Estimation method, estimation program, estimation device, and additive for hydraulic composition

Abstract

To estimate additives for hydraulic compositions for exhibiting suitable performance.SOLUTION: An estimation method includes a model generation step for generating a trained model for estimating an objective variable from an explanatory variable by using, as teacher data, a plurality of sets of explanatory variables showing a producing condition of additives for hydraulic compositions and objective variables showing characteristics of the additives for hydraulic compositions, and an estimation step for outputting an estimation value of the objective variable showing the characteristics of the additives for hydraulic compositions from the explanatory variable showing the producing condition of additives for hydraulic compositions for which the characteristics are to be estimated by using the trained model.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an estimation method, an estimation program, and an estimation device for estimating the performance of an additive for a hydraulic composition, and an additive for a hydraulic composition. [Background technology]

[0002] Hydraulic compositions such as concrete are generally prepared using aggregates such as sand and gravel, binders such as cement, and water as main materials, and additives are further mixed in. Such additives are used for the purpose of improving various properties of the hydraulic composition, such as fluidity, air content, strength, and viscosity. Additives for hydraulic compositions are exemplified in Patent Documents 1 to 3, for example. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2022-45007 [Patent Document 2] International Publication No. 2021 / 024853 [Patent Document 3] JP 2020-15656 A Summary of the Invention [Problem to be solved by the invention]

[0004] The development of additives for hydraulic compositions had to rely on classical experimental scientific techniques. In other words, it was necessary to repeatedly manufacture and evaluate candidate compositions to narrow down the compositions that would exhibit suitable performance. However, since additives for hydraulic compositions contain many components, the chemical species and composition ratios of each component are diverse, and it was necessary to find a suitable composition from an infinite number of candidates. Therefore, the development of a new additive for hydraulic compositions required a huge amount of work, time, and cost.

[0005] Therefore, there is a need for a method, a program, and an apparatus for estimating additives for hydraulic compositions that exhibit favorable performance. In addition, there is a need for new additives for hydraulic compositions manufactured using the method. [Means for solving the problem]

[0006] The estimation method according to the present invention includes a model generation step of using a plurality of pairs of explanatory variables indicating manufacturing conditions of the hydraulic composition additive and objective variables indicating characteristics of the hydraulic composition additive as training data to generate a trained model that estimates the objective variables from the explanatory variables; a generation step of generating a plurality of explanatory variables indicating manufacturing conditions of the hydraulic composition additive; an estimation step of using the trained model to output a plurality of objective variables estimated from the plurality of explanatory variables generated in the generation step; a step of specifying a predetermined criterion that the objective variables should satisfy; and a step of identifying an objective variable that satisfies the predetermined criterion from among the plurality of objective variables output in the estimation step, and outputting the objective variables generated in the generation step. and an output step of outputting, as a solution, explanatory variables used when outputting a target variable that satisfies the predetermined criterion, among the multiple explanatory variables generated, wherein the generating step and the estimating step are executed multiple times, and in the second or subsequent generating step, multiple explanatory variables are generated using a genetic algorithm that utilizes results of the generating step and the estimating step executed prior to the generating step, and the target variable includes at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in the hydraulic composition additive.

[0007] The estimation program according to the present invention, when executed by a computer, provides the computer with the following functions: a model generation function for generating a trained model that estimates a dependent variable from an explanatory variable by using, as training data, a plurality of pairs of explanatory variables indicating manufacturing conditions of the hydraulic composition additive and a dependent variable indicating characteristics of the hydraulic composition additive; a generation function for generating a plurality of explanatory variables indicating manufacturing conditions of the hydraulic composition additive; an estimation function for outputting, by using the trained model, a plurality of dependent variables estimated from the plurality of explanatory variables generated by the generation function; a function for receiving designation of a predetermined criterion that the dependent variable should satisfy; and an output function that outputs, as a solution, explanatory variables used when outputting a target variable that satisfies the predetermined criterion, from among the multiple explanatory variables generated by the generation function, wherein the generation function and the estimation function are realized multiple times, and in the generation function realized from the second time onwards, multiple explanatory variables are generated using a genetic algorithm that utilizes the results of the generation function and the estimation function realized before the generation function, and the target variable includes at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in an additive for hydraulic composition.

[0008] The estimation device according to the present invention is an estimation device including a calculation device and a storage device, the storage device storing an estimation program, and when the estimation program is executed by the calculation device, the calculation device is provided with a model generation function for generating a trained model for estimating the objective variables from the explanatory variables by using a plurality of pairs of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition and objective variables indicating characteristics of the additive for hydraulic composition as training data, a generation function for generating a plurality of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition, an estimation function for outputting a plurality of objective variables estimated from the plurality of explanatory variables generated by the generation function using the trained model, a function for accepting designation of a predetermined standard that the objective variables should satisfy, and a function for selecting an objective variable from among the plurality of objective variables output by the estimation function. and an output function of identifying a response variable that satisfies the predetermined criterion, and outputting, as a solution, an explanatory variable used when outputting the response variable that satisfies the predetermined criterion, among a plurality of explanatory variables generated by the generation function, wherein the generation function and the estimation function are realized a plurality of times, and in the generation function realized from the second time onwards, a plurality of explanatory variables are generated using a genetic algorithm that utilizes results of the generation function and the estimation function realized before the generation function, and the response variable includes at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in an additive for hydraulic composition.

[0009] According to these configurations, the manufacturing conditions (explanatory variables) that can obtain the additive for hydraulic composition exhibiting the desired performance are estimated using a learning model, so that suitable manufacturing conditions can be specified without manufacturing the additive for hydraulic composition and measuring its physical properties. This can shorten the trial and error required to determine suitable manufacturing conditions for the additive for hydraulic composition. In addition, it is expected that the obtained solution will be narrowed down to a more suitable range by using a genetic algorithm.

[0010] In one aspect, the estimation method according to the present invention preferably further includes a model regeneration step of adding a pair of explanatory variables output as a solution in the output step and a target variable indicating the performance of the additive for hydraulic composition manufactured under the manufacturing conditions indicated by the explanatory variables to training data, and regenerating the trained model.

[0011] With this configuration, the accuracy of the trained model can be improved by outputting a solution, verifying the solution, and feeding back the verification results to the trained model.

[0012] In one aspect of the estimation method according to the present invention, the explanatory variables preferably include chemical species information determined by chemical species of compounds contained in the hydraulic composition additive, and ratio information specifying the content ratio of each of the compounds.

[0013] According to this configuration, estimation is performed taking into consideration chemical species information and ratio information, which often have a particularly large effect on the performance of the hydraulic composition additive, making it easier to obtain highly accurate estimation results.

[0014] In one aspect of the estimation method according to the present invention, the chemical species information preferably includes a molecular descriptor of the compound.

[0015] According to this configuration, chemical species that can provide suitable objective variables can be systematically understood.

