Mixture descriptor-based modeling of highly formulated products

Abstract

A method for mixture descriptor-based modeling of highly compounded products such as paints is disclosed. The method includes categorizing the components of a dataset into multilevel classifications to generate a reduced dataset; incorporating one or more descriptors associated with the components into the reduced dataset to generate a modified dataset; receiving predictions of product characteristics from a machine learning module; and adjusting the chemical formulation and / or process based on the predictions of product characteristics to produce or reject the product.

Description

Technical Field

[0001] The present disclosure relates to mixture descriptor-based modeling of highly formulated products. Such techniques can be particularly useful for predicting product properties in order to adjust the chemical formulations used to produce a product or to determine whether to reject a particular chemical formulation for producing a product.

Background Art

[0002] Modern chemical products are often highly formulated to contain a number of components with diverse functions and compositions. For example, conventional architectural paints can contain a range of 10 to 20 individual components from functional additives such as inorganic pigments, binders, dispersants, rheology modifiers, adhesion promoters, etc. The individual components form a large formulation space in which combinations and specific component amounts are selected to balance the desired properties for each application. Further complexity includes variations in product properties between formulations containing different components of the same type. In the case of extender calcium carbonate and silica, substitution between components of similar size can result in performance changes because other characteristics, such as sphericity, surface roughness, free ions, etc., promote different interactions between other components.

[0003] In the case of binders, synthesis has an effect that is the same as or greater than the composition of the binder itself. When conceptualizing complex formulations such as architectural paints, the vastness of the combinatorial space (e.g., very high dimensionality) requires large datasets to accurately capture the dynamics of the system. In addition to the vastness of the formulation space, the time and energy required to generate data should not be ignored. Many researchers will consider the time and energy required to generate an amount of data sufficient to adequately describe an infeasible space.

[0004] One approach currently employed is to limit the range of the formulation space or to limit the complexity of the desired output. Constraining the formulation space often results in models with a limited range of applicability. [Overview of the Initiative]

[0005] This disclosure relates to the use of improvements in machine learning techniques for predicting the properties of products produced by chemical processes. Predictions may be based on a dataset containing multiple descriptors categorized based on material type and associated with the components of the product. A dataset containing multiple descriptors may be generated and input into an artificial neural network (ANN) trained to predict product properties based on the components of the product. In addition, a dataset containing multiple descriptors may be generated and input into an ANN trained to predict the components of a product containing desired properties.

[0006] The above summary of this disclosure is not intended to describe each disclosed embodiment, nor to describe all embodiments of this disclosure. More specifically, this specification illustrates exemplary embodiments. In several places throughout this application, guidance is provided through lists of examples, which can be used in various combinations. In all cases, the enumerated lists serve only as representative groups and should not be construed as exclusive lists. [Brief explanation of the drawing]

[0007] [Figure 1] One exemplary model 100 is shown for reducing the component-associated dataset based on the material type of the component. [Figure 2] This is an illustrative diagram showing an approach to generate a reduced dataset from an initial dataset and add component characteristic descriptors to the reduced dataset. [Figure 3]This is an illustrative diagram illustrating an approach to adding component characteristic descriptors to a model that can be used as part of a machine learning module. [Figure 4] This document presents an example of a method for descriptor-based modeling. [Figure 5] An example of a machine-readable medium for descriptor-based modeling is shown. [Figure 6] An example of a device for descriptor-based modeling is shown. [Modes for carrying out the invention]

[0008] This disclosure relates to a method and device for mixture descriptor-based modeling of highly formulated products that can use a machine learning model to predict product characteristics for one or more predicted formulations.

[0009] A machine learning model can be a function or equation for identifying patterns in data. A machine learning module can be multiple machine learning models used together to identify patterns in data. In certain examples, a machine learning module can be organized as a neural network. A neural network can contain a set of instructions that can be executed to recognize patterns in data. Some neural networks can be used to recognize underlying relationships in a set of data in a manner that mimics how the human brain works. Neural networks can adapt to fluctuating or changing inputs so that they can produce the best possible results without the neural network having to redesign its output criteria.

[0010] A neural network may contain multiple neurons, which can be represented by one or more equations or functions. In the context of neural networks, a neuron may receive a number or vector of a certain quantity as input and generate an output based on the properties of the neural network. For example, a neuron may have X kA neuron can receive n inputs, where k corresponds to the index of the input. For each input, the neuron generates a weight vector W. k The input can be assigned to a weight vector (e.g., weight values). In some embodiments, a neuron in a neural network can be distinguished from one or more different neurons in the network. In some neural networks, each input vector can be multiplied by its respective weight vector to obtain a value, as shown by Equation 1, which illustrates an example of a linear combination of input vectors and weight vectors. f(x1,x2) = w1x1 + w2x2 formula 1

[0011] In some neural networks, a nonlinear function (e.g., an activation function) can be applied to the values ​​fx1, x2) obtained from Equation 1. An example of a nonlinear function that can be applied to the values ​​obtained from Equation 1 is the rectified linear unit function (ReLU). The application of the ReLU function shown by Equation 2 results in a value input to the function when the value is greater than zero, or zero when the value input to the function is less than zero. The ReLU function is used herein only as an illustrative example of an activation function and is not intended to be an limitation. Other non-restrictive examples of activation functions that can be applied in the context of neural networks may include, among others, the sigmoid function, the binary step function, the linear activation function, the hyperbolic function, the leaky ReLU function, the parametric ReLU function, the softmax function, and / or the swish function. ReLU(x) = max(x, 0) formula 2

[0012] During the process of training a neural network, the input vector and / or weight vector may be modified to "tune" the network. In at least one example, the neural network may be initialized with random weights. Over time, the weights may be tuned to improve the accuracy of the neural network. This can result in a neural network with high accuracy over time. This disclosure utilizes machine learning, such as neural networks, to predict product characteristics through modeling of input data. In these embodiments, the weights may be tuned based on several factors. For example, the weights may be tuned with respect to descriptors based on the amount of an additional component containing the descriptor (e.g., weight fraction).

[0013] Embodiments of this disclosure include mixed descriptor-based modeling of highly formulated products utilizing machine learning modules such as neural networks. The methods disclosed herein may be used to improve product characteristic predictions from limited datasets of multi-component chemical formulations in a variety of applications, including paints, household and personal care compositions. Product characteristic predictions may be generated from one or more machine learning models, such as neural networks, random forests, or others, and may include one or more qualitative or quantitative values. The methods disclosed herein may also be applied to inverse strategies in which target product characteristics are input into one or more machine learning modules to generate outputs for one or more predicted product formulations.

