Systems and methods for generating recipes for thermoplastic compounds - Patents.com

Abstract

A processor system and method (100) for generating a recipe for a thermoplastic compound is provided, where the recipe defines a set of ingredients and the relative contributions of the ingredients for producing a thermoplastic compound. The ingredients may include additives that are added to a base polymer. The recipe can be generated by training (110) a machine learning enabled model on compound data (20) of existing (historical) compounds to predict values ​​of compound material properties from an input recipe, providing (120) candidate recipes, selecting (140) the best recipe based on a scoring function, outputting (150) the selected recipe, for example via a display, to enable a sample of the compound to be produced (200) and measured (210), receiving (160) the measurement data of the sample, determining deviations to a target specification, and determining (170) whether the recipe is acceptable.

Description

Detailed Description of the Invention

[0001] [Field of the Invention] The present invention relates to a computer-implemented method and a processor system for generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound comprising a base polymer, the recipe defining a set of ingredients and the relative contributions of the ingredients for producing the compound. The present invention further relates to a computer readable medium comprising a computer program for performing the computer-implemented method, a method further comprising the step of generating a sample of the thermoplastic compound, a method for producing a thermoplastic compound using the recipe, and a thermoplastic compound obtainable by the method.

[0002] [Background of the invention] A compound, such as a thermoplastic polymer mixed with other ingredients, such as glass fibers, stabilizers, flame retardants, colorants, etc., can be produced according to a target specification, for example, at the request of a customer. For example, a customer can request that the compound have a certain color, which can be expressed, for example, as a CIELAB value, which can be measured on a target plate produced from the compound under a specific light source. A recipe for the production of the compound can be determined so that the compound meets the target specification completely or at least to an acceptable extent. Such a recipe can define a set of ingredients used in the production of the compound, such as, for example, a colorant to produce a colored polymer, and the relative contribution of the ingredients to each other, such as the mass fraction of the ingredients.

[0003] It is known that such recipes are determined empirically. For example, in the case of colored compounds, the color expert can search his (physical) database for an existing color plate that visually most closely matches the customer's color target specification. The recipe of the existing color plate, i.e. the compound from which the plate is made, may be known. Empirically, using smart theory, the color expert can generate a colorant recipe using the recipe of the best matching color plate as a starting point. This colorant recipe can then be used to make test plates, for example by compounding (extrusion) and injection molding, The color values ​​can be measured under the same light source as the customer target plate. If the color values ​​are too far off, the development cycle can be repeated until the deviation of the color values ​​from the target specification is acceptable.

[0004] However, the target specification often covers not only one material property, e.g. color, but also several material properties of the compound, e.g. mechanical properties, UV stability performance, heat aging performance, viscosity, CTI (Comparative Tracking Index) value, flammability, etc., which are also referred to elsewhere as "compound material properties". For example, in the case of a colored compound, if the color target specification is met, several mechanical tests can be performed on a sample of the colored compound, and if the mechanical properties of the colored compound deviate too much from the target values ​​of the mechanical properties, the development process can be repeated until both the color value and the mechanical properties of the compound comply with the target specifications. It can often happen that the mechanical properties are insufficient when the color is optimized. For example, in the case of a colored compound based on a high temperature polymer, hard inorganic pigments that may be added to the base polymer to obtain the desired color may cause the glass fibers in the compound to decompose during compounding, which may result in a decrease in mechanical properties. Also, in general, when optimizing the recipe of a compound to meet the target specification with respect to one material property, other material properties may also be affected by changes in the recipe during optimization, and therefore may need to be checked for compatibility.

[0005] The development process described above can be time consuming, expensive, and have longer throughput times, while target specifications may not always be met. For example, typical difficulties that may be encountered during the color matching process are the selection of the best existing recipe to use as a starting point, and the estimation of the effect of changing colorants on the color of the compound. Similar difficulties may be encountered when optimizing other material properties, such as mechanical properties.

[0006] Therefore, it may be desirable to be able to generate recipes for the manufacture of compounds that take less time and / or have shorter throughput times.

[0007] [Summary of the invention] According to a first aspect of the invention there is provided a computer implemented method for generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound comprising a base polymer, the recipe defining a set of ingredients and relative contributions of the ingredients for producing the compound, the set of ingredients comprising one or more additives to be added to the base polymer; acquiring compound data for compounds produced using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties, and b) a recipe for each compound; training a machine learning enabled model on the compound data to predict values ​​of one or more compound material properties from an input recipe, thereby obtaining a predictive model; To approximate a target value of one or more compound material properties. i) generating candidate recipes for compounds and predicting values ​​of one or more compound material properties for each candidate recipe using a predictive model; ii) selecting a recipe from the candidate recipes, the recipe comprising evaluating a scoring function, the scoring function configured to reward correspondence between the predicted values ​​and target values ​​of one or more compound material properties of the candidate recipes; iii) outputting the selected recipe to enable production of a sample of the compound and measurement of one or more material properties of the sample; iv) receiving sample measurements of one or more material properties of the sample and comparing the sample measurements to target values ​​to determine deviations from the target values; and v) outputting the selected recipe as the compound recipe if the deviation meets the acceptability criteria (170); generating a recipe for the compound by A computer-implemented method is provided, comprising:

[0008] According to a further aspect of the present invention, there is provided a temporary or non-transitory data computer readable medium, the computer readable medium comprising data representing a computer program, the computer program comprising instructions for causing a processor system to perform any one of the computer-implemented methods described herein.

