Trained model, calculation device, calculation method, calculation program, and non-transitory storage medium

Abstract

Provided are a trained model, a calculation device, a calculation method, a calculation program, and a non-transitory storage medium that enable accurate calculation of culture-related concentrations using a biochemical reaction model. The trained model is trained using culture conditions serving as explanatory variables for when cells are cultured in a culture vessel, and a parameter of a physical model serving as the response variable that is used when a relevant concentration relevant to the culture conditions is calculated. The calculation device is provided with a culture condition input unit, a parameter acquisition unit, and a first concentration calculation unit. The culture condition input unit inputs a first culture condition as an explanatory variable to the trained model. When the first culture conditions are inputted to the trained model, the parameter acquisition unit acquires a first parameter outputted as the response variable from the trained model. The first concentration calculation unit calculates a first relevant concentration relevant to the first culture condition on the basis of the acquired first parameter and the physical model.

Description

Trained model, calculation device, calculation method, calculation program, and non-transitory storage medium

[0001] The present disclosure relates to a trained model, a calculation device, a calculation method, a calculation program, and a non-transitory storage medium.

[0002] Conventionally, under culture conditions in which cells are cultured in a culture vessel (e.g., a culture tank), concentrations related to the culture (e.g., cell concentration, substrate concentration, product concentration, by-product concentration, inhibitory product concentration, etc., hereinafter also referred to as related concentrations) are calculated using, for example, a biochemical reaction model (physical model).

[0003] In order to accurately calculate the relevant concentrations using a biochemical reaction model, it is necessary to optimize the parameters (model constants) used in the biochemical reaction model by parameter fitting (optimization calculation) using the actual measured values ​​of the relevant concentrations prior to the calculation.

[0004] The fitted parameters were then used to calculate concentrations under multiple different culture conditions, but even slight changes in the culture conditions could change the optimal parameters, making it difficult to accurately calculate the relevant concentrations.

[0005] Patent Document 1 discloses a technology for estimating the quality and properties of a target substance produced in a culture process in real time and feeding the quality and properties back to culture control. In the technology disclosed in Patent Document 1, after individual control values ​​are input, the quality and properties of the target substance in the culture tank are estimated using the control values ​​and a statistical numerical calculation model, and if the quality and properties fall below a preset reference value, the control target value is corrected.

[0006] JP 2019-110767 A

[0007] An object of the present disclosure is to provide a trained model, a calculation device, a calculation method, a calculation program, and a non-transitory storage medium that are capable of accurately calculating concentrations related to culture using a biochemical reaction model.

[0008] According to the present disclosure, there is provided a trained model having the following configuration: [1] A trained model that has been trained using culture conditions when culturing cells in a culture vessel as explanatory variables and parameters of a physical model used when calculating relevant concentrations related to the culture conditions as objective variables.

[0009] According to the present disclosure, culture conditions (explanatory variables) can be input into a trained model to obtain parameters of a physical model output as target variables; that is, parameters optimized according to the culture conditions can be obtained, and the culture-related concentrations can be calculated with high accuracy using the parameters and the physical model (biochemical reaction model).

[0010] Various embodiments of the present disclosure are exemplified below. The embodiments described below can be combined with each other. [2] A calculation device including a culture condition input unit, a parameter acquisition unit, and a first concentration calculation unit, wherein the culture condition input unit inputs a first culture condition as the explanatory variable to the trained model described in [1], the parameter acquisition unit acquires a first parameter output from the trained model as the objective variable when the first culture condition is input to the trained model, and the first concentration calculation unit calculates a first related concentration related to the first culture condition based on the acquired first parameter and the physical model. [3] The calculation device according to [2], further comprising a first setting unit, a second setting unit, and a second concentration calculation unit, wherein the first setting unit sets, in a calculation model, an equation related to the first related concentration and taking into account advection or diffusion between a first region in the culture vessel and a second region adjacent to the first region; the second setting unit sets, in the calculation model, a concentration change equation that uses the first parameter and shows a change over time of the first related concentration in the first region; and the second concentration calculation unit calculates, based on the set calculation model, a second related concentration in the first region that is related to the first culture condition. [4] The calculation device according to [3], further comprising an actual measurement value acquisition unit, a parameter calculation unit, and a learning unit, wherein the actual measurement value acquisition unit acquires actual measurement values ​​of a related concentration related to the first culture condition in the first region, the parameter calculation unit calculates parameters of the physical model so as to reduce a difference between the acquired actual measurement values ​​and the second related concentration, and the learning unit trains the trained model using the calculated parameters and the first culture condition as training data. [5] The calculation device according to [2], further comprising a second trained model, a second culture condition input unit, and a third related concentration acquisition unit, wherein the second trained model is trained using a second culture condition related to the first culture condition as an explanatory variable and the related concentration as a target variable, the second culture condition input unit inputs the second culture condition to the second trained model as the explanatory variable, and the third related concentration acquisition unit acquires the third related concentration output from the second trained model as the related concentration.[6] A calculation method having a culture condition input step, a parameter acquisition step, and a first concentration calculation step, wherein in the culture condition input step, first culture conditions are input as the explanatory variables to the trained model described in [1], in the parameter acquisition step, when the first culture conditions are input to the trained model, a first parameter output from the trained model as the objective variable is acquired, and in the first concentration calculation step, a first related concentration related to the first culture condition is calculated based on the acquired first parameter and the physical model. [7] A calculation program that causes a processor to execute the calculation method described in [6]. [8] A computer-readable non-transitory storage medium storing a calculation program that causes a processor to execute the calculation method described in [6].

[0011] According to the present disclosure, concentrations related to culture can be calculated with high accuracy using a biochemical reaction model.

[0012] FIG. 1 is a block diagram showing a hardware configuration of a calculation device in this embodiment. FIG. 2 is a block diagram showing an example of the functional configuration of a control unit provided in the calculation device in this embodiment. FIG. 3 is a diagram showing an example of the configuration of a culture vessel in this embodiment. FIG. 4 is a diagram showing a plurality of regions into which the contents contained in the culture vessel in this embodiment are divided. FIG. 5 is a flowchart showing an example of learning data preparation processing in this embodiment. FIG. 6 is a flowchart showing an example of learning processing in this embodiment. FIG. 7 is a flowchart showing an example of first concentration calculation processing in this embodiment. FIG. 8 is a flowchart showing an example of second concentration calculation processing in this embodiment. FIG. 9 is a flowchart showing an example of relearning processing in this embodiment. FIG. 10 is a block diagram showing a modified example of the functional configuration of a control unit provided in the calculation device in this embodiment.

[0013] Embodiments of the present disclosure will be described with reference to the accompanying drawings, in which the same reference numerals denote the same or similar configurations.

[0014] FIG. 1 is a block diagram showing the hardware configuration of a calculation device 10 according to this embodiment. The calculation device 10 calculates relevant concentrations related to culture conditions when culturing cells in a culture vessel. Here, the cells are not limited to, for example, bacterial cells, but also include cells extracted from animals or plants, and single-celled organisms (microorganisms). Examples of relevant concentrations include cell concentration, substrate concentration, product concentration, by-product concentration, and inhibitor product concentration. Furthermore, relevant concentrations may also be the dissolved oxygen concentration, dissolved carbon dioxide concentration, or dissolved hydrogen ion concentration contained in the culture solution.