[0014] Before training a machine learning model, formulation components may be organized into a multilevel classification to generate a reduced dataset. The multilevel classifications disclosed herein may include two or more classes arranged at one or more levels. At the first level, a class is generated for each component of any size (e.g., binders, surfactants, water, pigments, etc.). As used herein, “class” refers to a category used to define the type of component. The model is used to predict one or more target properties and determine the degree of fit. More detailed models are generated by further subdividing the classes (e.g., subdividing the class of binders into subclasses of epoxy, polyether, vinyl, acrylic, and polyurethane; subdividing the class of surfactants into subclasses of anionic, cationic, and amphoteric). The model may be further evolved with more specific classifications depending on the results of the fit to one or more target properties. The number of levels of each classification required depends on the property being modeled. For example, additional levels of classification can improve the fit when surfactants can be subdivided based on their chemical type (e.g., charge such as anionic or cationic, molecular weight, functional value, etc.) or function (e.g., effect on rheology, emulsion type, etc.). In other examples, a single level may be used for a component. For example, a component such as water, though not limited to this, may have a single level. Multiple models are generated and sorted (keep / discard) based on their fit to the characteristic data and the uncertainty associated with each model. In some cases, multiple models may be generated and applied as modules, and the result is the combined product of two or more models (e.g., mean, weighted mean, etc.).

[0015] As used herein, the singular forms “a,” “an,” and “the” include singular and plural nouns unless the context expressly indicates otherwise. Furthermore, the word “may” is used throughout this application in an allowable sense (e.g., possible, may) rather than an obligatory sense (e.g., must). The term “includes,” and its derivatives, means “includes, but not limited to.”

[0016] As should be understood, elements shown in various embodiments of this specification may be added, replaced, and / or excluded to provide some additional embodiments of the present disclosure. In addition, as should be understood, the proportions and relative scales of the elements provided in the figures are intended to illustrate certain embodiments of the invention and should not be construed as limiting.

[0017] Figure 1 shows one exemplary model 100 for reducing a dataset associated with components based on the material type of the components. Each labeled box in Figure 1 represents a different class of component. A class can be a category or description used to organize components based on their properties (e.g., material type, composition, etc.). Classes are represented by different levels 102-1, 102-2, 102-3, ..., 102-N, also called class levels. Model 100 includes a first class level 102-1 ("Level 1") and may include a second level 102-2 ("Level 2"), a third level 102-3 ("Level 3"), and additional levels such as an additional level 102-N. Additional levels below the first level 102-1 may be called subclasses. Subclasses can be extended to any number of levels necessary to capture the details of the associated components.

[0018] As used herein, “components” refers to the physical substances (e.g., chemical species, reactants, raw materials, etc.) used in a formulation to produce a product. The components of a formulation used to produce a product (e.g., “final product,” “finished product,” etc.) may be described by multiple different data metrics within a dataset. A dataset may contain multiple different components for producing a product. As described herein, a “dataset” may be an exhaustive list of possible components that may be difficult to organize and / or make available for training a machine learning module.

[0019] As described herein, a machine learning module can include a plurality of functions or equations that can be organized as nodes that recognize underlying relationships within a dataset in a manner that mimics how the human brain operates. The machine learning module can be trained using data compiled by model 100. Model 100 can be compiled at different levels 102-1, 102-2, 102-3, ..., 102-N, and descriptors can be added to one or more of the classes at the different levels 102-1, 102-2, 102-3, ..., 102-N. In at least one example, related historical formulation data, including the names of all components, is collected. The components are then categorized using the multi-level class definitions of model 100. In some examples, the descriptors can include at least one of a continuous descriptor, an ordinal descriptor, a binary descriptor, or a categorical descriptor. For example, the descriptors can be continuous values, ordinal values, binary values, and / or categorical values.

[0020] The class definition of a component can be used to define the type of the component and the reactants used to generate the component, among other categorizations of the component. The class definition must include a level 1 class and can include one class for each level below level 1 within the same branch of the illustrated tree structure. Referring to model 100, some examples of class definitions include (A), (B), (C), (A, D), (A, D, N), etc. In some examples, each class definition can include one or more descriptors.

[0021] As used herein, “descriptor” refers to a quantity or quality attributable to a given class of component. For example, a component having class definition (A) may have a specific characteristic or attribute assigned to class definition (A). In this way, descriptors can be used to further describe the quality or attribute of a component designated to a particular class definition. Descriptors may include, but are not limited to, measured characteristics, empirical characteristics, calculated characteristics, categorical characteristics, and / or compositional characteristics. Measured characteristics may be measurements collected during a particular test of a component. For example, the descriptor “acidic” may be assigned to a particular class. In this example, a component designated to that class will have a measured pH of less than 7 under specific conditions. Empirical characteristics may be experimental results of a particular class. Empirical characteristics of a component can be collected through experimental data. Empirical characteristics may be the result of combining different quantities of a component or a class of components with different components. In this way, experimental results of a particular class or a particular component may be assigned to a class or a particular component in Model 100. Compositional characteristics may be the compositional characteristics of a class of components. For example, a specific property can be associated with a binder, such as an acrylic binder. In this example, the specific property is a compositional property and can be specified as a class of acrylic binder.

[0022] In some examples, descriptors can be calculated descriptors. For example, a calculated descriptor for a paint pigment can be calculated based on the amount and density value of the pigment, using the concentration or volume solid of a particular pigment. In these examples, the calculated descriptor can be used for a particular component so that the calculation can be applied to the component. In some examples, descriptors can be categorical descriptors. For example, the descriptor can be the type of monomer (e.g., PEM, MMA, AA, etc.) and whether the monomer is compound-forming or not (e.g., a yes / no category). In these examples, the categorical descriptor can be one-hot coded based on the yes / no category description, and a weighted average can be used if a blend or mixture of components (e.g., a mixture of different monomers, etc.) exists.

[0023] Considering FIG. 1, an example is presented where the components may have a multi-level class definition such as (A), (A, D), or (A, D, N). The second level 102-2 (e.g., level 2) classes (D), (E), and (F) are available for the first level 102-1 class (A). However, as shown in FIG. 1, the second level 102-2 class (D) may also be available for the first level 102-1 class (B).

[0024] Each class at each level may be associated with one or more descriptors. Descriptors may or may not be unique for each class; that is, a particular descriptor may be assigned to multiple different classes. Each machine learning module may include one or more machine learning models, such as artificial neural networks (ANNs) including deep neural networks (DNNs), recurrent neural networks (RNNs) including symbolic regression, long short-term memory (LSTM) networks, or gated recurrent unit (GRU) networks, decision trees, random forests, boost trees such as gradient boost trees (XGBoost), linear regression, partial least squares regression, support vector machines, ridge regression, multilayer perceptrons (MLPs), autoencoders (e.g., denoising autoencoders such as stacked denoising autoencoders), Bayesian networks, support vector machines (SVMs), and hidden Markov models (HMNIs). Commercially available software packages may include JMP software, Microsoft AzureML, SAP data analysis tools, Python, R Project, and soft independent modeling by class analogy (SIMCA) using Sartorius. While specific types of machine learning models are described, other types of machine learning models may implement descriptors in a similar manner.