[0009] According to a further aspect of the invention, there is provided a processor system for generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound comprising a base polymer, the recipe defining a set of ingredients and relative contributions of the ingredients to a recipe for producing the compound, the set of ingredients including one or more additives to be added to the base polymer; a data storage interface configured to access compound data for compounds generated using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties, and b) a recipe for each compound; training a machine learning enabled model on the compound data to predict values ​​of one or more compound material properties from an input recipe, thereby obtaining a predictive model; To approximate a target value of one or more compound material properties. i) generating candidate recipes for compounds and predicting values ​​of one or more compound material properties for each candidate recipe using a predictive model; ii) selecting a recipe from the candidate recipes, the recipe comprising evaluating a scoring function, the scoring function configured to reward correspondence between the predicted values ​​and target values ​​of one or more compound material properties of the candidate recipes; iii) outputting the selected recipe to enable production of a sample of the compound and measurement of one or more material properties of the sample; iv) receiving sample measurements of one or more material properties of the sample and comparing the sample measurements to target values ​​to determine deviations from the target values; and v) outputting the selected recipe as the recipe for the compound if the deviation meets the acceptability criteria. To generate compound recipes, The processing subsystem that is configured A processor system is provided, comprising:

[0010] The above means may include a step of generating a recipe for the production of a thermoplastic compound. The recipe may define a set of components of the compound that can be used in the production of the compound. Such components may include additives added to the base polymer, for example colorants such as dyes and pigments ("colorants"), glass fibers, stabilizers, flame retardants, processing aids, flow improvers, etc. In some examples, the recipe may define all components of the compound including the base polymer, while in other examples, the recipe may only define additives added to the base polymer or the base polymer and other base constituents. An example of the latter type of recipe is a recipe that defines which colorants are added to the base polymer to obtain a colored polymer. The recipe may further define the relative contribution of the components, for example as mass fractions. It is noted that recipes of the type defined in this paragraph are known per se. For the avoidance of doubt, it is further noted that the set of components may include only a single component, in that the recipe may identify one component to be added and the amount of the component to be added. In other words, the set may be a singleton set.

[0011] The above means may further include obtaining compound data for compounds previously produced using known recipes. Such compound data may relate to existing compounds, i.e., compounds previously produced, e.g., by mass production or as individual samples, and may therefore be referred to elsewhere as "historical" compound data. From these existing compounds, one or more compound material properties may have been measured. With continued reference to the example given in the Background section, color values ​​of the compounds may have been measured, e.g., using test plates produced from the compounds. Furthermore, recipes for such existing compounds may be known, e.g., by compounds produced internally using that recipe or recipes obtained in other ways. Such compound data may be, for example, either a physical database, a digital database, or a combination of both, and may be retrieved from an existing database that was built over time during the process of generating a recipe for the compound, e.g., empirically producing samples of the compound and measuring one or more compound material properties of the samples. In some examples, providing compound data may include digitizing data previously existing only in physical form, or structuring existing digital data, such that both the measured material properties and the recipe for the compound can be accessed in conjunction. In other examples, providing such compound data may include accessing existing samples of the compounds, for example as test plates, and measuring such existing compounds de novo.

[0012] The above means may further include providing a machine-learnable model and training the machine-learnable model on the compound data. In particular, the machine-learnable model may be trained on known recipes and material properties of compounds produced using such known recipes. By training, the machine-learnable model may be trained to predict values ​​of one or more compound material properties from recipes provided as input to the machine-learnable model. For example, the machine-learnable model may include a neural network that may receive as input an existing recipe in computer-readable form and may be trained using backpropagation to predict compound material properties of compounds produced using the recipe. In another example, the machine-learnable model may include a Gaussian process whose covariance function may be selected based on the compound data. As a result of such training, a machine-learnable model may be obtained, which is also referred to below simply as a "trained model" or a "predictive model", the latter referring to a model trained to predict one or more compound material properties based on a recipe provided as input.