[0015] The culture vessel is used as a culture tank for culturing microorganisms, animal cells, etc. in a liquid by supplying oxygen, etc. or degassing carbon dioxide, etc., and is used in the fields of petrochemistry, chemistry, pharmaceuticals, life sciences, etc.

[0016] Next, a problem solved by the calculation device 10 in this embodiment will be described. Conventionally, under culture conditions in which cells are cultured in a culture vessel, concentrations related to the culture (relevant concentrations) are calculated using, for example, a biochemical reaction model. To accurately calculate the relevant concentrations using a biochemical reaction model, it is necessary to optimize the parameters (model constants) used in the biochemical reaction model through parameter fitting (optimization calculation) using actual measured values ​​of the relevant concentrations prior to the calculation. Once fitted, the parameters are then used to calculate the relevant concentrations for multiple different culture conditions. However, even slight changes in the culture conditions can change the optimal parameters, making it difficult to accurately calculate the relevant concentrations using this method.

[0017] To address the above problem, it is conceivable to obtain actual measured values ​​of the relevant concentrations under the changed culture conditions each time the culture conditions are changed and then optimize the parameters used in the biochemical reaction model. However, in this case, the cost required to accurately calculate the relevant concentrations increases.

[0018] Therefore, in order to solve the above problem, the calculation device 10 in this embodiment is configured to accurately calculate the relevant concentrations related to culture using a biochemical reaction model, and moreover, to reduce the cost required for accurately calculating the relevant concentrations. As shown in Figure 1, the calculation device 10 is configured to include a control unit 11, a storage unit 12, a communication unit 13, an operation input unit 14, and a display unit 15.

[0019] The control unit 11 is, for example, a CPU (Central Processing Unit), a microprocessor, a DSP (Digital Signal Processor), or the like, and controls the overall operation of the calculation device 10 .

[0020] A part of the storage unit 12 is configured with, for example, a RAM (Random Access Memory) or a DRAM (Dynamic Random Access Memory), and is used as a work area when the control unit 11 executes processes based on various programs.

[0021] Furthermore, a part of the storage unit 12 is, for example, a non-volatile memory such as a ROM (Read Only Memory) or an HDD (Hard Disk Drive), and stores various data and programs used in the processing of the control unit 11. The storage unit 12 can hold a database including one or more tables for recording various information, processing results, and the like.

[0022] The programs stored in the storage unit 12 include, for example, an operating system (OS) for implementing the basic functions of the calculation device 10, drivers for controlling various hardware, and programs for implementing various functions, including a program that functions as the "calculation program" of the present disclosure. This calculation program executes the characteristic processing of this embodiment, i.e., a series of algorithms for predicting parameters of a physical model from input culture conditions and using the predicted parameters to calculate relevant concentrations, as well as the construction and update of a trained model. This calculation program may be stored and provided on a computer-readable non-transitory storage medium such as a CD-ROM, DVD, or semiconductor memory, and the storage unit 12 of the calculation device 10 can function as an example of such a non-transitory storage medium. The control unit 11 (processor) reads and executes the calculation program from the storage medium to implement the "calculation method" of the present disclosure.

[0023] The communication unit 13 is, for example, a network interface controller (NIC) and has a function of connecting to a communication line (not shown). Instead of or in addition to the NIC, the communication unit 13 may have a function of connecting to a wireless local area network (LAN), a function of connecting to a wireless wide area network (WAN), a function of enabling short-range wireless communication such as Bluetooth (registered trademark), and infrared communication. The computing device 10 is connected to an external device (not shown) via a communication line, such as a database server storing experimental data, a control system controlling a culture process, another analysis system, or cloud-based computing resources, and can transmit and receive various data to and from the external device. This makes it possible, for example, to automatically acquire new culture data from another database and add it to the training data, or to transmit calculated related concentrations and prediction results to other systems for use. It is also possible that the calculation program and the trained model are downloaded via a network and stored in the memory unit 12.

[0024] The operation input unit 14 is configured with a keyboard, a mouse, and the like, and accepts input of various operations by the user of the computing device 10 .

[0025] The display unit 15 is, for example, a liquid crystal display device, and displays various images.

[0026] The control unit 11, storage unit 12, communication unit 13, operation input unit 14, and display unit 15 are electrically connected to one another via a system bus 16. Therefore, the control unit 11 can access the storage unit 12, display images on the display unit 15, grasp the operation status of the operation input unit 14 by the user, and access various communication networks and external devices via the communication unit 13.

[0027] FIG. 2 is a block diagram showing an example of the functional configuration of the control unit 11 included in the calculation device 10 according to this embodiment.

[0028] 2, the control unit 11 includes, as its functional configuration, an actual measurement value acquisition unit 11a, a parameter calculation unit 11b, a learning unit 11c, a culture condition input unit 11d, a parameter acquisition unit 11e, a first concentration calculation unit 11f, a first setting unit 11g, a second setting unit 11h, and a second concentration calculation unit 11i. Note that, although the control unit 11 generally has various functions in addition to those described above, only the characteristic functions of the calculation device 10 of this embodiment will be described here, and illustration and description of other known functions will be omitted.

[0029] The actual measurement value acquisition unit 11a acquires actual measurement values ​​of relevant concentrations (e.g., cell concentration, substrate concentration, product concentration, by-product concentration, inhibitory product concentration, dissolved oxygen concentration, dissolved carbon dioxide concentration, dissolved hydrogen ion concentration, etc.) related to the culture conditions when culturing cells in a culture vessel.

[0030] Fig. 3 is a diagram showing an example of the configuration of a culture vessel 20 according to this embodiment. As shown in Fig. 3, the culture vessel 20 contains a content 21 (e.g., a culture solution) and is provided with an aeration sparger 23 that blows (aerates) a gas 22 (e.g., oxygen) into the content 21, and a stirring blade 24 (agitator) that rotates in the direction of rotation in the figure to apply an external force to the content 21 in the culture vessel 20, thereby efficiently stirring the content 21 and the gas 22.

[0031] In the present embodiment, the actual measurement value acquiring unit 11a acquires actual measured values ​​of relevant concentrations by measuring, in the culture vessel 20 during actual operation, relevant concentrations related to the culture conditions when cells are cultured in the culture vessel 20, for one or more regions out of a plurality of regions (a plurality of positions) obtained by dividing the contents 21 contained in the culture vessel 20. Here, the culture conditions include data on the shape of the agitator, data on the shape of the aeration sparger 23, data on the shape of the culture vessel 20, operating conditions in the culture vessel 20 (e.g., operating conditions of the aeration impeller 24, aeration conditions of the aeration sparger 23), physical properties, composition, chemical properties, biochemical properties, or concentration of the contents 21 contained in the culture vessel 20, the elapsed time since the start of culture, information on the cells to be cultured (omics information), and culture conditions in a culture step performed on the contents 21 prior to the current culture step (operating conditions, medium composition, culture time in a pre-culture step), etc.