[0016] In one embodiment of the estimation method according to the present invention, the objective variable preferably includes at least one physical property value selected from the group consisting of fluidity, fluidity retention, air entrainment, air amount retention, viscosity, pumpability, material separation resistance, finishability, adhesion, setting characteristics, strength characteristics, shrinkage characteristics, frost resistance, carbonation characteristics, chloride penetration resistance, steel protection, alkali-silica reaction resistance, crack resistance, watertightness, fire resistance, abrasion resistance, and surface aesthetics of the hydraulic composition to which the hydraulic composition additive has been added.

[0017] According to this configuration, a suitable additive for a hydraulic composition can be derived by using physical properties that are particularly noteworthy as functions of the additive for a hydraulic composition as evaluation criteria.

[0018] Further features and advantages of the present invention will become more apparent from the following description of exemplary and non-limiting embodiments, which are given with reference to the drawings. [Brief description of the drawings]

[0019] [Figure 1] 1 is a graph showing the verification results of the trained model in Example 1. [Diagram 2] 13 is a graph showing the verification results of the trained model in Example 2. [Diagram 3] 13 is a graph showing the verification results of the trained model in Example 3. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] BEST MODE FOR CARRYING OUT THE PRESENTLY PREFERRED EMBODIMENTS An embodiment of an estimation method, an estimation program, an estimation device, and an additive for a hydraulic composition according to the present invention will be described.

[0021] [Explanatory variables and objective variables related to additives for hydraulic compositions] In the estimation method according to the present embodiment, explanatory variables indicating the manufacturing conditions of the additive for hydraulic composition and response variables indicating the performance of the additive for hydraulic composition are used. Prior to describing the embodiment of the estimation method, the explanatory variables and response variables will be described.

[0022] (Explanatory variables) The hydraulic composition additive is a composition used for the purpose of imparting some function to the hydraulic composition, and includes components (compounds) such as surfactants (anionic, cationic, amphoteric, nonionic), air entraining agents (AE agents), defoamers, shrinkage reducing agents, retarders, hardening accelerators, thickeners, preservatives, and rust inhibitors. Examples of surfactants include, but are not limited to, polycarboxylic acid ether compounds, lignin sulfonates, naphthalene sulfonic acid formaldehyde condensates, melamine sulfonic acid formaldehyde condensates, and polyalkoxy ether compounds, or mixtures thereof. The chemical species, combinations, content ratios, number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity (molecular weight distribution), melting point, glass transition point, HLB, SP value, water solubility, foamability, wettability, dispersibility, emulsification, particle size, particle size distribution, and particle shape of each component are explanatory variables that can affect the performance (objective variable) of the hydraulic composition additive.

[0023] In addition, the process conditions for the process of producing the hydraulic composition additive may also affect the performance (objective variable) of the hydraulic composition additive. Therefore, for example, the explanatory variables may be various conditions in the reaction that produces each component, such as the reaction method, reaction temperature, reaction time, reaction pressure, volume and shape of the reaction vessel, stirring speed, the presence or absence of a catalyst, the type and concentration of the catalyst, the reaction atmosphere, and the order in which the raw materials are added, as well as various conditions when mixing each component, such as the temperature, pressure, volume and shape of the vessel, stirring speed, the presence or absence of an auxiliary agent, the type and concentration of the auxiliary agent, and the order in which the raw materials are added.

[0024] In addition, the conditions related to the hydraulic composition to which the additive for hydraulic composition is applied can be used as explanatory variables. Examples of such conditions include, but are not limited to, the type, specific surface area, and chemical composition of cement, the type, density, particle size distribution, and surface water ratio of fine aggregate, the type, density, surface water ratio, and maximum size of coarse aggregate, the type, specific surface area, and chemical composition of admixture, the temperature and humidity used for the hydraulic composition, the type of device used for kneading the hydraulic composition, the kneading speed, kneading time, and kneading amount, the blending conditions of the hydraulic composition (design standard strength, target slump, water-cement ratio, fine aggregate ratio, unit amount of each material, amount of water-reducing component added, and the presence or absence, chemical species, and amount of special components added, etc.), the method, temperature, humidity, and time of curing. On the other hand, it is also possible to fix these conditions as prerequisites and consider only explanatory variables indicating the manufacturing conditions of the additive for hydraulic composition itself.

[0025] Since the explanatory variables are used to generate a trained model, they are preferably quantified. For example, when chemical species information determined by the chemical species of the components (compounds) contained in the hydraulic composition additive is treated as an explanatory variable, a character string indicating the name of the chemical species may be used as the chemical species information, but it is preferable to use a molecular descriptor as the chemical species information. The molecular descriptor can be obtained by expressing the molecular structure of the chemical species in a notation such as SMILES notation, SMARTS notation, InChI notation, or the like, and then using a known tool such as RDKit or Dragon.

[0026] (Objective variable) The hydraulic composition additive is used for the purpose of imparting some function to the hydraulic composition, and the variable that evaluates whether or not the purpose is achieved is the objective variable. In other words, the objective variable can be a physical property value that indicates the performance of the hydraulic composition additive. Examples of such physical property values ​​include the number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity (molecular weight distribution), melting point, glass transition point, HLB, SP value, water solubility, foamability, wettability, dispersibility, emulsification ability, particle size, particle size distribution, and particle shape of the surfactant contained in the hydraulic composition additive, and the fluidity, fluidity retention, air entrainment, air amount retention, viscosity, pumpability, material separation resistance, finishability, adhesion, setting characteristics, strength characteristics, shrinkage characteristics, frost resistance, neutralization characteristics, chloride penetration resistance, steel protection, alkali-silica reaction resistance, crack resistance, watertightness, fire resistance, abrasion resistance, and surface aesthetics of the hydraulic composition to which the hydraulic composition additive has been added, and any of these can be objective variables.

[0027] The method of identifying the objective variables including the items exemplified above is not particularly limited as long as the objective variables can be uniquely identified. For example, when the objective variables are items for which industrial standards such as JIS, ASTM, and ISO standards exist, or standards independently established between traders, the physical property values ​​identified according to these standards can be used as the objective variables. In addition, even if there are no industrial standards for an item, the item can be treated as the objective variable if it can be uniquely determined.

[0028] [Pre-trained model] There is a correlation between the explanatory variables and the objective variables described above. For example, it is known that the fluidity of a hydraulic composition (an example of an objective variable) is correlated with the chemical species and concentration of a surfactant contained in an additive for a hydraulic composition (an example of an explanatory variable). As in this example, the relationship between the explanatory variables and the objective variables is known or predictable theoretically or empirically. However, the additive for a hydraulic composition that is the subject of the present embodiment is often a mixture of many components, and the explanatory variables are diverse. Therefore, it is unrealistic or very difficult for a person to specify the correlation between the explanatory variables and the objective variables related to the additive for a hydraulic composition. Therefore, in this embodiment, a plurality of pairs of explanatory variables and objective variables are used as training data to generate a trained model that estimates the objective variable from the explanatory variables, and this model is utilized.