[0025] Methods further described herein may include training or constructing one or more machine learning modules, and then evaluating the effectiveness of the machine learning modules using a measure of goodness of fit (e.g., R-squared). Machine learning modules are constructed with a user-defined number of classes and level depth associated with the model. For each module, referring to Figure 1, classes may be considered at various level depths. For example, a first module may be constructed considering the components of each class at a first level 102-1 ((A), (B), and (C)). A second module may be constructed where components of classes (A) and (C) are considered at the first level 102-1, while (B) is considered at the second level 102-2 (in particular, the class definitions of (A), (B), (B, G), (B, H), (B, D), and (C)). In the second module example, components of (B, G), (B, H), or (B, D) that do not have a second level 102-2 class definition retain the first level 102-1 class definition of (B).

[0026] In the third module, the components of class (A) are considered at the first level 102-1, the components of class (B) are considered at the second level 102-2, and the components of class (C) are considered at the third level 102-3. In this example, all components will have one of the following class definitions: (A), (B), (B, G), (B, H), (B, D), (C), (C, J), (C, K), (C, L), (C, L, S), and (C, L, T). Thus, the components considered at the first level 102-1 are defined by the categorization of components at the first level 102-1. In addition, the components considered at the second level 102-2 are defined by the categorization of components at the first level 102-1, the second level 102-2, and / or a combination thereof. Furthermore, the components considered in the third level 102-3 will be defined by the categorization of components in the first level 102-1, the second level 102-2, the third level 102-3, and / or combinations thereof.

[0027] A machine learning module is constructed / trained using one or more of the following: component classes, class definitions for each component and associated descriptors, and component quantities (e.g., component quantities, component ratios, etc.). As further described herein, depending on the modeling task, descriptors may be weighted by the quantities of components in the product using various algebraic approaches (e.g., generalized mean approaches), or they may be used without further modification. The resulting machine learning module is then evaluated for effectiveness, either individually or in an ensemble (a combination of results from multiple models), by considering metrics relevant to the experimental data and assessing the goodness of fit (e.g., fitted values, etc.).

[0028] Traditional approaches to modeling a formulation space typically weight each component as a unique factor. The method disclosed herein constructs a machine learning module that allows specific information or data to be distilled into a dataset as descriptors, enabling the generalization of a subset of unique components into multilevel classes. Generalization reduces the overall data requirements for effective modeling and increases computational speed while preserving chemical information and relationships between components. For example, instead of including a descriptor for each possible component of a formulation in the machine learning module, the number of input variables from the dataset is reduced by including descriptors for different generalized classifications of components at different generalization levels for each model. As described herein, descriptors may be a combination of measured values, chemical properties, empirical chemical relationships, application-specific qualities, etc.

[0029] In certain cases, descriptors may be used to describe the relationships between different rheological modifiers in a paint formulation. Paints are typically formulated using three characteristic viscosity measurements: cone-and-plate (ICI), stormer (KU), and Brookfield. These measurements represent fluid behavior at high, medium, and low shear, respectively. Depending on the application, the viscosity of the paint formulation is adjusted with various rheological modifiers that control viscosity within a specific shear regime or across multiple shear regimes. The efficiency of a rheological modifier depends on several factors, including its chemical structure, polarity, number of active groups, bonding behavior (e.g., interaction with other components in the formulation), and rheological modification mechanism.

[0030] These factors can be used as descriptors to place rheological modifiers (components) into a multilevel classification. The descriptors may also include additional behavioral aspects that provide a measure of convolutional compounding behavior not included by the characterization of the isolated components. The measure of convolutional compounding behavior may be a measure of a specific property of a component corresponding to a particular compounding of that component. That is, a particular component may have a first behavior or a first measured behavior when formed by a first method, while a particular component may have a second behavior or a second measured behavior when formed by a second method. For example, the behavior of a rheological modifier may be placed in a class having descriptors that define the performance of the rheological modifier as a function of its behavior in a reference shear regime, such as "ICI modification efficiency in a given stormer / KU".

[0031] In another example, descriptors may be used to describe the empirical relationships and properties of binders in paint formulations. Simply describing all binders as acrylics, or even defining them by their composition, does not accurately reflect their behavior / performance. The synthesis of building binders is as technical as it is science, due to the dynamics used in synthesis and the dynamic phenomena resulting from the synthesis procedure. An example of the role kinetics play in binders can be seen in particle morphology, where certain conditions can lead to either a homogeneous (one-phase) or heterogeneous (multiphase) form. An example of dynamics in binders can be seen in the emergence of polar groups on the surface of binder particles. In another example, acids may embed or emerge themselves as a function of pH or other local interactions on the surface. This phenomenon makes the measurement / quantification of polarity subjective to how the measurement is performed. However, a rough empirical assessment can provide information that could not be included in the initial dataset, as it would be in relation to other binders, by aggregating the tendencies of this phenomenon.

[0032] The use of empirical, measurable, computational, and categorical descriptors provides breadth and depth of information to a dataset, maximizing the usefulness of otherwise sparse datasets. This is especially true when each component is used as a column within the dataset. Not only is rich domain knowledge excluded from the dataset, but this approach dramatically increases the dimensionality of the system to be modeled.

[0033] The methods described herein overcome the challenges associated with sparse, high-dimensional datasets. Complex formulations often result in a small number of mixtures with many components, yielding minimal data. Conventional modeling techniques capture component amounts in multiple columns with many columns into which null values ​​are input. Many coating components are unique, requiring numerous formulations to distinguish the performance impact of each component, which in turn increases the amount of computational power required.

[0034] This disclosure focuses on reducing the complexity of training and testing datasets by generating one or more classes for each component of a formulation. Data dimensions are reduced by replacing component names with class specifiers and grouping them as "levels," where the first level 102-1 is the broadest class (e.g., binders, pigments, thickeners, etc.), the second level 102-2 includes subgrouping by features such as component type (e.g., polymer type, function, partial content, etc.), and is followed by an Nth level 102-N as needed to characterize the input data. In this way, descriptors can be used in addition to reduced datasets to generate more accurate machine learning modules for predicting the properties of highly formulated products.

[0035] In other words, a model generated using multilevel class definitions and characteristic descriptors (e.g., Model 100) can improve the data used by the machine learning module. A machine learning model or module that utilizes multilevel class definitions and characteristic descriptors can improve the prediction of characteristics of highly formulated products compared to using conventional models. Furthermore, the multilevel class definition models described herein provide improved formulation results with selected characteristics compared to conventional models by compressing data to reduce complexity and adding descriptors to improve accuracy. In this way, the accuracy of characteristic prediction and / or formulation prediction is improved while also reducing the amount of computational resources required to run the machine learning module.

[0036] Figure 2 is an exemplary Figure 210 illustrating an approach for generating a reduced dataset 214 from an initial dataset 212 and adding characteristic descriptors for components to the reduced dataset 214. Figure 210 includes a model similar to Model 100 shown in Figure 1, along with a method flow from the initial dataset 212 to the reduced dataset and descriptor 216. Figure 210 may illustrate how the initial dataset 212 can be reduced to the reduced dataset 214. The reduced dataset 214 can be categorized into multiple levels, as shown by Model 100 as referenced in Figure 1. As further described herein, the reduced dataset 214 may have descriptors added to multiple classes and / or levels of the model in order to generate the reduced dataset and descriptor 216.