[0032] Data relating to the shape of the agitator includes, for example, data relating to the shape of the agitator blades 24 constituting the agitator, such as the type of agitator blades 24, the agitator blade diameter, and the number of agitator blades, which are set according to the shape of the agitator blades 24 (paddle (inclined / not inclined), disk turbine (inclined / not inclined), propeller, anchor, etc.). The data relating to the shape of the agitator blades 24 includes data in which a numerical label is attached for each type of shape, or includes data quantified by assigning a first numerical value (e.g., 1) to the area occupied by the agitator blades 24 in three-dimensional space and a second numerical value (e.g., 0) to the area not occupied by the agitator blades 24 in three-dimensional space. The shape data of the agitator blades 24 quantified by the first numerical value and the second numerical value can be expressed as a single numerical value, for example, by taking the average value and standard deviation for the entire device (culture vessel 20). It is also possible to combine a method of assigning a numerical label to each type of shape of the agitator blade 24 with a method of quantifying by assigning a first numerical value (e.g., 1) to the area occupied by the agitator blade 24 in three-dimensional space and a second numerical value (e.g., 0) to the area not occupied by the agitator blade in three-dimensional space. In this case, by assigning a numerical label to each type of shape of the agitator blade 24 (paddle (inclined / not inclined), disk turbine (inclined / not inclined), propeller, anchor, etc.) and further expressing the presence or absence of occupation of the agitator blade 24 in three-dimensional space by a numerical value, the shape of the agitator blade 24 can be expressed more precisely as numerical data.

[0033] Examples of data regarding the shape of the aeration sparger 23 provided in the culture vessel 20 include the type of aeration sparger 23 set according to the sparger shape (ring-shaped, cylindrical, circular), the aeration sparger diameter, the aeration sparger hole diameter, and the number of aeration sparger holes. The data regarding the shape of the aeration sparger 23 includes data in which a numerical label is attached for each type of shape, or includes data quantified by assigning a first numerical value (e.g., 1) to an area occupied by the sparger in three-dimensional space and a second numerical value (e.g., 0) to an area not occupied by the sparger in three-dimensional space.

[0034] The data regarding the shape of the culture vessel 20 is data regarding the shape of the vessel itself that contains the contents 21. The data regarding the vessel shape includes, for example, the type of vessel shape (cylindrical, rectangular, etc.), the vessel dimensions (inner diameter (in the case of a cylinder), width (in the case of a rectangle), height, etc.), and the presence or absence of a baffle (in the case of a stirring vessel).

[0035] Data regarding the physical properties, composition, chemical properties, biochemical properties, or concentration of contents 21 includes, for example, the physical properties, composition, chemical properties, and biochemical properties specific to the materials contained in contents 21, the initial values ​​of the physical properties of the entire contents within culture vessel 20, and the initial values ​​of the concentrations of each material contained in contents 21.

[0036] The physical properties inherent to the material contained in the contents 21 include, for example, the density, viscosity, elasticity, non-Newtonianity, etc. inherent to the material. The composition inherent to the material contained in the contents 21 includes, for example, the elements constituting the material and their contents, the ions dissolved in the material and their contents, the presence or absence of surfactants, the sugar content, the amino acid content, the vitamin content, the inorganic salt content, the lipid component content, and the trace metal content. The chemical properties inherent to the material contained in the contents 21 include, for example, the reaction activity, activation energy, reaction rate, reaction rate constant, etc. in the chemical reaction of the material. The biochemical properties inherent to the material contained in the contents 21 include, for example, the presence or absence of mycelium formation (if the material is a cell line), the doubling time, and the growth factors contained (if the material is a culture solution). The materials contained in the contents 21 are, for example, a culture medium / culture solution and cells. In this case, the physical properties, composition, chemical characteristics, biochemical characteristics, or concentrations inherent to the materials contained in the contents 21 include the fluid physical properties of the culture medium / culture solution (density, viscosity, non-Newtonianity, etc.), the culture medium composition (presence or absence of surfactants, type of growth factor, concentration of sugars contained in the culture medium, concentration of amino acids contained in the culture medium, concentration of vitamins contained in the culture medium, concentration of inorganic salts contained in the culture medium, concentration of lipid components contained in the culture medium, concentration of trace metals contained in the culture medium, etc.), and the type of cell line / strain (mass, viscosity, elasticity, size, shape, presence or absence of mycelium formation, doubling time, etc.). The physical properties, composition, chemical characteristics, and biochemical characteristics inherent to the materials contained in the contents are parameters that do not change even after the device containing the contents is operated, or parameters whose values ​​under operating conditions such as temperature and pressure are known.

[0037] The initial values ​​of the physical properties of the entire contents in the culture vessel 20 include, for example, the temperature, osmotic pressure, oxidation-reduction potential (ORP), and pH of the contents at the start of operation of the culture tank (culture vessel 20). The initial values ​​of the concentrations of the materials contained in the contents 21 include, for example, the seeding amount and cell concentration at the start of operation of the culture tank (culture vessel 20).

[0038] The actual measurement value acquiring unit 11a may also acquire other data such as gas superficial velocity, aeration rate, turbulent energy dissipation rate, gas hold-up, bubble diameter, dissolved oxygen (DO) control method (upper and lower limits, etc.), agitator rotation speed, culture medium inlet / outlet conditions, shear stress, and culture medium feed rate / perfusion rate (in the case of fed-batch culture / perfusion culture).

[0039] The gas superficial velocity is a set value calculated from the aeration rate and vessel diameter entered by the user. The aeration rate, rotation speed, dissolved oxygen control method, and fed-batch rate / perfusion rate (culture medium input / output conditions) are set by values ​​entered by the user. The turbulent energy dissipation rate is a value theoretically calculated from P / V (agitation power) or the in-apparatus mean value + standard deviation value obtained by multi-point measurement techniques in CFD simulations or experiments. The gas holdup is a value calculated from the liquid level difference with / without aeration or the in-apparatus mean value + standard deviation value obtained by multi-point measurement techniques in CFD simulations or experiments. The bubble diameter is a value measured using photographic images, etc., or the in-apparatus mean value + standard deviation value obtained by multi-point measurement techniques in CFD simulations or experiments. The shear stress is set by a value entered by the user or the in-apparatus mean value + standard deviation value obtained by multi-point measurement techniques in CFD simulations or experiments.

[0040] The culture conditions may include information about the cells being cultured (so-called omics information). The cellular information may include information about intracellular substances or cellular activity, such as the genome (gene information), transcriptome (mRNA information), proteome (protein information), metabolome (metabolite information), interactome (information about biomolecular interactions), and cellome (information about the transmission of substances and information within and between cells). The culture conditions may include individual pieces of information such as the genome, or may include the results of analyzing omics information including the genome, transcriptome, proteome, metabolome, interactome, and cellome. Furthermore, such cellular information may constitute one of the culture conditions as information indicating mutations, genetic modifications, and the like that have occurred in the target cells.

[0041] In addition, the actual measurement value acquisition unit 11a may acquire the actual measurement value of the related concentration by measuring the related concentration related to the culture conditions once under the same culture conditions for one or more of the multiple regions into which the contents 21 are divided, or may acquire the actual measurement value of the related concentration by measuring the related concentration multiple times.

