraints of Yersinia lipophila with the industrial scale-up fluid dynamics model, and outputs the optimal experimental design scheme after multi-index collaborative optimization.

[0095] Specifically:

[0096] S41, Prompt word encapsulation: The system formats and encapsulates the production target (resveratrol) and time constraint (144 hours) parsed from S1, the knowledge-enhanced prompt words output from S2 (historical high-yield range and composite substrate experience), and the real-time resource status (FQ-60T fermenter, strain YL-RSV-M03) obtained from S3 to construct an integrated prompt word;

[0097] S42, Optimization Reasoning and Metabolic Constraint Evaluation: The system calls a finely tuned LLM for logical reasoning.

[0098] Metabolic constraint assessment: The LLM was fine-tuned to identify the lipid metabolism network characteristics of *Yarrowia lipolytica*, particularly the flux distribution patterns of the acetyl-CoA precursor pool. During inference, the model automatically correlated historical data: it was identified that while high concentrations of glycerol could significantly increase product flux, under 60-ton hydrodynamic conditions (risk of kLa decay), excessively high feed rates could lead to carbon flux redirection to the byproduct citric acid.

[0099] Multi-indicator collaborative optimization (Pareto optimization): Addressing the potential conflict between "high yield" and "low cost" objectives for users, LLM employs a multi-objective trade-off based on the Pareto optimization rule. The model uses multiple rounds of parameter combination reasoning based on historical experimental data correlations to evaluate the economics and expected yield under different concentration combinations of the composite carbon source (glucose and glycerol). Ultimately, while ensuring precursor supply, it selects the optimal solution that aligns with the current resource cost weighting.

[0100] S43, Output the optimal experimental design scheme: Based on the above reasoning, the system generates a complete industrial-grade process scheme, including:

[0101] (1) Basic culture medium formulation and composite substrate: The initial carbon source was determined to be a composite substrate of 40 g / L glucose and 20 g / L purified glycerol, and the nitrogen source was 10 g / L yeast extract and 5 g / L ammonium sulfate.

[0102] (2) Dynamic feeding strategy: When the residual sugar drops below 5 g / L, feeding is triggered. The feeding material is mainly high-concentration glycerol, and the volumetric flow rate is set to 5 g / (L·h). At the same time, 25% ammonia water is used to keep the pH constant at 6.0.

[0103] (3) Coupling control of dissolved oxygen and stirring: The ventilation rate is set to 0.5-1.0 vvm, and the lower limit threshold of dissolved oxygen (DO) is 30%. When DO approaches the threshold, the stirring speed is controlled to be increased nonlinearly in a stepwise manner between 200 rpm and 400 rpm.

[0104] S44, Solution Summary: The system ultimately generates a structured summary: "It is recommended to use *Yarrowia lipolyticis* strain, a glucose and glycerol composite substrate, temperature controlled at 29℃, pH maintained at 6.0, dissolved oxygen controlled above 30%, combined with variable frequency stirring, fermentation time of 144 hours, and equipment of FQ-60T. The expected final resveratrol concentration is 53.2 g / L."

[0105] S5 sends the optimal experimental design to the user and provides a visual interface for the user to confirm or modify the experimental design.

[0106] Specifically:

[0107] The system will display the design scheme in the form of tables / process curves;

[0108] Users can modify parameters or switch candidate schemes;

[0109] Provide risk warnings for sensitive items (such as the ammonia flow acceleration at a scale of 60 tons and drastic fluctuations in dissolved oxygen);

[0110] The changes will be fed back to the system for updating the task schedule.