[0037] In some examples, the initial dataset 212 may contain data associated with the components of a particular product (e.g., a compound, mixture, formulation). For example, the initial dataset 212 may contain multiple components that can be used to form a particular final product. In a particular example, the particular product is a highly formulated product such as a paint. In this example, the paint may contain multiple components, including, but not limited to, water, surfactants, neutralizers, solvents, defoamers, thickeners, binders, biocides, fillers, pigments, and / or other components based on the type of paint to be produced. In some examples, multiple options may be provided for each of the multiple components. For example, multiple binders may be provided as choices for a particular type of binder having corresponding physical and / or chemical properties. In this example, each of the multiple binder options may have a particular manufacturer having corresponding ratios of reactants for producing the binder. In this way, each of the multiple components may have corresponding multiple options that should be used to produce a final product (e.g., paint).

[0038] The initial dataset 212 may include each of the multiple component options for each of the multiple components. That is, the initial dataset 212 may include all possible component options for each of the multiple components having a corresponding weight fraction. Furthermore, the initial dataset 212 may include additional information to describe each of the multiple options. In this way, the initial dataset 212 may be a broad dataset containing information that may be useful for producing a particular final product, as well as information that may not be useful or relevant for producing a particular final product. In a particular example, the initial dataset 212 may include weight fractions for all potential components to produce a particular final product. As used herein, weight fraction may be expressed as the weight of a component relative to the total weight of the final product.

[0039] In some examples, the initial dataset 212 can be reduced to a reduced dataset 214 by categorizing the components of the initial dataset 212 based on material type. That is, each of several components may be categorized into a specific material type or Level 1 category (e.g., binder 218-1, pigment 218-2, and / or thickener 218-3, as shown by the model), for some components. In addition, several components may be categorized into subclasses of material types or Level 2 categories. For example, the Level 1 category of binder 218-1 may be further categorized into Level 2 categories such as acrylic 220-1, vinyl acrylic 220-2, and styrene acrylic 220-3. This reduction may continue up to the Nth level of subclasses based on the Level 1 class. This reduction can significantly reduce the amount of information provided by the reduced dataset 214 compared to the initial dataset 212. Descriptors may be added to the reduced dataset 214 to increase the amount and / or quality of information available for each of several components within each of the Level 1 categories. As used herein, a descriptor is a description of a specific property of a corresponding component, which may be designated to a particular level (e.g., Level 1, Level 2, Level N, etc.). In some examples, a descriptor may be based on different combinations of reactants to produce a particular component. As described herein, different ratios and / or different combinations of reactants may produce components having different properties that may affect the properties of the final product.

[0040] As shown in the model, multiple components can be categorized into general categories (e.g., high-level descriptive categories). For example, the model may include, but is not limited to, a Level 1 category containing binder 218-1, pigment 218-2, and thickener 218-3. Although three components are shown in the model, additional or fewer components may be used to generate or train machine learning modules. Multiple subcategories or Level 2 categories may be generated for each of the Level 1 categories. A Level 2 category may be a category within a Level 1 category. For example, the Level 1 category for binder 218-1 may include Level 2 categories such as acrylic 220-1, vinyl acrylic 220-2, and / or styrene acrylic 220-3. Although three Level 2 categories are shown for the Level 1 category of binder 218-1, additional or fewer Level 2 categories may be available. Furthermore, each Level 2 category may include additional Level 3 categories and / or specific components such as component 222. Component 222 may be a specific component that can be used to produce the final product. In particular, component 222 lists Acousticryl AV-1120, Avance 311, and Rhoplex 585, but embodiments herein are not limited in that respect, and additional or fewer components 222 may be provided.

[0041] In different examples, a Level 1 category may be pigment 218-2. Pigment 218-2 may include titanium dioxide 224-1 and filler 224-2 from a Level 2 category. Furthermore, filler 224-2 from a Level 2 category may include a Level 3 category containing clay 226-1, calcium carbonate 226-2, and / or aluminum silicate 226-3. In some examples, each or part of a Level 3 category may include a Level 4 category. In contrast, each of a Level 3 category may be specified in a corresponding list of components.

[0042] One or more descriptors may be added to multiple categories within each level. A descriptor may further describe or include information about the properties of the components in the corresponding category and / or how the components in the corresponding category affect the properties of the final product when used. In some examples, as further illustrated with reference to Figure 3, a descriptor may add information about different combinations of reactants used to produce the corresponding component. As used herein, a combination of components may include the ratio of the component and associated reactants and / or the formation process. This combination may include a corresponding list of component properties, as further illustrated herein.

[0043] As described herein, each component may have different ratios for each combination of reactants used to produce a particular component. For example, a binder may have different properties for different combinations of reactants for specific properties such as surface stabilization. Diagrams may be created to show how specific properties of a particular component are divided based on the combination of reactants used to form that component. This type of diagram may be created for each of multiple components. Thus, a first category may show the properties of a first binder, which is a single binder having a specific surface stabilization. In addition, the diagram may include a second category that may show the properties of a second binder, which includes a binder mixture having a specific mixing ratio. The properties of the first and second binders may be used to predict how the binders will change or influence the properties of the final product.

[0044] Figure 3 is an exemplary Figure 340 illustrating an approach to adding component characteristic descriptors to a machine learning module model. In some examples, Figure 340 may represent an exemplary portion of an ANN or similar machine learning module. In some examples, Figure 340 includes multiple nodes associated with an input layer 346, multiple nodes associated with an output layer 350, and multiple nodes associated with a hidden layer 348 between the input layer 346 and the output layer 350.

[0045] Multiple nodes can be represented by circles in the hidden layer 348 in Figure 340. As described herein, multiple nodes can function as neurons that can be represented by one or more equations or functions. In this way, the output value of a particular node among the multiple nodes can be used as an input value for a subsequent node among the multiple nodes, based on the equation or function in that particular node. In another example, the output value of a particular node can be used to determine which subsequent node to activate or use in the hidden layer 348. In this way, a particular node among the multiple nodes in the hidden layer 348 can correspond to a particular weight or bias to be applied to the input or input value. In some examples, the input layer 346 may include a first set of multiple input values ​​representing a material class 342 and a second set of multiple input values ​​representing weighted descriptors 344. The material class 342 may be a class of identified components, as described herein, and the weighted descriptors 344 may be descriptors associated with components having corresponding weighted values. In some examples, input values ​​from material class 342 and / or input values ​​from weighted descriptor 344 may be provided as input values ​​to multiple nodes or portions of multiple nodes in the hidden layer 348 to generate output values ​​in output layer 350.

[0046] As described herein, the output layer 350 may contain values ​​that can be used to produce one or more formulations containing the desired properties. In some examples, the output layer 350 may contain several different output values. For example, several output values ​​may correspond to different levels of fit or fit values ​​that may be within the tolerance range of the input values. In some examples, the output values ​​of the output layer 350 may be analyzed to determine the best fit value for the desired properties. In some examples, the output values ​​of the output layer 350 can be used to modify several input values ​​of the input layer 346. For example, the output values ​​of the output layer 350 may be used to select or deselect a particular material class 342 and / or to change the weighted values ​​of a weighted descriptor 344. In this way, the ANN shown by Figure 340 may be further adjusted to improve the accuracy of the output values ​​of the output layer 350 produced by the ANN.