[0042] 4 is a diagram showing a plurality of regions into which the contents 21 contained in the culture vessel 20 in this embodiment are divided. As shown in FIG. 4, the contents 21 are divided into 36 regions including, for example, regions 21A, 21B, 21C, 21D, and 21E. Note that the contents 21 contained in the culture vessel 20 may be divided into fewer than 36 regions or more than 36 regions.

[0043] The parameter calculation unit 11b calculates parameters of a biochemical reaction model (corresponding to the "physical model" in this disclosure) used to calculate relevant concentrations related to culture conditions by optimization calculation so that the actual measured values ​​of the relevant concentrations acquired by the actual measured value acquisition unit 11a can be calculated.

[0044] As a biochemical reaction model, for example, a model expressed by the following equations (1) and (2) is used, with respect to the cell concentration (viable cell concentration) as the relevant concentration: where X is the cell concentration [g cell / l], μ is the specific growth rate [1 / h = g cell / l / h / (g cell / l)], and μ ｍａｘ is the maximum specific growth rate [1 / h], Ks is the affinity constant [g substrate / l], and kd is the death rate constant [1 / h].

[0045]

[0046]

[0047] Furthermore, with respect to the substrate concentration as the relevant concentration, the model expressed by the following equation (3) is used: where S is the substrate concentration [g substrate / l], and Y Ｇwhere m is the growth yield [g cell / g substrate] and m is the maintenance constant [g substrate / l / h / (g cell / l)]. The substrate concentration may include the concentration of the limiting substrate. The substrate concentration may also include the concentration of cofactors such as inorganic salts.

[0048]

[0049] Furthermore, for the product concentration as the relevant concentration, a model expressed by the following equation (4) is used: where P is the product concentration [g product / l], α is a constant representing the linkage between product production and cell growth [g product / l / (1 / h)], and β is a constant representing the linkage between product production and cell concentration X [g product / l / (g cell / l)].

[0050]

[0051] Furthermore, with respect to the inhibitory product concentration as the relevant concentration, the model expressed by the following equation (5) is used: ｉ is the inhibitory product concentration [g product / l], and α ｉ is a constant [g product / l / (1 / h)] that represents the coupling between the production of inhibitory products and cell growth, and β ｉ is a constant [g product / l / (g cell / l)] that represents the correlation between the production of inhibitory products and the cell concentration X.

[0052]

[0053] The parameter calculation unit 11b calculates the actual measured values ​​of the relevant concentrations acquired by the actual measured value acquisition unit 11a in the above-mentioned biochemical reaction model by calculating the affinity constant Ks, the death rate constant kd, the growth yield Y Ｇ , the maintenance constant m, and the constants α and β are calculated by optimization calculation.

[0054] Furthermore, for the dissolved oxygen concentration as the relevant concentration, a model expressed by the following equation (6) is used: where DO is the dissolved oxygen concentration, X is the cell concentration, and α ＤＯ is a constant that represents the link between cell growth and the increase or decrease of oxygen, and β ＤＯ is a constant that represents the relationship between the increase or decrease in oxygen and the cell concentration. ２ ) and dissolved hydrogen ion concentration (C Ｈ＋ ) can also be expressed by an equation consisting of a term representing the linkage between cell growth and oxygen increase / decrease, and a term representing the linkage between oxygen increase / decrease and cell concentration, similar to equation (6).

[0055] The parameter calculation unit 11b calculates the actual measured value of the relevant concentration acquired by the actual measured value acquisition unit 11a in the above-mentioned biochemical reaction model by calculating a constant α ＤＯ and β ＤＯ is calculated by optimization calculation.

[0056] However, the biochemical reaction model for calculating the relevant concentrations is not limited to the above example. The biochemical reaction model may be, for example, an enzyme reaction model, or a more detailed model including multiple metabolic pathways and metabolic reactions.

[0057] Examples of optimization calculation algorithms include the least squares method and weighted least squares method, gradient descent search method, Newton's method (Newton-Raphson method), Gauss-Newton method, Levenberg-Marquardt method, genetic algorithm, ant colony optimization (ACO), simulated annealing (SA), memetic algorithm (MA), evolutionary computation method, evolutionary strategy method, and evolutionary programming.

[0058] The parameter calculation unit 11b stores data indicating pairs of the calculated parameters of the biochemical reaction model and the culture conditions corresponding to the actual measured values ​​of the relevant concentrations used to calculate the parameters by optimization calculation (the culture conditions in the culture vessel 20 when the actual measured values ​​were measured) as learning data in the storage unit 12. In this embodiment, data indicating multiple pairs of parameters of the biochemical reaction model and the culture conditions corresponding to the actual measured values ​​of the relevant concentrations is stored in the storage unit 12 as learning data.

[0059] The learning unit 11c performs machine learning of a learning model, using the training data (teacher data) stored in the memory unit 12 by the parameter calculation unit 11b, to output, as a response variable, parameters of a biochemical reaction model used to calculate relevant concentrations related to the culture conditions (parameters for one or more regions among the multiple regions into which the contents 21 contained in the culture vessel 20 are divided) when culture conditions for culturing cells in the culture vessel 20 are input as explanatory variables (input data). That is, the learning unit 11c inputs the training data into the learning model, and generates a trained model by having the learning model perform machine learning (so-called supervised learning) on ​​the correlation between the input data (culture conditions) and output data (parameters of the biochemical reaction model) that constitute the training data. The learning unit 11c stores the generated trained model in the memory unit 12.

[0060] The trained model may be, for example, a convolutional neural network (CNN). The machine learning learning algorithm may be, for example, deep learning. In this embodiment, the trained model is described as a CNN. However, the trained model is not limited to a CNN, and may be trained using any learning algorithm, such as a neural network other than a CNN, a support vector machine (SVM), a Bayesian network, or a regression tree.

[0061] The culture condition input unit 11d retrieves the trained model generated by the learning unit 11c from the storage unit 12 and inputs first culture conditions (new culture conditions) as explanatory variables into the trained model. The first culture conditions include, for example, a value indicating shape information regarding the shape of the agitator blades 24 (e.g., a1), a value indicating shape information regarding the shape of the aeration sparger 23 (e.g., b1), a value indicating the medium composition (e.g., c1), a value indicating the viscosity of the medium (e.g., d1), a value indicating the initial bacterial cell concentration (e.g., e1), and a value indicating the elapsed time from the start of culture (e.g., f1). The operation input unit 14 accepts an operation to input the first culture conditions from a user of the computing device 10 and outputs the input first culture conditions to the culture condition input unit 11d.

[0062] When the first culture condition is input into the trained model by the culture condition input unit 11d, the parameter acquisition unit 11e acquires the first parameter (a parameter optimized in one or more of the multiple regions into which the contents 21 contained in the culture vessel 20 are divided) output from the trained model as a target variable.

[0063] The first concentration calculation unit 11f inputs the first parameters acquired by the parameter acquisition unit 11e into the biochemical reaction model expressed by the above formulas (1) to (5) based on the first parameters acquired by the parameter acquisition unit 11e and the biochemical reaction model expressed by the above formulas (1) to (5), and calculates the cell concentration X, substrate concentration S, product concentration P, and inhibitory product concentration P that satisfy the above formulas (1) to (5). ｉBy performing calculations to obtain each of the above, first related concentrations (cell concentration, substrate concentration, product concentration, inhibitory product concentration, dissolved oxygen concentration, dissolved carbon dioxide concentration, and dissolved hydrogen ion concentration in one or more regions among the multiple regions into which the contents 21 contained in the culture vessel 20 are divided) related to the first culture condition input by the culture condition input unit 11 d are calculated.