[0029] The algorithm used to generate a trained model from training data is not particularly limited. Examples include, but are not limited to, support vector machines (regression, classification), decision trees, random forests, gradient boosting, logistic regression, neural networks (simple perceptron, multi-layer perceptron), Gaussian process regression, Bayesian networks, k-nearest neighbors, Lasso regression, multiple regression analysis, ridge regression, elastic net, partial least squares regression, etc.

[0030] [First embodiment] The estimation method according to the first embodiment is a method for estimating the performance of an additive for a hydraulic composition, and includes a model generation step and an estimation step. Note that the estimation method according to the first embodiment is performed by using a computer.

[0031] The model generation process is a process of generating a trained model that estimates a response variable from an explanatory variable. In the first embodiment, a data group in which the correspondence between the explanatory variables and the response variable has been clarified by a method such as an experiment or a simulation is used as training data.

[0032] When obtaining training data through an experiment, first, a plurality of hydraulic composition additives with different manufacturing conditions, i.e., different explanatory variables, are prepared. At this time, the manufacturing conditions used as explanatory variables are recorded. Next, for the plurality of hydraulic composition additives prepared, the physical property values ​​that represent the performance required for the hydraulic composition additive, i.e., the objective variables, are measured and recorded. Through the above operations, training data, which are a plurality of pairs of explanatory variables and objective variables, is obtained. Note that, although some or all of the variables constituting the training data may be obtained by simulation, it should be noted that if the relationship between the explanatory variables and the objective variables can be clarified without using a trained model, there is little benefit in generating a trained model.

[0033] The amount of training data to be obtained can be determined appropriately depending on the combination of explanatory variables and objective variables used, the required accuracy of estimation, the algorithm used to generate the trained model, etc.

[0034] In the model generation step, it is preferable that the explanatory variables include chemical species information determined by the chemical species of the compounds contained in the hydraulic composition additive, and ratio information specifying the content ratio of each of the compounds. This is because the chemical species of the compounds contained in the hydraulic composition additive and their content ratios are particularly dominant as factors that determine the performance of the hydraulic composition additive. In addition, as described above, it is more preferable to use molecular descriptors obtained from the molecular structures of the chemical species as the chemical species information, since this allows a systematic understanding of chemical species that can provide suitable target variables.

[0035] The response variable may be a physical property value selected according to the purpose of the estimation method according to the present embodiment. The physical property values ​​exemplified above are all preferred examples of the response variable in the estimation method according to the present embodiment.

[0036] The above algorithm is applied to the obtained teacher data to obtain a trained model. Typically, a trained model is obtained using a part of the obtained teacher data as training data, and the remaining part is used as test data to verify the trained model. At this time, if the validity of the obtained trained model is low, measures such as changing the number of explanatory variables used in learning, performing dimensionality reduction on the explanatory variables by principal component analysis or the like, adding or changing teacher data, and changing parameters when applying the algorithm are taken to improve the validity of the trained model. Note that preprocessing may be applied to the teacher data as appropriate.

[0037] The estimation process is a process of outputting an estimated value of a dependent variable indicating the performance of a hydraulic composition additive, from explanatory variables indicating the manufacturing conditions (hereinafter referred to as given manufacturing conditions) of the hydraulic composition additive, whose performance is to be estimated, using the trained model generated in the generation process.

[0038] Conventionally, in order to verify the performance of a hydraulic composition additive obtained under given manufacturing conditions, it was necessary to actually manufacture the hydraulic composition additive and measure its physical properties. On the other hand, according to the present embodiment, the objective variables of the hydraulic composition additive obtained under given manufacturing conditions are estimated using a learning model, so that the performance of the hydraulic composition additive can be verified without manufacturing it and measuring its physical properties. This can shorten the trial and error required to determine suitable manufacturing conditions for the hydraulic composition additive.

[0039] An estimation program that causes a computer to realize functions corresponding to the generating step and the estimating step is also an embodiment of the present invention. Also, an estimation device including a storage device that stores the estimation program and a computing device that executes the estimation program is also an embodiment of the present invention.

[0040] Second Embodiment The estimation method according to the second embodiment is an estimation method for outputting explanatory variables related to manufacturing conditions for obtaining an additive for hydraulic composition that satisfies a predetermined standard as a solution from among a plurality of candidates for explanatory variables, and includes a generating step, an estimating step, and an outputting step. Note that the estimation method according to the second embodiment is implemented using a computer.

[0041] In the estimation method according to the second embodiment, a trained model that has already been generated is used. Here, a description will be given assuming that the trained model generated in the model generation step of the estimation method according to the first embodiment is used.

[0042] The generating step is a step of generating a plurality of explanatory variables indicating the manufacturing conditions of the additive for hydraulic composition. In the estimation method according to the present embodiment, explanatory variables related to the manufacturing conditions for obtaining the additive for hydraulic composition that satisfies a predetermined standard are ultimately output as a solution, and the generating step generates explanatory variables that are candidates for the solution.

[0043] There is no particular limitation on the method of generating multiple explanatory variables in the generating step and the number of explanatory variables generated. However, the more explanatory variables generated in the generating step, the greater the amount of computational processing in the estimating step and the outputting step, while the greater the possibility of obtaining a suitable solution. The explanatory variables can be generated by, but are not limited to, an artificial method or computational processing.

[0044] The estimation step is a step of outputting a plurality of objective variables estimated from the plurality of explanatory variables generated in the generation step using a trained model. Compared to the estimation step in the first embodiment, there is a difference in the origin of the explanatory variables given to the trained model, but the procedure itself of giving explanatory variables to the trained model to obtain an estimate of the objective variable is the same.

[0045] The output step is a step of outputting explanatory variables related to manufacturing conditions for obtaining an additive for hydraulic composition that satisfies a predetermined criterion as a solution. The predetermined criterion is, for example, a criterion determined based on the performance desired for the additive for hydraulic composition. The solution output in the output step may be, for example, explanatory variables selected based on criteria such as explanatory variables indicating manufacturing conditions for the additive for hydraulic composition under which one or more predetermined physical property values ​​(objective variables) exceed a predetermined criterion value, explanatory variables indicating manufacturing conditions for the additive for hydraulic composition under which the rank for a predetermined physical property value (objective variable) exceeds a predetermined criterion value, etc.

[0046] In the output step, first, an objective variable that satisfies a predetermined criterion is identified from among the multiple explanatory variables output in the estimation step. Next, for the identified objective variable, an explanatory variable that was input to the trained model to output the objective variable is identified. Then, the explanatory variable identified here is output as a solution.

[0047] Conventionally, in order to obtain a hydraulic composition additive exhibiting desired performance, it was necessary to manufacture various hydraulic composition additives under various manufacturing conditions, measure physical property values, clarify the correlation between manufacturing conditions and performance, and identify suitable manufacturing conditions. On the other hand, according to the present embodiment, manufacturing conditions (explanatory variables) that can obtain a hydraulic composition additive exhibiting desired performance are estimated using a learning model, so that suitable manufacturing conditions can be identified without manufacturing the hydraulic composition additive and measuring its physical properties. This can shorten the trial and error required to determine suitable manufacturing conditions for the hydraulic composition additive.