[0047] Figure 4 shows an example of Method 460 for descriptor-based modeling. Method 460 may be implemented to train one or more machine learning models of a machine learning module or ANN to predict the properties of a product containing multiple components. In some examples, Method 460 may be performed by a computing device as described herein. Method 460 may be used to adjust chemical formulations and / or to adjust processes for producing a desired product. Method 460 may enable users to optimize or increase the presence of desired properties in complex or highly formulated products.

[0048] Step 462 may include providing a dataset from a predicted formulation containing two or more components. As described herein, the dataset (e.g., an initial dataset 212 as referenced in Figure 2) may include formulation data, chemical data, physical data, and / or data relating to the properties of the multiple components. In this way, the components or formulations of a product may be identified based on two or more components of the product to be formed. In some examples, a formulation can be used to identify two or more components of a product. In these examples, the dataset may include information relating to two or more components.

[0049] As described herein, a dataset for two or more components may include all weight fractions of the components for a particular product produced by the formulation. In this way, the amount of data in the dataset may be relatively large for highly formulated products. In some examples, Method 460 includes providing a historical dataset from historical formulation data. A historical dataset from historical formulation data may be data collected by producing a product from the historical formulation data and performing characterization tests on the product. In this way, the historical formulations and corresponding product characteristics can be used to determine which components affect specific characteristics of the product. For example, a portion of two or more components in the dataset may have a relatively greater influence on the gloss characteristics of the final product compared to other components.

[0050] In step 464, method 460 may include categorizing the components of a dataset into a multilevel classification in order to generate a reduced dataset. As described herein, a multilevel classification may include one or more classes placed at two or more levels. As shown in Figures 1 and 2, a multilevel classification can reduce an initial dataset by creating different categories at each of the multiple levels and assigning multiple components to one or more of the multiple levels. For example, a dataset may be reduced to a reduced dataset (e.g., reduced dataset 214) by categorizing multiple components based on material type. In this way, the dataset can be more easily managed and implemented in a machine learning module.

[0051] In an example utilizing historical data, Method 460 includes categorizing the components of the historical dataset into a multilevel classification in order to generate a reduced dataset. As described herein, the historical data may be data collected through previous experiments with multiple components and / or experiments on products resulting from the generation of multiple components. By implementing the historical dataset into a multilevel classification, the fit for predicting the properties of products produced by different formulations can be increased.

[0052] In some examples, categorizing data into one or more classes further involves determining feature selection for one or more classes based on the influence of one or more classes on the predicted properties. In some examples, feature selection may be based on values ​​indicating higher or lower importance. In this case, feature selection may be assigned to classes based on the influence or importance of the components within one or more classes associated with the predicted properties. Thus, components within a class with a higher feature importance value may have greater importance within the categorization compared to components within a class with a lower feature importance (e.g., feature importance value). In some examples, feature selection may allow some categories to be ignored or have a relatively less influence on the machine learning prediction module than others. In some examples, feature selection may allow adding or excluding material classes from the machine learning prediction module. For example, the desired property may be a particular gloss level of a paint. In this example, the gloss of a paint may be highly correlated with the pigment volume concentration (PVC) of the paint. In this way, other material classes, such as rheological modifiers, may have little effect on the gloss of the paint and may be excluded from the machine learning prediction module, while pigment volume may be added to the machine learning prediction module.

[0053] In some examples, material classes may be added to or excluded from a machine learning prediction module in response to their predicted importance to a desired property. In other examples, material classes may be added to or excluded from a machine learning prediction module in response to results associated with multiple different multilevel classification models used to train the machine learning prediction module. Thus, feature selection of material classes may be assigned based on multiple different factors, including the results of previously run machine learning prediction modules.

[0054] In other examples, categorizing data into one or more classes further involves assigning feature importance values ​​to one or more classes based on the impact of one or more classes on the predicted formulation. Similar to feature importance assigned based on the impact of predicted characteristics, one or more classes may be assigned feature importance based on their impact on the predicted formulation. In some examples, certain categories of components may not have any impact on the predicted formulation. In these examples, certain categories may be assigned a relatively low feature importance so that the computing resources required to run the machine learning prediction module may be relatively low. In another example, classes or categories that may have a greater impact on the predicted formulation may be assigned a relatively high feature importance to ensure that the machine learning prediction module prioritizes the classes that have a greater impact on the predicted formulation.

[0055] Step 466 of Method 460 may include incorporating one or more descriptors associated with components into a reduced dataset in order to generate a modified dataset. In some examples, one or more descriptors may include, but are not limited to, measured, empirical, or compositional properties for generating a modified dataset. One or more descriptors may include data describing the properties of a component. Descriptors may be assigned to or incorporated into a class of the modified dataset. In this way, descriptors may describe the properties of components to which they are assigned within a particular category. One or more descriptors may be incorporated into a particular component in the reduced dataset if the descriptor is specific to a particular component. In other examples, descriptors may identify multiple combinations for each of multiple reactants or reactant ratios for generating a component. In this way, components having different properties under different conditions may be identified by descriptors in the reduced dataset.

[0056] Categorizing the components of a dataset may further involve generating one or more descriptors that include characterization data based on the combination of reactants associated with the component. In these examples, the characterization data includes a characterization description of the component based on the ratio of reactants used to form the component. As described herein, components used to form a product may have different properties based on how the component is formed. That is, the reactants used to form the component may affect the properties of the component and thus have different effects on the properties of the product. For example, the pH of a component may affect certain properties of the product produced by the component. In this example, the formation of the component may affect the pH. Thus, the characterization data may include descriptions of different reactants used to form the component, and the resulting properties of the component for different reactants or different ratios of reactants.

[0057] Step 468 may include inputting the modified dataset into a machine learning module trained to predict the properties of products produced from a predicted formulation. The modified dataset, having descriptors as described herein, may be provided to a machine learning module run by a computing device. In some examples, the machine learning module may identify fit values ​​for several different functions, and particular functions may be selected based on complexity values ​​and fit values ​​as described herein. The machine learning module may be trained through the values ​​associated with the modified dataset. In some examples, levels, classes, and / or parts of the descriptor may be selected based on feature importance or weighted average to generate functions with fit values ​​above a threshold fit value.

[0058] In some examples, Method 460 involves inputting a modified dataset of one or more classes into a machine learning module to generate an updated, trained machine learning module for predicting product characteristics. Historical data may be added to the modified dataset provided to the machine learning module, as described herein. The machine learning module may be trained using the historical data added to the modified dataset to generate an updated, trained machine learning module. In some examples, the updated, trained machine learning module may produce a relatively higher fit compared to the previous machine learning module.