[0064] The first concentration calculation unit 11f displays the calculated first related concentration on the display unit 15. By referring to the first related concentration displayed on the display unit 15, the user can grasp the cell concentration, substrate concentration, product concentration, inhibitory product concentration, dissolved oxygen concentration, dissolved carbon dioxide concentration, and dissolved hydrogen ion concentration in one or more regions among the multiple regions into which the contents 21 contained in the culture vessel 20 is divided, which are related to the first culture condition, during actual operation of the culture vessel 20.

[0065] Other indices related to the culture may be calculated based on the related concentrations calculated by the first concentration calculation unit 11f. Examples of other indices include oxygen partial pressure, carbon dioxide partial pressure, pH, oxidation-reduction potential, osmotic pressure, cell weight, and cell viability. For example, the oxygen partial pressure, carbon dioxide partial pressure, and pH are calculated based on the dissolved oxygen concentration, dissolved carbon dioxide concentration, and dissolved hydrogen ion concentration calculated by the first concentration calculation unit 11f, respectively. The oxidation-reduction potential is calculated based on the substrate concentration and product concentration calculated by the first concentration calculation unit 11f. The cell viability is calculated based on the cell concentration calculated by the first concentration calculation unit 11f.

[0066] The first setting unit 11g sets, in a calculation model (e.g., a CFD (Computational Fluid Dynamics) model), an advection-diffusion equation that is related to the first related concentration (e.g., cell concentration) calculated by the first concentration calculation unit 11f and that takes into account, for example, advection and diffusion of cells between a first region (e.g., region 21A shown in FIG. 4 ) and a second region adjacent to the first region (e.g., regions 21B, 21C, and 21D shown in FIG. 4 ) in the culture vessel 20. When setting the advection-diffusion equation in the calculation model, the first setting unit 11g sets an initial value of the first related concentration (e.g., cell concentration) in the calculation model.

[0067] As shown in Figure 4, in region 21A, not only does cell proliferation occur within region 21A, but also cell advection and diffusion between adjacent regions 21B, 21C, and 21D cause an increase or decrease in cells within region 21A. An advection-diffusion equation that considers the advection and diffusion of cells between adjacent regions and quantifies the increase or decrease in cells is expressed, for example, by the following equation (7). In equation (7), the first term on the left side represents the change in cell concentration due to cell proliferation, and the second term on the left side represents the advection term (c ２１Ａ : advection velocity), and the right-hand side is the diffusion term (D ２１Ａ : diffusion coefficient). However, the advection-diffusion equation is not limited to equation (7), and other models representing advection and diffusion may be used. Furthermore, an equation representing only advection or an equation representing only diffusion may be used as the advection-diffusion equation.

[0068]

[0069] Furthermore, the first setting unit 11g sets, in the calculation model, an advection-diffusion equation that is related to the first related concentration (e.g., substrate concentration) calculated by the first concentration calculation unit 11f and that takes into account the advection and diffusion of, for example, a substrate between a first region (e.g., region 21A shown in FIG. 4 ) and a second region adjacent to the first region (e.g., regions 21B, 21C, and 21D shown in FIG. 4 ) in the culture vessel 20. When setting the advection-diffusion equation in the calculation model, the first setting unit 11g sets, in the calculation model, an initial value of the first related concentration (e.g., substrate concentration).

[0070] Furthermore, the first setting unit 11g sets, in the calculation model, an advection-diffusion equation that is related to the first related concentration (e.g., product concentration) calculated by the first concentration calculation unit 11f and that takes into account, for example, advection and diffusion of the product between a first region (e.g., region 21A shown in FIG. 4 ) and a second region adjacent to the first region (e.g., regions 21B, 21C, and 21D shown in FIG. 4 ) in the culture vessel 20. When setting the advection-diffusion equation in the calculation model, the first setting unit 11g sets, in the calculation model, an initial value of the first related concentration (e.g., product concentration).

[0071] When the first culture conditions are input to the trained model, the second setting unit 11h uses the first parameters acquired by the parameter acquisition unit 11e, and sets a concentration change equation (biochemical reaction equation) indicating a change over time of the first related concentration in the first region (e.g., region 21A shown in FIG. 4 ) in the calculation model. The concentration change equation indicating a change over time of the cell concentration as the first related concentration is expressed, for example, by the following equation (8). Furthermore, the concentration change equation indicating a change over time of the substrate concentration as the first related concentration is expressed, for example, by the following equation (9). Furthermore, the concentration change equation indicating a change over time of the product concentration as the first related concentration is expressed, for example, by the following equation (10). However, as with the biochemical reaction model described above, the concentration change equations are not limited to these equations. In addition, when the parameter acquisition unit 11e acquires the first parameters at multiple times (for example, time t1, time t2), the second setting unit 11h sets the calculation model in the CFD calculation (simulation) performed by the second concentration calculation unit 11i so that the first parameters change at the multiple times to the first parameters acquired corresponding to each of the times.

[0072]

[0073]

[0074]

[0075] When there is only one trained model corresponding to one of the multiple regions obtained by dividing the contents 21 contained in the culture vessel 20, the second setting unit 11h may use the first parameter acquired by the parameter acquisition unit 11e from the trained model for the multiple regions (all regions) in the concentration change equation. When there are two trained models corresponding to two of the multiple regions obtained by dividing the contents 21 contained in the culture vessel 20, the second setting unit 11h may appropriately use the two first parameters acquired by the parameter acquisition unit 11e from the trained model for the region close to the two regions in the concentration change equation, while for the region not close to the two regions, the second setting unit 11h may weight the region based on the distance from the two regions and use a parameter calculated from the two first parameters as the first parameter in the concentration change equation.

[0076] The second concentration calculation unit 11i performs CFD calculations based on the calculation model set by the first setting unit 11g and the second setting unit 11h, and calculates second related concentrations (e.g., related concentrations at each time from the start of culture at time t0 to the end of culture at time tf) related to the first culture condition in a first region (e.g., region 21A shown in FIG. 4 ). That is, the second concentration calculation unit 11i calculates the changes over time in the cell concentration, substrate concentration, and product concentration (concentration distribution within the culture vessel 20) in each of the multiple regions obtained by dividing the contents 21 contained in the culture vessel 20 by solving, through calculation, the related concentrations (cell concentration, substrate concentration, and product concentration) that satisfy both the advection-diffusion equation set by the first setting unit 11g and the concentration change equation set by the second setting unit 11h.

[0077] The second concentration calculation unit 11i displays the calculated second related concentrations on the display unit 15. By referring to the second related concentrations displayed on the display unit 15, the user can grasp the distribution of each related concentration (cell concentration, substrate concentration, and product concentration) in the culture vessel 20 as a result of CFD calculation at a desired time during actual operation of the culture vessel 20.

[0078] The actual measurement value acquisition unit 11a acquires actual measurement values ​​of relevant concentrations (e.g., cell concentration, substrate concentration, and product concentration) related to the first culture condition when culturing cells in the culture vessel 20 for a first region (e.g., region 21A shown in Figure 4) among multiple regions into which the contents 21 contained in the culture vessel 20 are divided, by measuring the relevant concentrations in the culture vessel 20 during actual operation.