[0048] An estimation program that enables a computer to realize the functions corresponding to the generating step, the estimating step, and the outputting step is also an embodiment of the present invention. Also, an estimation device including a storage device that stores the estimation program and a computing device that executes the estimation program is also an embodiment of the present invention. Furthermore, when a hydraulic composition additive is manufactured by a manufacturing method that includes a manufacturing step of manufacturing a hydraulic composition additive based on the explanatory variables output as a solution in the outputting step in addition to the generating step, the estimating step, and the outputting step, the hydraulic composition additive is also an embodiment of the present invention.

[0049] (Variation 1: Use of optimization algorithm) In order to obtain a more suitable solution, it is preferable to execute the generating step and the estimating step a number of times. As an example, a method using a genetic algorithm (which is an example of an optimization algorithm) will be described.

[0050] The procedure of the first generation step and estimation step is as described above. Here, the group of explanatory variables generated in the first generation step is referred to as the first generation explanatory variable group, and the group of dependent variables output in the first estimation step is referred to as the first generation dependent variable group. Here, a predetermined number of dependent variables having a high degree of adaptability to a predetermined criterion in the output step are identified from the first generation dependent variable group, and a group of explanatory variables that give the identified dependent variables is identified. This is referred to as the first generation solution. Note that the term n-th generation is used hereafter for the explanatory variable group, dependent variable group, and solution generated in the nth time.

[0051] Since the first-generation solution is a group of explanatory variables that give the additive for hydraulic composition with the best performance (objective variable) among the group of explanatory variables of the first generation, it can be said that the best manufacturing conditions have been narrowed down to a certain extent at the time when the first-generation solution is obtained. On the other hand, the first-generation solution is merely a best range selected from the group of explanatory variables of the first generation, and is not a best range for the whole group of possible explanatory variables. Therefore, if the group of explanatory variables of the first generation itself is far outside the best range, the first-generation solution may not be said to be in a very best range among the whole group of possible explanatory variables. In addition, there may be room for optimization even in the first-generation solution.

[0052] Therefore, in the second generation step, multiple explanatory variable groups (second-generation explanatory variable groups) are generated again based on the first-generation solution. Specifically, explanatory variable groups are generated by expanding the search range centered on the first-generation solution. That is, the first-generation explanatory variable group is a list of explanatory variable candidates that are comprehensively extracted from the entire set of explanatory variables that can be taken in a state where there are no clues or few clues, whereas the second-generation explanatory variable group is a list of explanatory variable candidates that are extracted from a narrowed-down range based on the first-generation solution as a tentative guideline. Therefore, it can be said that the second-generation explanatory variable group is more likely to include a more suitable solution than the first-generation explanatory variable group.

[0053] By performing the second estimation step using the second generation explanatory variable group thus generated, a second generation response variable group is obtained. Also, as in the first step, a second generation solution is obtained. By repeating the generation step and the estimation step in the same manner, and stacking the third, fourth, and subsequent generations, it is expected that the obtained solution will be narrowed down to a more suitable range.

[0054] Although the above description has been given of an example of the application of a genetic algorithm as an optimization algorithm, the optimization algorithm that can be used is not limited to the genetic algorithm. For example, Bayesian optimization and the steepest descent method are also examples of optimization algorithms that can be used.

[0055] (Variation 2: Regeneration of trained models) When the generating step, the estimating step, and the outputting step are performed as described above, explanatory variables are output as a solution (referred to as a first solution). The explanatory variables output here are expected to indicate manufacturing conditions for a hydraulic composition additive with favorable performance (objective variable), but the hydraulic composition additive actually manufactured under those manufacturing conditions may not actually exhibit favorable performance. In this case, it can be said that there is room for improvement in the accuracy of the trained model.

[0056] Therefore, it is preferable to further carry out a model regeneration step. The model regeneration step is a step of adding a pair of explanatory variables output as the first solution and a target variable indicating the performance of the additive for hydraulic composition manufactured under the manufacturing conditions indicated by the explanatory variables to the training data to regenerate the trained model. In other words, an actual evaluation (typically experimental verification) of the explanatory variables output as the first solution is added, and the trained data is regenerated (updated) to reflect the results.

[0057] After performing the regeneration process to regenerate the trained model, the generation process, the estimation process, and the output process may be performed again to obtain a solution (referred to as a second solution). Since the trained model is regenerated through the regeneration process, the second solution may differ from the first solution. Since the accuracy of the trained model is expected to be improved through the regeneration process, the second solution is expected to be a more preferable solution than the first solution.

[0058] If satisfactory performance is not obtained even with the second solution, the model regeneration step can be performed again using a set of explanatory variables output as the second solution and objective variables indicating the performance of the hydraulic composition additive manufactured under the manufacturing conditions indicated by the explanatory variables. In other words, the accuracy of the trained model can be improved by repeating the output of the solution, verification of the solution, and feedback of the verification results to the trained model.

[0059] Other embodiments Regarding other configurations, it should be understood that the embodiments disclosed in this specification are illustrative in all respects and that the scope of the present invention is not limited thereto. A person skilled in the art would easily understand that appropriate modifications are possible without departing from the spirit of the present invention. Therefore, other embodiments modified without departing from the spirit of the present invention are naturally included in the scope of the present invention.

[0060] As one aspect, the present invention can be an estimation method for estimating characteristics of an additive for hydraulic composition, the estimation method including: a model generation step of using multiple pairs of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition and objective variables indicating characteristics of the additive for hydraulic composition as training data to generate a trained model that estimates the objective variables from the explanatory variables; and an estimation step of using the trained model to output an estimated value of the objective variable indicating the characteristics of the additive for hydraulic composition from the explanatory variables indicating manufacturing conditions of the additive for hydraulic composition whose characteristics are to be estimated.

[0061] As one aspect, the present invention can be an estimation program for estimating the performance of an additive for hydraulic composition, which, when executed by a computer, causes the computer to realize a model generation function for using, as training data, a plurality of pairs of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition and a target variable indicating the performance of the additive for hydraulic composition, to generate a trained model for estimating the target variable from the explanatory variables, and an estimation function for using the trained model to output an estimated value of the target variable indicating the performance of the additive for hydraulic composition, from the explanatory variables indicating the manufacturing conditions of the additive for hydraulic composition, whose performance is to be estimated.

[0062] As one aspect, the present invention can be an estimation device including a calculation device and a storage device, wherein the storage device stores an estimation program, and the estimation program, when executed by the calculation device, causes the calculation device to realize a model generation function of using a plurality of pairs of explanatory variables indicating manufacturing conditions of a hydraulic composition additive and a target variable indicating performance of the hydraulic composition additive as training data to generate a trained model that estimates the target variable from the explanatory variables, and an estimation function of using the trained model to output an estimated value of the target variable indicating performance of the hydraulic composition additive from the explanatory variables indicating manufacturing conditions of the hydraulic composition additive, the performance of which is to be estimated.