[0059] Step 470 may include Method 460 receiving predictions of product properties from a machine learning module. In these examples, receiving predictions of product properties may further include receiving a prediction of one of a group of properties, including molecular weight, density, quality, performance, and identification. In some examples, the property prediction may be a prediction of a selected property. For example, a surface stabilization category may be selected, and a prediction of the surface stabilization category for a product utilizing a predicted formulation may be calculated. In this way, a desired property may be selected, and the machine learning module may generate values ​​associated with the desired property based on a modified dataset, including a modified dataset and / or a historical dataset.

[0060] In some cases, Method 460 may proceed from step 470 to step 464 using different multilevel classifications. In these examples, Method 460 may perform steps 466, 468, and 470 utilizing different multilevel classifications. In these examples, Method 460 may include a step of comparing the results of a first multilevel classification with different multilevel classifications. In some examples, Method 460 performs steps 466, 468, and 470 over multiple additional time periods, utilizing multiple different multilevel classifications to obtain multiple results that can be compared.

[0061] In some examples, method 460 includes a step of comparing multiple results from several different multilevel classifications to determine the best-fit model for a particular desired characteristic. Based on the comparison, method 460 can proceed to step 472.

[0062] In step 472, method 460 may include adjusting the chemical formulation and / or process based on predictions of product characteristics to produce or reject a product. In some examples, adjusting the chemical formulation includes receiving output suggestions from a machine learning module to change the weight fraction of a particular component. For example, the machine learning module may generate predicted characteristic values. In this example, the machine learning module may generate output suggestions to change, add, or remove components to change the predicted characteristics to desired characteristics. In this way, the chemical formulation may be adjusted to improve the predicted characteristic values ​​of the product produced.

[0063] In another example, the process for producing a product can also be adjusted. In some examples, a machine learning module may generate output suggestions for modifying, adding, or removing specific steps in a process to produce a product that can improve desired characteristics of the product. In some examples, the specifications of the process may be modified based on suggestions provided by the machine learning module. For example, the temperature of a particular process may be changed based on suggestions from the machine learning module. Process steps can be analyzed from historical data to determine how specific steps in the process affect specific characteristics of the product. In this way, the process can be modified to improve desired characteristics of the product.

[0064] Figure 5 shows an example of a machine-readable medium 580 for descriptor-based modeling. The machine-readable medium 570 may be communicatively connected to a processor resource 582 by a communication path 584. In some examples, the communication path 584 may include wired or wireless connections that can enable communication between devices and / or between components within a single device. As used herein, the processor resource 582 may include, but is not limited to, a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a metal-programmable cell array (MPCA), a semiconductor-based microprocessor, or other combinations of circuitry and / or logic for coordinating the execution of instructions 586, 588, 590, 592, 594. In a specific example, processor resource 582, when executed, utilizes a non-temporary computer-readable medium that stores instructions 586, 588, 590, 592, and 594 that cause the processor resource 582 to perform its corresponding function.

[0065] The machine-readable medium 580 may be an electronic, magnetic, optical, or other physical storage device that stores executable instructions. Therefore, a non-transitory machine-readable medium (MRM) (e.g., machine-readable medium 580) may include, for example, random-access memory (RAM), read-only memory (ROM), electrically erasable programmable ROM (EEPROM), storage drives, optical discs, etc. The machine-readable medium 580 may be located within a controller and / or computing device. In this example, executable instructions 586, 588, 590, 592, and 594 may be "installed" on the device. Additionally, and / or alternatively, the machine-readable medium 580 may be, for example, a portable, external, or remote storage medium that enables a computing system to download instructions 586, 588, 590, 592, 594 from a portable / external / remote storage medium. In this context, the executable instructions may be part of an “installation package”.

[0066] The machine-readable medium 580 includes instructions 586 for determining two or more components of a formulation to produce a product having desired properties. In some examples, a specific product having desired properties can be selected. The specific product may contain a general formulation or have a set of components that are typically used to produce the specific product. The desired properties may be properties of a specific product that can be affected or modified based on the formulation. For example, properties such as the drying time of a paint product may be modified based on the components used to produce the paint product. In this example, a general-purpose paint product may have multiple components that can be used. However, specific properties can be modified by changing the weight fractions and / or ratios of the multiple components. In this way, components can be produced by utilizing product selection, while components that change the desired properties can be identified by utilizing desired property selection.

[0067] The machine-readable medium 580 includes instructions 588 for generating a multilevel classification that includes one or more classes arranged at one or more levels for two or more components. In some examples, the multilevel classification includes one or more respective descriptors for each of the two or more components. In these examples, one or more descriptors include one or more properties of the product associated with corresponding combinations of reactants for producing the two or more components. In some examples, one or more descriptors describe corresponding properties of the product associated with different ratios of the combination of reactants. As described herein, the multilevel classification can be a model of multiple possible components that may be used to produce a product.

[0068] The machine-readable medium 580 may include instructions for generating weighted values ​​for two or more components based on their effect on modifying an identified characteristic. In these examples, components with a relatively greater effect on modifying the identified characteristic are assigned higher weights, and components with a relatively smaller effect on modifying the identified characteristic are assigned lower weights. As described herein, weighted values ​​may be assigned to components and / or component descriptors based on the effect associated with the component or class of components having the desired characteristic. To modify the characteristics of a product to have the desired characteristic, components that have a greater effect on modifying the characteristic may be given higher weights or priority.

[0069] The machine-readable medium 580 includes instructions 590 for selecting one or more parts of a multilevel classification descriptor for two or more components based on the characteristics of the product. In some examples, the descriptors may include multiple combinations or cases that identify different characteristics of a component based on the composition of the components. In this way, a particular component may have different characteristics based on how that particular component is formed. In this example, a first combination for a particular component may have a relatively large impact on the production of a product having the desired characteristics, while a second combination for a particular component may have a relatively small or less impact on the production of a product having the desired characteristics. In this example, the first combination descriptor may be selected over the second combination descriptor.

[0070] The machine-readable medium 580 may include instructions for selecting a portion of one or more descriptors of a multilevel classification based on weighted values ​​assigned to one or more descriptors. As described herein, different descriptors may have different levels of importance or effect on the production of a product having desired characteristics. In this way, specific weight values ​​can be assigned to descriptors, and the portion of descriptors to be selected may be based on these weight values.

[0071] The machine-readable medium 580 includes instructions 592 for inputting a selected portion of a multilevel classification into a machine learning module trained to predict the properties of a product produced using a formulation. As described herein, a selected portion of a multilevel classification may have a relatively greater impact on the production of a product having the desired properties compared to other portions of the multilevel classification. The selected portion may be provided to the machine learning module to predict the properties of a product produced with different formulations and / or under different processes.

[0072] The machine-readable medium 580 includes instructions 594 for receiving a predicted formulation, which includes the ratio of two or more components, in order to produce a product having desired properties. In some examples, the machine learning module may generate weight fractions of the components of the formulation to produce a product having desired properties. In these examples, multiple formulations may be provided, each having a range of properties within a desired property range. For example, a selected portion of a multilevel classification may be associated with a first portion of the component, and an unselected portion of the multilevel classification may be associated with a second portion of the component. In this example, the weighted fraction of the first portion of the component may be more specific than that of the second portion, since the first portion has a greater influence on producing the desired properties of the product.