[0079] The parameter calculation unit 11b calculates, by optimization calculation, parameters of a biochemical reaction model used to calculate the relevant concentration related to the culture conditions so that the difference between the actual measurement value acquired by the actual measurement value acquisition unit 11a and the second relevant concentration calculated by the second concentration calculation unit 11i is small.

[0080] The parameter calculation unit 11b stores in the memory unit 12 as learning data data, data indicating a pair of the parameters of the biochemical reaction model calculated by the optimization calculation and the first culture conditions (culture conditions in the culture vessel 20 when the actual measured value was measured) corresponding to the actual measured value of the second related concentration used to calculate the parameters.

[0081] The learning unit 11c performs machine learning (relearning) of the trained model stored in the memory unit 12 using training data (teacher data) stored in the memory unit 12 by the parameter calculation unit 11b. That is, the learning unit 11c inputs the training data into the trained model, and generates a new trained model by having the trained model perform machine learning (so-called supervised learning) to determine the correlation between the input data (first culture conditions) and output data (parameters of the biochemical reaction model) that constitute the training data. This makes it possible to obtain more optimal parameters for the biochemical reaction model output from the trained model, and ultimately to more accurately calculate concentrations related to culture using the biochemical reaction model. The learning unit 11c stores the newly generated trained model in the memory unit 12.

[0082] If the actual measurement values ​​acquired by the actual measurement value acquisition unit 11a correspond to only one of the multiple regions obtained by dividing the content 21, the parameters of the biochemical reaction model calculated by the parameter calculation unit 11b using the actual measurement values ​​may be applied to a region other than the one region, and machine learning (relearning) of the trained model corresponding to the different region may be performed. If the multiple actual measurement values ​​acquired by the actual measurement value acquisition unit 11a correspond to all of the regions obtained by dividing the content 21, the parameters of the biochemical reaction model calculated by the parameter calculation unit 11b using the actual measurement values ​​may be applied to each of the regions, and machine learning (relearning) of the trained model corresponding to all of the regions may be performed. In this case, the parameter calculation unit 11b may calculate parameters common to all regions by optimization calculation so that the difference between the actual measurement values ​​and the second related concentration is uniformly small for all regions, or so that weights are set for each region so that the difference between the actual measurement values ​​and the second related concentration is small for each region, or so that parameters corresponding to each of the regions may be calculated by optimization calculation.

[0083] FIG. 5 is a flowchart showing an example of a learning data preparation process performed by the calculation device 10 in this embodiment.

[0084] First, the actual measurement value acquisition unit 11a acquires actual measurement values ​​of relevant concentrations (e.g., cell concentration, substrate concentration, product concentration, by-product concentration, inhibitory product concentration, etc.) related to the culture conditions when culturing cells in the culture vessel 20 (step S100).

[0085] Next, the parameter calculation unit 11b calculates the parameters of the biochemical reaction model used to calculate the relevant concentrations related to the culture conditions by optimization calculation so that the actual measured values ​​of the relevant concentrations obtained in step S100 are calculated (step S110).

[0086] Finally, the parameter calculation unit 11b stores, as training data, data indicating a pair of the parameters of the biochemical reaction model calculated in step S110 and the culture conditions corresponding to the actual measured values ​​of the relevant concentrations used to calculate the parameters by optimization calculation (the culture conditions in the culture vessel 20 when the actual measured values ​​were measured) in the storage unit 12 (step S120). Upon completion of the process of step S120, the calculation device 10 ends the training data preparation process shown in FIG.

[0087] FIG. 6 is a flowchart showing an example of a learning process performed by the calculation device 10 in this embodiment, following the learning data preparation process shown in FIG.

[0088] First, when the culture conditions when culturing cells in the culture vessel 20 are input as explanatory variables, the learning unit 11c performs machine learning of a learning model that outputs, as objective variables, parameters of a biochemical reaction model used to calculate relevant concentrations related to the culture conditions, using the learning data (teacher data) stored in the memory unit 12 in step S120 of Figure 5 (step S200).

[0089] Next, the learning unit 11c stores the trained model that has been subjected to machine learning in step S200 in the storage unit 12 (step S210). When the process of step S210 is completed, the computing device 10 ends the learning process shown in FIG. 6.

[0090] FIG. 7 is a flowchart showing an example of a first concentration calculation process (corresponding to the "calculation method" of the present disclosure) performed by the calculation device 10 in this embodiment, following the learning process shown in FIG.

[0091] First, the culture condition input unit 11d inputs the first culture conditions (new culture conditions) as explanatory variables to the trained model stored in the storage unit 12 in step S210 of FIG. 6 (step S300).

[0092] Next, the parameter acquiring unit 11e acquires the first parameter output as the objective variable from the trained model to which the first culture condition was input in step S300 (step S310).

[0093] Finally, the first concentration calculation unit 11f calculates a first related concentration related to the first culture condition input in step S300 based on the first parameter acquired in step S310 and the biochemical reaction model (step S320). Upon completion of the process in step S320, the calculation device 10 ends the first concentration calculation process shown in FIG.

[0094] FIG. 8 is a flowchart showing an example of a second concentration calculation process performed by the calculation device 10 following the first concentration calculation process shown in FIG. 7 in this embodiment.

[0095] First, the first setting unit 11g sets, in the calculation model, an advection-diffusion equation that is related to the first related concentration calculated by the first concentration calculation unit 11f and that takes into account advection and diffusion between a first region (e.g., region 21A shown in Figure 4) in the culture vessel 20 and a second region adjacent to the first region (e.g., regions 21B, 21C, and 21D shown in Figure 4) (step S400).

[0096] Next, the second setting unit 11h uses the first parameters acquired in step S310 of Figure 7 to set a concentration change equation showing the change over time of the first related concentration in a first region (e.g., region 21A shown in Figure 4) within the culture vessel 20 in the calculation model (step S410).

[0097] Finally, the second concentration calculation unit 11i performs CFD calculations based on the calculation model set in steps S400 and S410 to calculate a second related concentration (e.g., related concentration at each time from the start of culture at time t0 to the end of culture at time tf) related to the first culture condition in a first region (e.g., region 21A shown in FIG. 4 ) in the culture vessel 20 (step S420). Upon completion of the process of step S420, the calculation device 10 ends the second concentration calculation process shown in FIG. 8.

[0098] FIG. 9 is a flowchart showing an example of a relearning process performed by the calculation device 10 following the second concentration calculation process shown in FIG. 8 in this embodiment.

[0099] First, the actual measurement value acquisition unit 11a acquires the actual measurement value of the related concentration related to the first culture condition when culturing cells in the culture vessel 20 for a first region (e.g., region 21A shown in Figure 4) among multiple regions into which the contents 21 contained in the culture vessel 20 are divided, by measuring the related concentration in the culture vessel 20 during actual operation (step S500).

[0100] Next, the parameter calculation unit 11b calculates parameters of the biochemical reaction model used to calculate the related concentration related to the culture conditions by optimization calculation so that the difference between the actual measurement value obtained in step S500 and the second related concentration calculated in step S420 of Figure 8 is small (step S510).