[0063] According to these configurations, the objective variables of the additive for hydraulic composition obtained under given manufacturing conditions are estimated using a learning model, so that the performance of the additive for hydraulic composition can be verified without manufacturing and evaluating the additive for hydraulic composition, thereby shortening the trial and error required to determine suitable manufacturing conditions for the additive for hydraulic composition.

[0064] As one aspect, the present invention may be an estimation method further including a model generation step of using multiple pairs of explanatory variables indicating manufacturing conditions of an additive for hydraulic composition and objective variables indicating characteristics of the additive for hydraulic composition as training data to generate a trained model that estimates the objective variables from the explanatory variables.

[0065] In one aspect of the present invention, the objective variable may include at least one physical property value selected from the group consisting of fluidity, fluidity retention, air entrainment, air amount retention, viscosity, pumpability, material separation resistance, finishability, adhesion, setting characteristics, strength characteristics, shrinkage characteristics, frost resistance, carbonation characteristics, chloride penetration resistance, steel protection, alkali-silica reaction resistance, crack resistance, watertightness, fire resistance, abrasion resistance, and surface aesthetics of the hydraulic composition to which the hydraulic composition additive has been added.

[0066] In one aspect, the present invention provides a model generation process using multiple pairs of explanatory variables indicating manufacturing conditions of a hydraulic composition additive and objective variables indicating characteristics of the hydraulic composition additive as training data to generate a trained model that estimates the objective variables from the explanatory variables; a generation process generating multiple explanatory variables indicating manufacturing conditions of the hydraulic composition additive; an estimation process using the trained model to output multiple objective variables estimated from the multiple explanatory variables generated in the generation process; a process specifying a predetermined criterion that the objective variables should satisfy; and a process for specifying an objective variable that satisfies the predetermined criterion from among the multiple objective variables output in the estimation process, the process being used when outputting the objective variable that satisfies the predetermined criterion from among the multiple explanatory variables generated in the generation process. The hydraulic composition additive may be characterized in that the hydraulic composition additive is manufactured by a manufacturing method including an output step of outputting the explanatory variables generated as a solution in the output step, and a manufacturing step of manufacturing the hydraulic composition additive based on the explanatory variables output as a solution in the output step, the generating step and the estimating step being executed multiple times, and in the second or subsequent generating step, a plurality of explanatory variables are generated using a genetic algorithm that utilizes the results of the generating step and the estimating step executed before the generating step, and the objective variable includes at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in the hydraulic composition additive. According to this configuration, an additive for hydraulic composition that is highly expected to exhibit suitable performance can be realized more easily than before. EXAMPLES

[0067] The present invention will be further described below with reference to examples, but the present invention is not limited to the following examples.

[0068] [Example 1] Estimation of the optimal composition of strength enhancer (Teacher data) Forty-six types of hydraulic composition additives were prepared, each with different chemical species and content ratios of compounds contained therein. For each hydraulic composition additive, water, cement, sand as fine aggregate, gravel as coarse aggregate, and the hydraulic composition additive were mixed in a conventional manner to prepare a hydraulic composition. The compressive strength of each hydraulic composition was measured at 1 day of age according to JIS A 1108:2018, and the compressive strength ratio was calculated based on the compressive strength when the hydraulic composition additive was not used. Based on the measurement results, training data was obtained with chemical species information related to the chemical species of compounds contained in the hydraulic composition additive, ratio information related to the content ratio of each compound, manufacturing condition information related to the manufacturing conditions of the hydraulic composition additive, and composition information related to the conditions related to the hydraulic composition as explanatory variables, and the compressive strength ratio as the objective variable. The chemical species information includes molecular descriptors, and the molecular descriptors were obtained from the molecular structures of each compound expressed in SMILES notation using the chemoinformatics tool RDKit in a Python environment.

[0069] (Generating and validating trained models) Using the machine learning library scikit-learn in a Python environment, we generated and verified trained models using the above training data. Five types of algorithms were selected as candidates for model generation: lasso regression, ridge regression, elastic net regression, support vector regression, and random forest regression. Multiple trained models were generated within the search range of 5 to 15 folds for cross-validation and 5 to 15 explanatory variables. Among the trained models generated, the coefficient of determination R 2The trained model with the largest ratio was adopted. The explanatory variables in the trained model adopted are, for example, fr_Al_OH (number of aliphatic hydroxyl groups), fr_COO (number of carboxylic acids), and PEOE_VSA2 (degree of electrostatic interaction), all of which are explanatory variables that affect adsorption to the binder (cement in this case), so it can be said that it is reasonable for these to have a correlation with the compressive strength ratio. In other words, the trained model adopted can be said to be a reasonable model for those skilled in the art.

[0070] The results of verifying the validity of the adopted trained model are shown in Figure 1. The graph in Figure 1 shows the actual measured value of the compressive strength ratio for each data that constitutes the training data on the horizontal axis, and the estimated value of the compressive strength ratio estimated from the explanatory variables using the adopted trained model on the vertical axis. The coefficient of determination R of the adopted trained model 2 The prediction accuracy was 0.88, which was confirmed to be sufficient for practical use.

[0071] (Generation of explanatory variables) Next, a group of explanatory variables was generated to estimate the manufacturing conditions for obtaining suitable additives for hydraulic compositions as a solution. The explanatory variables were generated using a genetic algorithm using the evolutionary computing framework DEAP in a Python environment. Ranges were set in advance for each of the chemical species information, ratio information, manufacturing condition information, and composition information of the additives for hydraulic compositions, and the initial population was generated so that each condition was randomly determined within this range. Fitness was evaluated as being higher in the order of the compressive strength ratio estimated using the above trained model. The selection of individuals to be left in the next generation was based on an elite preservation strategy. The termination condition was set to when the number of generations reached 100.

[0072] (Estimation and verification of favorable manufacturing conditions) The compressive strength ratio was estimated for each explanatory variable included in the above explanatory variable group using the above trained model. Among the estimated values ​​of the compressive strength ratio obtained, the explanatory variables that gave the two largest estimates were selected as candidates. The compressive strength ratio estimated from the first candidate explanatory variable was 115%, and the compressive strength ratio estimated from the second candidate explanatory variable was 113%.

[0073] Next, the hydraulic composition additive was actually manufactured according to the manufacturing conditions indicated by the explanatory variables of the first and second candidates, and the compressive strength ratio of the hydraulic composition prepared by blending the obtained hydraulic composition additive was actually measured. The actual measured value of the compressive strength ratio was 108% for the manufacturing conditions of the first candidate, and 105% for the manufacturing conditions of the second candidate. It can be said that there was a large discrepancy between the estimated value of the compressive strength ratio and the actual measured value for both the first and second candidates. Next, the trained model was regenerated using the pairs of explanatory variables and the measured compressive strength ratios obtained for the first and second candidates.