[0073] Figure 6 shows an example of a device 601 for descriptor-based modeling. In some examples, device 601 is a computing device that includes a processor resource 682 and a machine-readable medium 680 for storing instructions 603, 605, 607, 609, 611, 613, 615, 617, and 619 that are executed by the processor resource 682 to perform a particular function. Figure 6 shows how the computing device may execute instructions to perform the functions described herein.

[0074] Device 601 includes instructions 603 stored in a machine-readable medium 680, which are executed by a processor resource 682 to provide a dataset from predicted formulations of products containing two or more components. The dataset from the predicted formulations may be an initial dataset (e.g., initial dataset 212). As described herein, the initial dataset may be a dataset containing the weight fractions of all components of several different formulations for different final products. As described herein, the initial dataset may include historical data from experiments or tests to determine the properties of different products produced by different formulations. Thus, the characteristic values ​​associated with the historical data may be part of the initial dataset.

[0075] Device 601 includes instructions 605 stored in a machine-readable medium 680, which are executed by a processor resource 682 to determine the respective category for each of two or more components based on the material type of the components. As described herein, a dataset or initial dataset may be reduced by categorizing the components based on the material type of the components. In some examples, the categorization of a dataset may be used to generate a reduced dataset (e.g., reduced dataset 214 as referenced in Figure 1).

[0076] Device 601 includes instructions 607 stored in a machine-readable medium 680, which are executed by a processor resource 682 to generate a multilevel classification of two or more components based on material type categories. In some examples, the multilevel classification of categories in a reduced dataset may be implemented as a model (e.g., model 100 as referenced in Figure 1). As described herein, each of the multiple levels may be a different category based on the material type of the previous category. In some examples, the multiple levels may be based on a classification of a particular material type.

[0077] Device 601 may include instructions for generating multiple categories corresponding to multiple material types for a multilevel classification. In these examples, device 601 may include instructions for assigning different components to corresponding categories among the multiple categories based on the material types of different components. Furthermore, device 601 may include instructions for incorporating multiple descriptors into the multilevel classification for different components based on the set of properties associated with those different components. In this way, the multilevel categorization or model can be updated to utilize additional components not present in the previous model.

[0078] A multilevel categorization model may include multiple classes organized into multiple levels. These classes can be specified by one or more descriptors. Descriptors can be used to incorporate empirical data, measurement data, and / or compositional data into the model. In some examples, empirical data may be incorporated into the model via material type descriptors to describe material types associated with components or classes of components, formation type descriptors to describe formations associated with components or classes of components, and / or property type descriptors to describe specific properties associated with components or classes of components.

[0079] Device 601 includes instructions 609 stored in a machine-readable medium 680, which are executed by processor resource 682 to incorporate a material type descriptor into a multilevel classification of two or more components. The material type descriptor may be incorporated to further describe the material properties of components within a particular class or level. The material type descriptor may include chemical composition, particle size, molecular mass, density, hardness, pH, melting point, boiling point, and / or other properties of the components. In some examples, the material type descriptor may be limited to features or properties that affect a particular product or type of product to be formed.

[0080] Device 601 includes instructions 611 stored in a machine-readable medium 680, which are executed by a processor resource 682 to incorporate a formation type descriptor into a multilevel classification for two or more components. As used herein, a formation type descriptor can be a description of properties associated with components formed by different processes and / or different reactants. As described herein, a particular component may have different component properties based on how that particular component was formed. As described herein, component properties may affect the properties of the final product. In this way, a formation type descriptor can be used to describe the component properties of a component under different formation conditions.

[0081] Device 601 includes instructions 613 stored in a machine-readable medium 680, which are executed by processor resource 682 to incorporate a characteristic type descriptor into a multilevel classification of two or more components. A characteristic type descriptor can describe how components affect specific properties of a final product when used to form the final product. For example, a characteristic type descriptor can identify or describe how components affect the gloss of a paint product. In other examples, a characteristic type descriptor can identify different ratios of components in the final product and how specific properties are affected by different ratios of components in the final product.

[0082] Device 601 includes instructions 615 stored in a machine-readable medium 680, which are executed by processor resource 682 to select a set of descriptors from material type descriptors, formation type descriptors, and characteristic type descriptors based on selected output characteristics of a product. Device 601 may include instructions for ranking the material type descriptors, formation type descriptors, and characteristic type descriptors based on the selected output characteristics. As described herein, weighted values ​​or priority levels may be assigned to multiple descriptors. Weighted values ​​may be used to rank different descriptors. In some examples, sets of descriptors may be assigned weighted values ​​and / or ranked based on selected output characteristics or desired characteristics of a product. That is, the selected output characteristics may be a specific desired product value or desired characteristic range for a product, and the set of descriptors may be selected based on how the set of descriptors affects the desired characteristic.

[0083] In some examples, sets of descriptors are positioned at different levels of a multilevel classification. In some examples, a first descriptor assigned to level 1 may relate to a selected characteristic of a product, and a second descriptor assigned to level 3 may also relate to a selected characteristic; thus, sets of descriptors can be assigned to different levels of a multilevel classification or model.

[0084] In these examples, device 601 may include instructions for incorporating numerical data descriptors for two or more components based on a set of descriptors. The numerical data descriptors may be numerical characteristic data based on historical data and / or tests performed on the components or products produced by the components. For example, the numerical data descriptors may be pH values, hardness values, or pigment values, among other numerical values ​​or ranges of values ​​that may be used to describe the properties or characteristics of the components.

[0085] Device 601 may include instructions for generating a set of ratios for combining two or more components to produce a product having predicted properties. For example, the set of ratios may be weight fractions for a formulation to produce a product having predicted properties. In some examples, the weight fractions or set of ratios may be used for formulations to produce components and for formulations to use components to produce a product.

[0086] Device 601 may include instructions for incorporating performance type descriptors into a multilevel classification of two or more components to identify different resulting performances for different combinations of reactants forming two or more components, based on performance data of different products that utilize two or more components and components within the same category. A performance type descriptor may be performance data for a particular combination of reactants and / or a particular component produced by a particular process. As described herein, a performance type descriptor may be a description or category specified based on performance data. For example, a performance type may be specified for components that fall within a specific range of values ​​for a particular performance test.

[0087] Device 601 includes instructions 617 stored in a machine-readable medium 680, which are executed by processor resource 682 to input a multilevel classification having a set of descriptors into a machine learning module trained to predict the properties of a product produced from a predicted formulation. As described herein, the multilevel classification and / or model may be used to train the machine learning module to improve its accuracy in predicting product properties. In these examples, the predicted properties may be compared to selected output properties.

[0088] Device 601 includes instructions 619 stored in a machine-readable medium 680, which are executed by a processor resource 682 to adjust a predictive formula based on predicted characteristics and selected output characteristics. As described herein, predicted characteristics can be compared with selected output characteristics, and the predictive formula may be modified if the predicted characteristics are not within a certain threshold of the selected output characteristics. In this way, the predictive formula can be adjusted to bring the predicted characteristics closer to or within the range of the selected output characteristics.