[0111] S6 transforms the user-confirmed or modified experimental design into a task structure recognizable by the experimental system and sends it to the experimental system to complete the execution of the experimental design. Specifically:

[0112] The user clicks "Confirm Execution";

[0113] The system encodes the design scheme into a structured task JSON;

[0114] Automatically sent to the underlying automated fermentation production system (such as MES / LabOS);

[0115] The experimental platform automatically performs substrate preparation, inoculation, sterilization, fermentation, and data recording;

[0116] Results: The production system strictly followed the optimal process path designed by the LLM (Liquidity Management Model). After a 144-hour fermentation cycle, the final resveratrol concentration of *Yarrowia lipolytica* in a 60-ton fermenter reached 53.2 g / L. As a control group (without an LLM design): the conventional process, which directly scaled up from traditional small-scale experience, failed to effectively predict kLa decay and carbon flux deviation at a 60-ton scale. Under the same conditions, the final resveratrol concentration after 144 hours of fermentation was only 31.5 g / L.

[0117] Comparative data shows that the proposed solution increases yield by approximately 69% compared to traditional solutions, improves the system's estimated carbon source utilization rate by 68%, and reduces by-product accumulation by 52%. This demonstrates that LLM effectively overcomes the challenge of carbon flux redirection at the fluid dynamics scale and verifies the effectiveness and feasibility of the invention under industrial scale-up conditions.

[0118] To further verify the accuracy of the metabolic kinetic constraints and fluid dynamics model of this system, a time-dimensional yield comparison was performed on the entire fermentation cycle of Example 2. For example... Figure 3 As shown, the predicted curve provided by this system exhibits an extremely high goodness of fit with the actual sampling and detection curve of a 60-ton fermenter. At the end of 144 hours of fermentation, the system predicted a final yield of 53.8 g / L, while the actual detected yield was 53.2 g / L, with a prediction error of only about 1.1%. This set of hyperbolas demonstrates that the LLM of this invention can accurately predict the metabolic performance of microorganisms under nonlinear fluid dynamics environments, possessing extremely high reliability for industrial guidance.

[0119] Example 3

[0120] This embodiment demonstrates the system's multi-objective optimization capability when faced with the dual constraints of output and cost. The specific process is as follows:

[0121] S1, the user enters a new natural language command: "At a scale of 60 tons, seek the optimal balance between resveratrol production and raw material costs."

[0122] After parsing the S2 and LLM instructions, two conflicting sub-objectives were extracted: "maximizing yield" and "minimizing production cost." System searches of historical data and real-time market material prices revealed that while a high proportion of glycerol significantly improves conversion rates, the current procurement cost of industrial-grade glycerol is far higher than that of glucose.

[0123] S3, the system initiates the Pareto optimal solution set calculation strategy to predict parameter combinations for multiple candidate carbon source ratio schemes. Based on the multi-index optimization logic, LLM finally outputs a compromise recommendation scheme: adjusting the composite substrate ratio to 50 g / L glucose and 10 g / L glycogen.

[0124] Implementation Results: The system transformed the compromise solution into structured instructions and issued them to the automated control system of the 60-ton fermenter. Based on the logical deduction and prediction made by LLM based on fluid dynamics constraints and metabolic kinetics, the expected final concentration of resveratrol fermentation after 144 hours of operation under this process was 48.5 g / L. The system automatically calculated that although the yield was slightly lower than that of Example 2 (53.2 g / L), the raw material cost per batch was reduced by 27.4%, successfully achieving an optimal balance between yield and economic benefits. This demonstrates that the intelligent design system possesses advanced reasoning capabilities for flexible decision-making and dynamic process adjustment.

[0125] To visually demonstrate the comprehensive technical effect of the multi-objective optimization of this invention, the kinetic data of the traditional control group, the scheme of Example 2, and the scheme of Example 3 are plotted as comparative curves, such as... Figure 4 As shown in the figure, the synthesis rate of the traditional control group slowed down significantly after 96 hours of fermentation due to the hydrodynamic decay effect of the large tank; while the schemes of Example 2 and Example 3 of the present invention effectively broke through the metabolic bottleneck in the later stage. Although the yield of Example 3 was slightly lower than that of Example 2 in the later stage, it achieved a significant reduction in raw material costs while sacrificing a small yield, perfectly meeting the demand for Pareto optimality of yield and cost in actual industrial production.