[0074] The compressive strength ratio was estimated for each explanatory variable included in the above explanatory variable group using the regenerated trained model. As in the first estimation, the explanatory variables that gave the two largest estimated values ​​were selected as candidates, and were designated as the third and fourth candidates. The compressive strength ratio estimated from the third candidate explanatory variable was 121%, and the compressive strength ratio estimated from the fourth candidate explanatory variable was 118%.

[0075] Similarly, for the third and fourth candidates, additives for hydraulic compositions were manufactured, hydraulic compositions were prepared, and the compressive strength ratio was measured. The measured compressive strength ratio was 120% for the third candidate and 118% for the fourth candidate. Estimation using the regenerated trained model showed good agreement between the estimated and measured values ​​of the compressive strength ratio. Compared to the first and second candidates, the deviation between the estimated and measured values ​​was smaller, which indicates that the trained model has been improved.

[0076] In addition, when attempting to maximize the compressive strength ratio using conventional experimental methods, 35 more experiments (manufacturing additives for hydraulic compositions, preparing hydraulic compositions, and measuring compressive strength ratios) were required starting from the point where the training data was obtained. In addition, the control sample obtained through conventional experimental methods had a compressive strength ratio of 115%, which was at the same level as the third and fourth candidates. For the third and fourth candidates, the manufacturing conditions for additives for hydraulic compositions with the level of performance that could ultimately be obtained using conventional experimental methods were identified through actual measurements of only four points.

[0077] [Example 2] Estimation of the optimal composition of water-reducing components (Teacher data) Ninety types of hydraulic composition additives were prepared, each having different chemical species and content ratios of compounds contained therein. For each hydraulic composition additive, a hydraulic composition was prepared by mixing water, cement, sand as fine aggregate, gravel as coarse aggregate, and the hydraulic composition additive in a conventional manner. For each hydraulic composition, the slump value immediately after preparation and 60 minutes after preparation was measured according to JIS A 1101, and the value obtained by subtracting the slump value 60 minutes after preparation from the slump value immediately after preparation (hereinafter referred to as "ΔSL") was determined. Based on the measurement results, chemical species information related to the chemical species of compounds contained in the hydraulic composition additive, ratio information related to the content ratio of each compound, reaction condition information related to the reaction conditions for generating each compound (including, in particular, information on the chemical species and amount of the monomer, polymerization initiator, and chain transfer agent used), manufacturing condition information related to the manufacturing conditions of the hydraulic composition additive, and composition information related to the conditions related to the hydraulic composition were used as explanatory variables, and training data was obtained with ΔSL as the objective variable. The chemical species information includes molecular descriptors, which were obtained from the molecular structure of each compound expressed in SMILES notation using the chemoinformatics tool RDKit in a Python environment.

[0078] (Generating and validating trained models) Using the machine learning library scikit-learn in a Python environment, we generated and verified trained models using the above training data. Five types of algorithms were selected as candidates for model generation: lasso regression, ridge regression, elastic net regression, support vector regression, and random forest regression. Multiple trained models were generated within the search range of 5 to 15 folds for cross-validation and 5 to 15 explanatory variables. Among the trained models generated, the coefficient of determination R 2 The trained model in which the largest value was adopted was used. The explanatory variables in the trained model adopted are, for example, the mass average molecular weight of the surfactant contained in the additive for hydraulic compositions, fr_COO (number of carboxylic acids), and PEOE_VSA3 (degree of electrostatic interaction), all of which are explanatory variables that affect adsorption to the binder (cement in this case), and therefore it can be said that these are all reasonable explanatory variables that have a correlation with the fluidity of the hydraulic composition. In other words, the trained model adopted can be said to be a reasonable model for those skilled in the art.

[0079] The results of verifying the validity of the adopted trained model are shown in Figure 2. The graph in Figure 2 shows the actual measured value of ΔSL for each data that constitutes the training data on the horizontal axis, and the estimated value of ΔSL estimated from the explanatory variables using the adopted trained model on the vertical axis. The coefficient of determination R 2 The prediction accuracy was 0.63, which was confirmed to be sufficient for practical use.

[0080] (Generation of explanatory variables) Next, a group of explanatory variables was generated to estimate the manufacturing conditions for obtaining suitable additives for hydraulic compositions as a solution. The explanatory variables were generated by a genetic algorithm using the evolutionary computing framework DEAP in a Python environment. Ranges were set in advance for each of the chemical species information, ratio information, reaction condition information, manufacturing condition information, and composition information of the additives for hydraulic compositions, and the initial population was generated so that each condition was randomly determined within this range. Fitness was evaluated as being higher in the order of the smaller ΔSL value estimated using the above trained model. The selection of individuals to be left in the next generation was based on an elite preservation strategy. The termination condition was set to when the number of generations reached 100.

[0081] (Estimation and verification of favorable manufacturing conditions) For each explanatory variable included in the explanatory variable group above, ΔSL was estimated using the trained model above. Among the estimated values ​​of ΔSL obtained, the explanatory variables that gave the two smallest estimated values ​​were selected as candidates. ΔSL estimated from the first candidate explanatory variable was 2.4 cm, and ΔSL estimated from the second candidate explanatory variable was 2.8 cm.

[0082] Next, the additive for hydraulic compositions was actually manufactured according to the manufacturing conditions indicated by the explanatory variables of the first and second candidates, and the ΔSL of the hydraulic composition prepared by blending the obtained additive for hydraulic compositions was actually measured. The measured value of ΔSL was 4.0 cm for the manufacturing conditions of the first candidate, and 4.5 cm for the manufacturing conditions of the second candidate. It can be said that the deviation between the estimated value of ΔSL and the measured value was large for both the first and second candidates. Next, the trained model was regenerated using the pairs of explanatory variables and the measured ΔSL obtained for the first and second candidates.

[0083] For each explanatory variable included in the explanatory variable group above, ΔSL was estimated using the regenerated trained model. As in the first estimation, explanatory variables that gave the two smallest estimated values ​​were selected as candidates, and were designated as the third and fourth candidates. ΔSL estimated from the third candidate explanatory variable was 1.1 cm, and ΔSL estimated from the fourth candidate explanatory variable was 1.4 cm.

[0084] Similarly, for the third and fourth candidates, additives for hydraulic compositions were manufactured, hydraulic compositions were prepared, and ΔSL was measured. The measured values ​​of ΔSL were 1.0 cm for the third candidate and 1.4 cm for the fourth candidate. Estimation using the regenerated trained model showed good agreement between the estimated and measured values ​​of ΔSL. Since the deviation between the estimated and measured values ​​was smaller than for the first and second candidates, it can be said that the trained model has been improved.

[0085] Furthermore, when attempting to minimize ΔSL using conventional experimental methods, 20 more experiments (manufacturing additives for hydraulic compositions, preparing hydraulic compositions, and measuring ΔSL) were required starting from the point where the training data was obtained. The control sample obtained through conventional experimental methods had a ΔSL of 1.2 cm, which was at the same level as the third and fourth candidates. For the third and fourth candidates, the manufacturing conditions for additives for hydraulic compositions with the level of performance that could ultimately be obtained through conventional experimental methods were identified through actual measurements of only four points.