[0101] Next, the parameter calculation unit 11b stores data indicating a pair of the parameters of the biochemical reaction model calculated in step S510 and the first culture conditions corresponding to the actual measured values ​​of the second related concentrations used to calculate the parameters in the memory unit 12 as learning data (step S520).

[0102] The learning unit 11c performs machine learning (relearning) of the learned model stored in the memory unit 12 in step S210 of Figure 6 using the learning data (teacher data) stored in the memory unit 12 in step S520 (step S530).

[0103] Finally, the learning unit 11c stores the trained model that has been subjected to machine learning in step S530 in the storage unit 12 (step S540). When the process of step S540 is completed, the computing device 10 ends the re-learning process shown in FIG.

[0104] As described in detail above, in this embodiment, the trained model is trained using the culture conditions used when culturing cells in the culture vessel 20 as explanatory variables and the parameters of a biochemical reaction model (physical model) used to calculate a relevant concentration related to the culture conditions as objective variables. The calculation device 10 includes a culture condition input unit 11d, a parameter acquisition unit 11e, and a first concentration calculation unit 11f. The culture condition input unit 11d inputs the first culture conditions as explanatory variables to the trained model. When the first culture conditions are input to the trained model, the parameter acquisition unit 11e acquires the first parameters output from the trained model as objective variables. The first concentration calculation unit 11f calculates the first relevant concentration related to the first culture condition based on the acquired first parameters and the biochemical reaction model.

[0105] According to this embodiment configured as described above, the first culture conditions (explanatory variables) can be input into the trained model to obtain the first parameters of the biochemical reaction model output as the objective variables. That is, the first parameters optimized under the first culture conditions can be obtained. Therefore, the first related concentration associated with the first culture conditions can be accurately calculated using the first parameters and the biochemical reaction model. Furthermore, there is no need to obtain actual measured values ​​of the related concentrations under the changed culture conditions and optimize the parameters used in the biochemical reaction model each time the culture conditions are changed, thereby preventing an increase in the cost required for accurately calculating the related concentrations.

[0106] In the present embodiment, the calculation device 10 further includes a first setting unit 11g, a second setting unit 11h, and a second concentration calculation unit 11i. The first setting unit 11g sets, in the calculation model, an advection-diffusion equation related to the first related concentration and taking into account advection and diffusion between a first region and a second region adjacent to the first region in the culture vessel 20. The second setting unit 11h uses first parameters acquired by inputting the first culture condition into the learned model to set, in the calculation model, a concentration change equation indicating a time-dependent change in the first related concentration in the first region. The second concentration calculation unit 11i calculates, in the first region, a second related concentration related to the first culture condition, based on the set calculation model. This embodiment configured in this manner makes it possible to accurately calculate, in the first region, a second related concentration related to the first culture condition and taking into account advection and diffusion between the first region and the adjacent second region.

[0107] In this embodiment, the calculation device 10 further includes an actual measurement value acquisition unit 11a, a parameter calculation unit 11b, and a learning unit 11c. The actual measurement value acquisition unit 11a acquires actual measurements of relevant concentrations associated with a first culture condition in a first region. The parameter calculation unit 11b calculates parameters of a biochemical reaction model (physical model) so as to reduce the difference between the acquired actual measurements and a second relevant concentration. The learning unit 11c uses the calculated parameters and the first culture condition as training data to train the trained model. According to this embodiment configured as described above, the parameters of the biochemical reaction model are calculated so as to reduce the difference between the actual measurements of the relevant concentrations under the first culture condition and the second relevant concentration calculated in consideration of advection and diffusion between adjacent regions, and the trained model is retrained using the parameters as training data. Therefore, when the first culture conditions are input into the learned model, a first parameter that is more optimized for the first culture conditions can be obtained from the biochemical reaction model to the extent that re-learning based on the second related concentration has been performed, and the first related concentration associated with the first culture conditions can be calculated more accurately using the first parameter and the biochemical reaction model.

[0108] In the above embodiment, a trained model (hereinafter referred to as a "first trained model") in which culture conditions are used as explanatory variables and parameters of a biochemical reaction model (physical model) are used as objective variables, and the calculation device 10 that calculates related concentrations using parameters output from this first trained model in the physical model have been described. The calculation device 10 of the present disclosure can also be configured to include a "second trained model" in addition to this.

[0109] This second trained model is a model trained using second culture conditions related to the first culture conditions as explanatory variables and the related concentration itself as the objective variable. Here, the "relationship" between the first culture conditions and the second culture conditions includes various aspects, such as a partially shared relationship where the first culture conditions are various conditions related to the main culture process and the second culture conditions are various conditions related to the pre-culture process and some conditions related to the main culture process, or an inclusive relationship where the first culture conditions are only various conditions related to the main culture process and the second culture conditions are only various conditions related to the pre-culture process, or an exclusive relationship where the first culture conditions are only various conditions related to the main culture process. Specifically, the second trained model can be a model trained by machine learning using training data where the culture conditions in the pre-culture process (e.g., medium composition, operating conditions, device shape, etc.) are explanatory variables and the related concentration in the subsequent main culture process is the objective variable. In this case, the learning data of the second trained model may further include some of the culture conditions in the main culture process (e.g., initial culture medium composition, initial operating conditions, device shape, etc.).

[0110] The step of constructing the second trained model first includes a training data preparation step of creating a training data set that directly links the culture conditions in the pre-culture step as described above with the corresponding measured values ​​or highly accurate estimates of the relevant concentrations in the main culture step. Next, the training unit 11c uses the prepared training data as training data to perform machine learning of a training model in which the culture conditions in the pre-culture step are explanatory variables and the relevant concentrations are objective variables, thereby generating a second trained model and storing it in the memory unit 12. The training data preparation step and the training step can be performed independently of, in parallel with, or sequentially with the process of constructing the first trained model (e.g., the process described in FIGS. 5 and 6 ).

[0111] In such a configuration including a second trained model, the first trained model uses first culture conditions as explanatory variables, including the culture conditions of the main culture process (e.g., the shape of the device, the physical properties of the contents, and the operating conditions (initial operating conditions and operating conditions changed after operation has started)). The relationship between the first culture conditions used as explanatory variables by the first trained model and the second culture conditions used as explanatory variables by the second trained model can take various forms, such as a partially common relationship (pattern 1), a relationship in which one is included in the other (pattern 2), or a mutually exclusive relationship (pattern 3), as described above.

[0112] In this way, by adding a second trained model to a configuration that combines a first trained model and a physical model (a type of "hybrid model"), it is possible to quickly build an entire model for estimating relevant concentrations. This is because building a highly accurate hybrid model (a combination of a first trained model and a physical model) that comprehensively considers all culture conditions requires a huge amount of training data. However, in practice, there may be a need to estimate relevant concentrations early, even in the early stages of development when sufficient training data has not yet been collected.