[0089] While specific embodiments have been described above, these embodiments are not intended to limit the scope of the disclosure, even if only a single embodiment describes a particular feature. Examples of features provided in this disclosure are intended to be illustrative, not restrictive, unless otherwise specified. The above description is intended to cover such alternatives, modifications, and equivalents, as will be obvious to those skilled in the art who are interested in this disclosure.

[0090] The scope of this disclosure includes any features or combinations of features disclosed herein (expressly or implicitly), or any generalization thereof, whether or not they mitigate any or all of the issues addressed herein. While various advantages of this disclosure have been described herein, embodiments may provide some or all of such advantages, or none of them, or other advantages.

[0091] In the details described herein, some features are grouped together in a single embodiment for the purpose of streamlining the disclosure. The method of this disclosure is not to be construed as reflecting an intention that the disclosed embodiments of this disclosure must use more features than are explicitly enumerated in each claim. Rather, as reflected in the following claims, the subject matter of the invention lies in fewer features than all the features of a single disclosed embodiment combined. Accordingly, the following claims are incorporated herein into a form for carrying out the invention, and each claim stands independently as a separate embodiment.

Claims

1. It is a method, To provide a dataset from predicted formulations containing two or more components, To generate a reduced dataset, the components of the dataset are categorized into a multilevel classification, wherein the multilevel classification includes one or more classes located at one or more levels, and at least one class includes two or more levels. To generate a modified dataset, one or more descriptors associated with the component, wherein the one or more descriptors include the compositional characteristics of the component, are incorporated into the reduced dataset. The modified dataset is input into a machine learning module trained to predict the characteristics of the product generated from the predicted formulation. Receiving the prediction of the characteristics of the product from the machine learning module, A method comprising adjusting chemical formulations and / or processes based on the prediction of the characteristics of the said product to produce or reject the said product.

2. The aforementioned machine learning module, To provide a historical dataset from historical formulation data, To generate the reduced dataset, the components of the historical dataset are categorized into the multilevel classification, The one or more descriptors, each containing at least one of measured characteristics, empirical characteristics, calculated parameters, or compositional characteristics, are incorporated into the reduced dataset to generate the modified dataset. The method according to claim 1, further comprising training the machine learning module by inputting the modified datasets of one or more classes into the machine learning module in order to generate an updated trained machine learning module for predicting the characteristics of the product.

3. The method according to claim 1, further comprising categorizing the data into one or more classes, and assigning feature importance to the one or more classes based on the influence of the one or more classes on the predicted characteristics.

4. The method according to claim 1, wherein receiving the prediction of the characteristics of the product includes receiving the prediction of one of the characteristics, which include molecular weight, density, quality, performance, and identification.

5. The method according to claim 1, further comprising categorizing the data into one or more classes and assigning feature importance to the one or more classes based on the influence of the one or more classes on the predicted formulation.

6. The method according to claim 1, further comprising categorizing the components of the dataset to generate one or more descriptors that include characteristic transformation data based on combinations of reactants associated with the components.

7. The method according to claim 6, wherein the characteristic conversion data includes a characteristic description of the component based on the ratio of the reactants used to form the component.

8. A machine-readable medium for storing machine-readable instructions, wherein when the machine-readable instructions are executed by the device's processor resources, the processor is provided with the following information: To produce a product with desired properties, determine two or more components of the formulation. A multilevel classification comprising one or more classes arranged at one or more levels for the two or more components, wherein the multilevel classification generates a multilevel classification comprising one or more descriptors for each of the two or more components, Based on the characteristics of the product, select the portion of each of the two or more descriptors of the multilevel classification for the two or more components. The selected portion of the multilevel classification is input into a machine learning module trained to predict the characteristics of a product produced using the formulation. A machine-readable medium for receiving a predicted formulation including the ratio of the two or more components in order to produce the product having the desired characteristics.

9. The machine-readable medium according to claim 8, wherein the one or more descriptors include one or more characteristics of a product associated with corresponding combinations of reactants for producing the two or more components.

10. The machine-readable medium according to claim 9, wherein one or more descriptors describe corresponding properties of a product associated with different ratios of the combination of reactants.

11. The machine-readable medium according to claim 8, comprising instructions for generating weighted values ​​for two or more components based on the effect of modifying the desired characteristic, wherein components with a relatively larger effect of modifying the identified characteristic are assigned a larger weight value, and components with a relatively smaller effect of modifying the identified characteristic are assigned a smaller weight value.

12. The machine-readable medium according to claim 8, wherein the descriptor includes at least one of a sequential descriptor, a sequence descriptor, a binary descriptor, or a categorical descriptor.

13. The machine-readable medium according to claim 11, comprising instructions for selecting the portion of the one or more descriptors of the multilevel classification based on the weighted values ​​assigned to the one or more descriptors.

14. It is a device, Processor resources and A non-temporary memory resource for storing machine-readable instructions, wherein when the machine-readable instructions are executed, the processor resource is used to store them. We request a dataset from the predicted formulations of products containing two or more ingredients. Based on the material type of the aforementioned component, the respective category for each of the two or more components is determined. Based on the material type category, a multilevel classification is generated for the two or more components, wherein the multilevel classification includes one or more classes arranged at one or more levels, and at least one class includes two or more levels. In order to identify the material properties of the two or more components at the corresponding levels of the multilevel classification, a material type descriptor is incorporated into the multilevel classification for the two or more components. In order to identify combinations of different reactants for forming the two or more components, a formation type descriptor is incorporated into the multilevel classification for the two or more components. In order to identify the different resulting properties for the combination of the different reactants that form the two or more components, a property type descriptor is incorporated into the multilevel classification for the two or more components. Based on the selected output characteristics of the product, a set of descriptors is selected from the material type descriptor, the formation type descriptor, and the characteristic type descriptor. The multilevel classification having the set of descriptors is input to a machine learning module trained to predict the characteristics of the product produced from the predicted formulation. A device that adjusts the predicted formulation based on the predicted characteristics and the selected output characteristics.

15. The device according to claim 14, wherein the processor is for ranking the material type descriptor, the formation type descriptor, and the characteristic type descriptor based on the selected output characteristics.

16. The device according to claim 14, wherein the processor is for generating a set of ratios for combining the two or more components in order to produce the product having the predicted characteristics.

17. The device according to claim 14, wherein the processor incorporates performance type descriptors into the multilevel classification of the two or more components in order to identify different resulting performances for different combinations of reactants forming the two or more components, based on performance data for different products that utilize components in the same category as the two or more components.

18. The aforementioned processor, Multiple categories corresponding to multiple material types for the aforementioned multilevel classification are generated, Based on the material types of different components, the different components are assigned to the corresponding category among the multiple categories. The device according to claim 14, for incorporating a plurality of descriptors into the multilevel classification for the different components based on a set of properties associated with the different components.

19. The device according to claim 14, wherein the processor is for incorporating numerical data descriptors for two or more components based on the set of descriptors.

20. The device according to claim 14, wherein the set of descriptors is positioned at different levels of the multilevel classification.

Images