[0126] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart design method for resveratrol fermentation experiments based on LLM, characterized in that, Includes the following steps: S1, the user's input natural language instructions are parsed into experimental objectives and constraints through LLM structured parsing; S2, based on the parsed experimental objectives and constraints, automatically retrieves the user's historical experimental database. It adopts BERT vector embedding and multi-field weighting mechanism, and determines the preset similarity by performing a three-level weighted score on strain genotype, carbon source type and fermentation scale. It also integrates historical information that meets the preset similarity into enhancement conditions and outputs prompt words containing the enhancement conditions. S3, obtain available resource information via API based on prompts containing enhanced conditions; S4 encapsulates the experimental objectives and constraints parsed from S1, the prompts containing enhancement conditions output from S2, and the available resource information obtained from S3 into an integrated prompt. It then performs instruction reasoning through LLM, combines the metabolic constraints of Yersinia lipophila with the industrial scale-up fluid dynamics model, and outputs the optimal experimental design scheme after multi-index collaborative optimization. S5 sends the optimal experimental design to the user and provides a visual interface for the user to confirm or modify the experimental design. S6 converts the user-confirmed or modified experimental design into a task structure that the experimental system can recognize, and sends it to the experimental system to complete the execution of the experimental design.

2. The intelligent design method for resveratrol fermentation experiments according to claim 1, characterized in that, The user's historical experiment database mentioned in S2 contains various information recorded in the user's historical experiments. The historical experiment data specifically refers to the experimental data on the production of resveratrol by fermentation of Yersinia lipolytica as a chassis cell, including but not limited to experimental objectives, strain information, culture medium information, process parameters, product information, SOP steps, and abnormal situations.

3. The intelligent design method for resveratrol fermentation experiments according to claim 1, characterized in that, The three-layer weighted scoring described in S2 is specifically measured using a multi-dimensional nested scoring model.

4. The intelligent design method for resveratrol fermentation experiments according to claim 1, characterized in that, In S2, historical information that meets the preset similarity is fused into enhancement conditions through multidimensional clustering statistics, parameter induction fusion, and natural language prompt template generation.

5. The intelligent design method for resveratrol fermentation experiments according to claim 1, characterized in that, The available resource information mentioned in S3 includes, but is not limited to, equipment information, microbial information, and material information.

6. The intelligent design method for resveratrol fermentation experiments according to claim 1, characterized in that, If the experimental objective in S4 includes multiple conflicting indicators such as output, cost, or cycle time, the LLM system performs multi-objective optimization based on the Pareto optimization rule. By performing non-dominated sorting on different combinations of process parameters, it generates a set of Pareto optimal solutions and selects the optimal experimental design scheme that meets the user's preference weights.

7. The intelligent design method for resveratrol fermentation experiments according to claim 1, characterized in that, The optimal experimental design scheme includes material selection, process parameters, equipment selection, and schedule arrangement.

8. A smart design system for resveratrol fermentation experiments based on LLM, characterized in that, include: The intent parsing module is configured to perform the steps of S1 as described in claim 1; The knowledge enhancement module is configured to perform the steps of S2 as described in claim 1; The resource scheduling module is configured to perform the steps of S3 as described in claim 1; The experimental design module is configured to perform the steps of S4 as described in claim 1; The human-machine collaboration module is configured to perform the steps of S5 as described in claim 1; The task execution module is configured to perform the steps of S6 as described in claim 1.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is running, it controls the device where the computer-readable storage medium is located to execute the intelligent design method for resveratrol fermentation experiments as described in any one of claims 1-7.

10. An electronic device, characterized in that, It includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the electronic device is triggered to execute the intelligent design method for resveratrol fermentation experiment as described in any one of claims 1-7.

Images