[0113] In such a situation, one embodiment of the model construction method of the present disclosure may include the following stepwise steps. First, in a first construction step, a second trained model that directly predicts a related concentration from a relatively small number of culture conditions is constructed in advance. Then, in a first estimation step, corresponding culture conditions (second culture conditions) are input into the constructed second trained model, and a related concentration (third related concentration) is obtained as an output and utilized. Specifically, a second culture condition input unit 11j (see FIG. 10) inputs the second culture conditions into the second trained model, and a third related concentration acquisition unit 11k (see FIG. 10) acquires the third related concentration from the second trained model. This third related concentration can be utilized as an early concentration estimation result. Next, in a second construction step, a first trained model is constructed that uses a larger number of culture conditions as explanatory variables and parameters of a biochemical reaction model as target variables, and this is combined with a physical model to form a hybrid model. Then, in the second estimation step, this hybrid model is used to obtain parameters (first parameters) of a physical model based on input culture conditions (first culture conditions), and a relevant concentration (first relevant concentration, or a second relevant concentration further using CFD or the like) is calculated from the first parameters and the physical model (e.g., the process shown in Figures 7 and 8). By adopting this stepwise approach of model construction and calculation, it is possible to meet early needs for concentration estimation in the first estimation step, while transitioning to or complementing more accurate or detailed relevant concentration estimation in the second estimation step as data collection and model development progress.

[0114] In another aspect of the model construction method of the present disclosure, or as an update step following the above aspect, the constructed second trained model can be retrained using input / output data of a subsequently constructed or improved hybrid model (a combination of the first trained model and the physical model), i.e., a set of culture conditions input to the hybrid model and the relevant concentrations estimated thereby, as new training data (updating step of the second trained model). In this update step, specific culture conditions are first input to the hybrid model, and corresponding estimates of the relevant concentrations are obtained. Next, a pair of the input culture conditions and the relevant concentration estimates obtained by the hybrid model is prepared as new training data. The learning unit 11c then performs retraining (machine learning) of the second trained model using this new training data. This allows new knowledge obtained by the hybrid model to be fed back to the second trained model, improving its prediction accuracy and expanding its scope of application. This retraining process can be realized, for example, within the framework of the retraining process shown in FIG. 9 , by using estimates by the hybrid model as part of the training data instead of actual measurements and executing the second trained model as the target.

[0115] This construction method, combining a second trained model with a hybrid model, is particularly effective when the culture process undergoes multiple scale-ups. Specifically, the advantages of this method are utilized when the culture conditions and culture results of one or more pre-culture steps preceding the final main culture step have a significant impact on estimating the relevant concentration in the final main culture step. Typically, building a single, advanced model using comprehensive training data covering all culture conditions and culture results across multiple culture steps at different scales not only requires a massive amount of experimental data, but also significantly complicates the parameter adjustment process of the trained model, resulting in a long time required to complete the model. Therefore, a second trained model is quickly constructed using the culture conditions of the pre-culture step (and, if necessary, some initial conditions of the main culture step) as explanatory variables and the relevant concentration in the main culture step as the objective variable, and is used for initial predictions (corresponding to the first construction step and first estimation step described above). In parallel, a more detailed hybrid model is constructed (corresponding to the second construction step described above), enabling a model for estimating the relevant concentration in the main culture step to be constructed quickly and efficiently.

[0116] Note that the second trained model does not necessarily need to be retrained using the input / output data of the hybrid model. For example, if sufficient prediction accuracy for practical use is achieved by using only the second trained model, which uses the culture conditions of the pre-culture step as explanatory variables and the relevant concentrations of the main culture step as objective variables, it may be possible to omit retraining using the data of the hybrid model (the above-mentioned step of updating the second trained model).

[0117] As described above, the second trained model can be designed so that the second culture conditions used as explanatory variables further include some data related to the main culture process (e.g., initial conditions, etc.) For example, initial conditions such as the initial operating conditions and initial medium composition of the main culture process can be included in the explanatory variables of the second trained model.

[0118] Furthermore, the above-described embodiments are merely examples of specific embodiments for carrying out the present disclosure, and the technical scope of the present disclosure should not be interpreted as being limited by these embodiments. In other words, the present disclosure can be carried out in various forms without departing from the gist or main features thereof.

[0119] 10: Calculation device, 11: Control unit, 11a: Actual measurement value acquisition unit, 11b: Parameter calculation unit, 11c: Learning unit, 11d: Culture condition input unit, 11e: Parameter acquisition unit, 11f: First concentration calculation unit, 11g: First setting unit, 11h: Second setting unit, 11i: Second concentration calculation unit, 11j: Second culture condition input unit, 11k: Third related concentration acquisition unit, 12: Memory unit, 13: Communication unit, 14: Operation input unit, 15: Display unit, 16: System bus, 20: Culture vessel, 21: Culture solution, 21A, 21B, 21C, 21D: Region, 22: Gas, 23: Aeration sparger, 24: Stirring blade

Claims

1. A trained model that has been trained using the culture conditions when culturing cells in a culture vessel as explanatory variables and the parameters of a physical model used to calculate relevant concentrations related to the culture conditions as objective variables.

2. A calculation device comprising a culture condition input unit, a parameter acquisition unit, and a first concentration calculation unit, wherein the culture condition input unit inputs first culture conditions as the explanatory variables to the trained model described in claim 1, the parameter acquisition unit acquires a first parameter output from the trained model as the objective variable when the first culture conditions are input to the trained model, and the first concentration calculation unit calculates a first related concentration related to the first culture condition based on the acquired first parameter and the physical model.

3. A calculation device according to claim 2, further comprising a first setting unit, a second setting unit, and a second concentration calculation unit, wherein the first setting unit sets, in a calculation model, an equation related to the first related concentration and taking into account advection or diffusion between a first region in the culture vessel and a second region adjacent to the first region, the second setting unit sets, in the calculation model, a concentration change equation that uses the first parameter and shows the change over time of the first related concentration in the first region, and the second concentration calculation unit calculates, based on the set calculation model, a second related concentration in the first region that is related to the first culture condition.

4. A calculation device according to claim 3, further comprising an actual measurement value acquisition unit, a parameter calculation unit, and a learning unit, wherein the actual measurement value acquisition unit acquires an actual measurement value of a relevant concentration associated with the first culture condition in the first region, the parameter calculation unit calculates parameters of the physical model so as to reduce the difference between the acquired actual measurement value and the second relevant concentration, and the learning unit uses the calculated parameters and the first culture condition as learning data to train the trained model.

5. A calculation device according to claim 2, further comprising a second trained model, a second culture condition input unit, and a third related concentration acquisition unit, wherein the second trained model is trained using a second culture condition related to the first culture condition as an explanatory variable and the related concentration as a target variable, the second culture condition input unit inputs the second culture condition as the explanatory variable to the second trained model, and the third related concentration acquisition unit acquires the third related concentration output from the second trained model as the related concentration.

6. A calculation method having a culture condition input step, a parameter acquisition step, and a first concentration calculation step, wherein in the culture condition input step, a first culture condition is input as the explanatory variable to the trained model described in claim 1; in the parameter acquisition step, when the first culture condition is input to the trained model, a first parameter output from the trained model as the objective variable is acquired; and in the first concentration calculation step, a first related concentration related to the first culture condition is calculated based on the acquired first parameter and the physical model.

7. A calculation program that causes a processor to execute the calculation method according to claim 6.

8. A computer-readable non-transitory storage medium storing a calculation program for causing a processor to execute the calculation method according to claim 6.

Images