[0013] The above method may further include a step of generating a recipe for the compound given a target specification. The target specification may be given, for example, in the form of a target value of one or more compound material properties. Such a target value may be a numerical value that may represent a desired material property, such as, for example, defining a desired color value in CIELAB units or a desired tensile strength in Pa units. In other examples, at least some target values ​​may represent a minimum value, in that a better value, for example a higher tensile strength, may be desired. Thus, the target value may represent an upper or lower boundary above which the compound may be considered unacceptable.

[0014] Given the predictive model and the target specification of the compound, a recipe for the compound can be generated as follows: Several candidate recipes can be generated for the compound. As explained elsewhere, such candidate recipes can be generated, for example, by starting from an initial recipe and modifying the initial recipe to better match the target specification. The predictive model can be used to predict the value of one or more compound material properties of a recipe, given a recipe (either the initial recipe or the modified recipe), so that it can be evaluated whether the modification brings about an improvement or not. Furthermore, a scoring function can be provided that can quantify the suitability of this recipe for meeting the target specification. For this purpose, the scoring function can reward a correspondence between the predicted and target values ​​of one or more compound material properties of the recipe. For example, if the target material property includes a color value, this target color value can be compared to the predicted color value of the recipe to determine the suitability of this recipe. That is, a large deviation to the target material property can indicate that the recipe is less suitable or may be inappropriate, and a small deviation or no deviation to the target material property can indicate that the recipe is (more) suitable. In some examples, the candidate recipes may be generated by using a scoring function, e.g., by improving a randomly generated initial recipe, while in other examples, the candidate recipes may be generated in a different manner, e.g., by a user manually specifying the candidate recipes.

[0015] From the set of candidate recipes, a recipe can then be selected using a scoring function, which candidate recipe can represent the "best" recipe according to the scoring function. To validate the recipe, the recipe can be output, for example in a human readable format, to allow a sample of the compound to be made using the recipe. Such a sample of the compound can be made using any known technique, including but not limited to those described herein. Once a sample of the compound is made, one or more compound material properties of the sample can be measured, for example, under the same or similar conditions under which the target specifications of the compound are defined, and the resulting sample measurements can be input again into the processor system or computer-implemented method. The degree of deviation from the target value may be predicted in advance, but the actual measurement may or may not confirm this prediction. In particular, the deviation between the actual measurement value and the target value can be determined, and if the deviation is acceptable, the recipe can be selected as the recipe for the compound. Such acceptability can be evaluated, for example, by computer-readable acceptability criteria, which can define an acceptable range and can be evaluated by the processor system and method. In another example, the acceptability criteria can be user criteria, where the user accepts or rejects the recipe based on the deviation from the target value.

[0016] The above means may have the effect that a machine learning-enabled model is provided that can be trained on past data so that it can predict the material properties of a compound based on its recipe. Often, a large amount of such past data is available or can be made available by integrating data from various existing physical and digital databases, so that the machine learning-enabled model can learn the relationships and patterns in such data to predict the material properties with sufficient accuracy. Thus, such selected recipes may represent a better starting point than those determined empirically, since the user may have a limited ability to recognize relationships and patterns in the data. In particular, for example, recipes that may include ingredients selected from a significant number of possible ingredients (meaning that the recipe may be a large-dimensional N-dimensional vector), the data space may be very large, so it may be difficult for the user to select a starting recipe for further optimization. However, it is recognized that the selected recipe may still be insufficient, for example due to the large dimensionality of the above-mentioned data space, which may cause the predictive model to lack accuracy in some parts of this data space. Thus, a user can have a recipe output by the processor system and method, use the recipe to manufacture a sample, measure its material properties, and feed the resulting measurements back to the processor system and method to enable the processor system and method to determine whether the compound satisfactorily meets the target specifications.

[0017] Because the process of generating recipes for manufacturing compounds is now at least partially automated by the above means, in particular by providing and using machine learning models trained on historical compound data, fewer samples may be sufficient to be manufactured in that the first samples may already provide a better match to the target specifications than previously possible when relying solely on empirical selection of a starting recipe. Advantageously, compound recipes may be generated in a less time-consuming manner and / or may require shorter throughput times.

[0018] If the deviation does not meet the acceptability criteria, the sample measurements and sample recipes can be used to retrain or update the predictive model, and steps i)-v) can be repeated using the predictive model. If the deviation of the manufactured sample is large, especially if this deviation is larger than originally predicted, this may indicate that the accuracy of the predictive model may be insufficient in that the predictive model may be insufficient to accurately capture the relationships in a particular portion of the data space. This may occur, for example, when the amount of training data is small relative to the size of the data space, e.g., the dimensions of the recipe, or when the training data is unevenly distributed across the data space, e.g., when the past compound data has only a few or no compounds with a particular component. To address this inadequacy, the predictive model can be retrained, in that the recipe and measurements of the sample compounds can be used to update the parameters of the predictive model. Such retraining can not only improve the accuracy of the predictive model when it is used to regenerate the recipe of the next compound, but also when the compound recipe is further optimized. That is, when the deviation to the target specification may be unacceptable, the recipe can be further optimized, which may include performing steps i)-v) again using the retrained predictive model. Here, previously selected recipes for a compound can be reused, for example, when generating candidate recipes, but can also be discarded and a new candidate recipe can be generated using the retrained predictive model. The recipe selection process can become more accurate because the predictive model can be more accurate in parts of the data space that are more likely to be relevant to a particular compound. Advantageously, the accuracy of the predictive model can be improved over time, not only between compounds, but also during recipe generation for a particular compound.