[0086] [Example 3] Optimization of the mass average molecular weight of the surfactant contained in the additive for hydraulic composition In the following, an example will be described in which the estimation method according to this embodiment is applied to the estimation of the mass average molecular weight of a surfactant (here, a polycarboxylic acid ether compound) contained in an additive for hydraulic compositions. When a polycarboxylic acid ether compound is used as an additive for hydraulic compositions, it is empirically known that when the mass average molecular weight is around 20,000, the compound has good adsorption to a binder (here, cement). In this example, the objective is to estimate the manufacturing conditions that give a polycarboxylic acid ether compound with a mass average molecular weight of around 20,000. Note that the setting of the suitable mass average molecular weight in this example is merely one example, and that the suitable mass average molecular weight may change depending on conditions such as the form of the hydraulic composition to which it is applied.

[0087] (Sample for creating training data) Regarding the polycarboxylic acid ether-based compound contained in the additive for hydraulic compositions, 27 types were prepared with different chemical species and ratios of raw materials (unsaturated polyalkylene glycol monomer, unsaturated carboxylic acid monomer, vinyl monomer copolymerizable with unsaturated polyalkylene glycol monomer, polymerization initiator, and chain transfer agent).

[0088] (Measurement of mass average molecular weight) The mass average molecular weight of each of the 27 polycarboxylic acid ether compounds was measured by gel permeation chromatography (GPC) under the following conditions. <Measurement conditions> Equipment: Shodex GPC-101 (Showa Denko) Column: OHpak SB-G + SB-806M HQ + SB-806M HQ (Showa Denko) Detector: Differential refractometer (RI) Eluent: 50mM sodium nitrate aqueous solution Flow rate: 0.7mL / min Column temperature: 40℃ Sample concentration: Eluent solution with a sample concentration of 0.5% by mass Standard material: PEG / PEO (Agilent Technologies)

[0089] (Teacher data) Based on the measurement results, the first chemical species information related to the chemical species of the polycarboxylic acid ether compound, the raw material information related to the chemical species and ratio of the raw materials of the polycarboxylic acid ether compound (unsaturated polyalkylene glycol monomer, unsaturated carboxylic acid monomer, vinyl monomer copolymerizable with the unsaturated polyalkylene glycol monomer, polymerization initiator, and chain transfer agent), and the manufacturing condition information related to the manufacturing conditions of the polycarboxylic acid ether compound were used as explanatory variables, and training data was obtained with the mass average molecular weight as the objective variable. Note that the chemical species information includes molecular descriptors, which were obtained from the molecular structure of each compound expressed in SMILES notation using the chemoinformatics tool RDKit in a Python environment.

[0090] (Generating and validating trained models) Using the machine learning library scikit-learn in a Python environment, trained models were generated and verified using the above training data. Five types of algorithms were selected as candidates for use in generating trained models: lasso regression, ridge regression, elastic net regression, support vector regression, and random forest. Multiple trained models were generated within the search range of 5 to 15 folds in cross-validation and 5 to 15 explanatory variables. Of the trained models generated, the coefficient of determination R 2 The trained model in which the largest value was adopted was used. The explanatory variables in the adopted trained model are, for example, fr_SH (the number of thiol groups), the reaction temperature during production, and the reaction time during production, all of which are explanatory variables that affect the reactivity of the raw materials of the polycarboxylic acid ether compound, and therefore it can be said that these are all reasonable explanatory variables that have a correlation with the mass average molecular weight of the polycarboxylic acid ether compound. In other words, the adopted trained model can be said to be a reasonable model for those skilled in the art.

[0091] The results of verifying the validity of the adopted trained model are shown in Figure 3. The graph in Figure 3 shows the actual measured value of the mass average molecular weight of the polycarboxylic ether compound for each data that constitutes the training data on the horizontal axis, and the estimated value of the mass average molecular weight of the polycarboxylic ether compound estimated from the explanatory variables using the adopted trained model on the vertical axis. The coefficient of determination R of the adopted trained model 2 The prediction accuracy was 0.96, which was confirmed to be sufficient for practical use.

[0092] (Generation of explanatory variables) Next, a group of explanatory variables was generated to estimate the manufacturing conditions for obtaining a polycarboxylic ether compound having a suitable mass average molecular weight as a solution. The explanatory variables were generated by a genetic algorithm using the evolutionary computing framework DEAP in a Python environment. Ranges were set in advance for the first chemical species information of the polycarboxylic ether compound, the raw material information of the polycarboxylic ether compound, and the manufacturing condition information, and an initial population was generated so that each condition was randomly determined within this range. Fitness was evaluated as being higher in the order of the mass average molecular weight estimated using the above trained model being closer to 20,000. The selection of individuals to be left in the next generation was based on an elite preservation strategy. The termination condition was set to when the number of generations reached 100.

[0093] (Estimation and verification of favorable manufacturing conditions) For each explanatory variable included in the explanatory variable group, the mass average molecular weight of polycarboxylic ether compounds was estimated using the trained model. Among the estimated values ​​of mass average molecular weight obtained, the explanatory variables giving the two estimates closest to 20,000 were selected as candidates. The mass average molecular weight estimated from the first candidate explanatory variable was 18,000, and the mass average molecular weight estimated from the second candidate explanatory variable was 17,500.

[0094] Next, a polycarboxylic acid ether compound was actually produced according to the production conditions indicated by the explanatory variables of the first and second candidates, and the mass average molecular weight of the obtained polycarboxylic acid ether compound was actually measured. The actual measured value of the mass average molecular weight was 17400 for the production conditions of the first candidate, and 17000 for the production conditions of the second candidate. It can be said that the deviation between the estimated value and the actual value of the mass average molecular weight was large for both the first and second candidates. Next, the trained model was regenerated using the pairs of explanatory variables and the measured mass average molecular weights obtained for the first and second candidates.

[0095] The mass average molecular weight was estimated for each explanatory variable included in the above explanatory variable group using the regenerated trained model. As in the first estimation, the explanatory variables giving the two estimated values ​​closest to 20,000 were selected as candidates, and these were designated as the third and fourth candidates. The mass average molecular weight estimated from the explanatory variable of the third candidate was 20,100, and the mass average molecular weight estimated from the explanatory variable of the fourth candidate was 19,800.

[0096] Similarly, for the third and fourth candidates, polycarboxylic ether compounds were produced and the mass average molecular weight was measured. The measured mass average molecular weight was 20,200 for the third candidate and 19,500 for the fourth candidate. In the estimation using the regenerated trained model, the deviation between the estimated and measured mass average molecular weights was reduced. In addition, in the initial trained model, the mass average molecular weight in the optimal solution was 18,000, which was relatively large from the value (20,000) set as the benchmark, but the optimal solution in the regenerated trained model obtained a mass average molecular weight of 20,100, which is close to the benchmark value. From these points, it can be said that the trained model has been improved.

[0097] In addition, when the optimization of the mass average molecular weight was attempted by the conventional experimental method, 12 more experiments (production of polycarboxylic ether compounds and measurement of mass average molecular weight) were required starting from the point where the teacher data was obtained. In addition, the control sample obtained through the conventional experimental method had a mass average molecular weight of 18,500, which was at the same level as the third and fourth candidates. For the third and fourth candidates, the production conditions of polycarboxylic ether compounds with performance exceeding the solution obtained by the conventional experimental method were identified through actual measurements of only four points. In other words, the present invention made it possible to arrive at a solution of a level that could be obtained by the conventional experimental method without large-scale experiments.

[0098] In addition, in Examples 1 and 2, the estimation was performed using the learning model for the manufacturing conditions of the entire additive for hydraulic compositions, whereas in Example 3, the estimation was performed using the learning model limited to the manufacturing conditions of the polycarboxylic acid ether compound, which is a main component of the additive for hydraulic compositions. In Example 3, it can be said that the range of the target of the estimation using the learning model was reduced compared to Examples 1 and 2 by utilizing the known knowledge that a polycarboxylic acid ether compound with a mass average molecular weight of about 20,000 is preferable for use as an additive for hydraulic compositions. This makes it possible to reduce the scale of the hardware resources and the calculation time required for the estimation. In addition, as for the sample for obtaining the training data, since it is not necessary to prepare the additive for hydraulic compositions and the hydraulic composition in Example 3, the procedure is simplified compared to Examples 1 and 2. In addition, the sample and the learned data of Example 3 may be shared when the polycarboxylic acid ether compound is used for purposes other than additives for hydraulic compositions. In this way, the conventional experimental method and the estimation method according to the present embodiment can be appropriately combined from the viewpoint of minimizing the necessary resources.

[0099] [Summary] As is clear from the above Examples 1, 2 and 3, the present invention makes it possible to arrive at a solution of a level that can be obtained by conventional experimental methods without large-scale experiments. This is expected to reduce the man-hours, time and costs required for developing a new additive for hydraulic compositions. [Industrial Applicability]

[0100] The present invention can be utilized, for example, for estimating the performance of an additive for a hydraulic composition and for estimating the manufacturing conditions that provide a suitable additive for a hydraulic composition.

Claims

1. A model generation process for generating a trained model that estimates the objective variable from the explanatory variables by using a plurality of pairs of explanatory variables indicating the manufacturing conditions of the additive for hydraulic composition and objective variables indicating the characteristics of the additive for hydraulic composition as training data; A generation step of generating a plurality of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition; an estimation step of outputting a plurality of objective variables estimated from a plurality of explanatory variables generated in the generation step, using the trained model; specifying predetermined criteria that the objective variable must satisfy; an output step of identifying a dependent variable that satisfies the predetermined criterion from among the plurality of dependent variables output in the estimation step, and outputting, as a solution, an explanatory variable that was used when outputting the dependent variable that satisfies the predetermined criterion from among the plurality of explanatory variables generated in the generation step, The generating step and the estimating step are performed a plurality of times; generating a plurality of explanatory variables using a genetic algorithm that utilizes results of the generating step and the estimating step executed before the second or subsequent generating step; The estimation method, characterized in that the objective variable includes at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in an additive for a hydraulic composition.

2. 2. The estimation method according to claim 1, further comprising a model regeneration step of adding a pair of explanatory variables output as a solution in the output step and a target variable indicating the performance of the additive for hydraulic composition manufactured under the manufacturing conditions indicated by the explanatory variables to training data, and regenerating the trained model.

3. The explanatory variables are Chemical species information determined by the chemical species of a compound contained in the additive for hydraulic composition; and ratio information specifying a content ratio of each of the compounds.

4. The method of claim 3 , wherein the chemical species information includes a molecular descriptor of the compound.

5. 2. The method according to claim 1, wherein the objective variable includes at least one physical property value selected from the group consisting of fluidity, fluidity retention, air entrainment, air retention, viscosity, pumpability, material separation resistance, finishability, adhesion, setting characteristics, strength characteristics, shrinkage characteristics, frost resistance, carbonation characteristics, chloride penetration resistance, steel protection, alkali-silica reaction resistance, crack resistance, watertightness, fire resistance, abrasion resistance, and surface aesthetics of the hydraulic composition to which the additive for hydraulic compositions has been added.

6. When executed by a computer, the computer: A model generation function that uses a plurality of pairs of explanatory variables indicating manufacturing conditions of the additive for hydraulic compositions and objective variables indicating characteristics of the additive for hydraulic compositions as training data to generate a trained model that estimates the objective variables from the explanatory variables; A generation function for generating a plurality of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition; an estimation function that outputs a plurality of objective variables estimated from a plurality of explanatory variables generated by the generation function using the trained model; A function of accepting a specification of a predetermined criterion that the objective variable should satisfy; an output function of identifying a response variable that satisfies the predetermined criterion from among the plurality of response variables output by the estimation function, and outputting, as a solution, explanatory variables that were used when outputting the response variable that satisfies the predetermined criterion from among the plurality of explanatory variables generated by the generation function, performing said generating function and said estimating function a plurality of times; an estimation program, characterized in that in the generation function realized for the second or subsequent time, a plurality of explanatory variables are generated using a genetic algorithm that utilizes the results of the generation function and the estimation function realized prior to the generation function, and the objective variable includes at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in an additive for hydraulic composition.

7. An estimation device comprising a computing device and a storage device, The storage device stores an estimation program, When the estimation program is executed by the arithmetic device, the arithmetic device A model generation function that uses a plurality of pairs of explanatory variables indicating manufacturing conditions of the additive for hydraulic compositions and objective variables indicating characteristics of the additive for hydraulic compositions as training data to generate a trained model that estimates the objective variables from the explanatory variables; A generation function for generating a plurality of explanatory variables indicating manufacturing conditions of the additive for hydraulic composition; an estimation function that outputs a plurality of objective variables estimated from a plurality of explanatory variables generated by the generation function using the trained model; A function of accepting a specification of a predetermined criterion that the objective variable should satisfy; an output function of identifying a response variable that satisfies the predetermined criterion from among the plurality of response variables output by the estimation function, and outputting, as a solution, explanatory variables that were used when outputting the response variable that satisfies the predetermined criterion from among the plurality of explanatory variables generated by the generation function, performing said generating function and said estimating function a plurality of times; An estimation device characterized in that, in the generation function realized for the second or subsequent time, a plurality of explanatory variables are generated using a genetic algorithm that utilizes the results of the generation function and the estimation function realized prior to the generation function, and the objective variable is a program including at least one physical property value selected from the group consisting of number average molecular weight, weight average molecular weight, viscosity average molecular weight, molecular weight dispersity, melting point, glass transition point, HLB, SP value, particle size, particle size distribution, and particle shape of a surfactant contained in an additive for hydraulic composition.

Images