[0019] It should be noted that some predictive models can be updated to provide more accurate predictions without the need for retraining. For example, in the case of Gaussian process and similar predictive models that use compound data during inference and thus effectively form part of the predictive model, sample measurements and sample recipes can be added to the compound data as new data points. Such updates can achieve the same or similar effect as retraining, but without the computational effort typically required to retrain a predictive model. In some embodiments, adding data points and retraining can be alternated in that the predictive model can be retrained on the compound data for every N new data points.

[0020] The following may define optional aspects of the present invention, which may represent steps of a computer-implemented method, and which may represent processing subsystems of a processor system configured to perform the respective steps.

[0021] Optionally, the predictive model includes a neural network, and the neural network is retrained. Such retraining may include updating parameters of the neural network, such as weights or bias values.

[0022] Optionally, the predictive model includes a Gaussian process, the Gaussian process uses compound data for inference, and the Gaussian process is updated by adding sample measurements and sample recipes to the compound data. Thus, to update the Gaussian process, the sample measurements and sample recipes can be added to the input data of the Gaussian process.

[0023] Optionally, generating candidate recipes for the compound includes providing a set of random recipes and generating candidate recipes based on the set of random recipes. The candidate recipes can be generated in a number of ways, for example, by providing a set of initial recipes as "seeds" and using predictive models and scoring functions to improve the initial recipes. Interestingly, it has been found that randomly generated recipes can represent a sufficiently suitable starting point, which may be preferred since such recipes can be easily generated, for example, computationally, instead of requiring user effort. That is, due to their random nature, such initial recipes can cover the data space of recipes, i.e., the data space constituted by all possible combinations and amounts of ingredients. This may be particularly important, since completely different recipes (e.g., data points in completely different parts of the data space, e.g., representing recipes using completely different ingredients and / or completely different amounts of ingredients) may generate similar compounds. For example, when using only one recipe as a starting point, e.g., a user-selected recipe, the optimization process may suffer from a "local minimum" problem in that the optimization process may result in recipes in the vicinity of the selected starting point (the term "neighborhood" refers to the vicinity in the N-dimensional data space of recipes), but is unlikely to reach a completely different part of the data space that may still provide a suitable, and perhaps globally better, recipe for the compound ("globally" refers to considering possible combinations of ingredients). Thus, by starting from a random set of recipes, the problem of local optima in the data space that result in the selection of a globally suboptimal recipe can be reduced.

[0024] Optionally, generating candidate recipes for the compound includes iteratively modifying a set of random recipes using a genetic algorithm to obtain an improved score by the scoring function. Although the random recipes may fully cover the data space of recipes, such randomly generated recipes may represent only a small number of data points of the entire data space, so that the randomly generated recipes are unlikely to represent the optimal recipe for the compound. This may be addressed by increasing the amount of randomly generated recipes, or in extreme cases, by considering all possible recipes, but this may be infeasible as the data space is too large. For example, if 50 ingredients are available, the data space may be 50-dimensional, which makes an exhaustive evaluation of the recipes infeasible. Rather, the set of randomly generated recipes may be used as a starting point, e.g., a seed, and further improved by modifying individual recipes. The scoring function may be evaluated to evaluate whether the modification has resulted in an improvement. The modification may be iterative, e.g., by modifying an initial randomly selected recipe and modifying the resulting recipe again. Several such iterations may be performed. During each iteration, the recipes may be evaluated using the predictive model and the scoring function. To efficiently structure this search for optimal recipes, genetic algorithms can be used that can use evolutionary principles to iteratively optimize the highest scoring recipes while discarding the lowest scoring recipes. Thus, an efficient evolutionary local search for an optimal solution can be performed starting from each randomly selected data point in the data space. In this way, the data space of recipes can be better covered in that there is a higher probability of finding a globally optimal solution, i.e., recipe.

[0025] Optionally, the step of iteratively varying the set of random recipes comprises: Randomly varying the relative contributions of ingredients in a recipe; Mixing two or more recipes; Steps for randomly omitting ingredients from recipes; Steps to randomly add ingredients to a recipe; modifying the recipe in a direction selected based on a previous modification direction in a previous iteration of the genetic algorithm; and Using gradient descent techniques to modify the recipe towards a local minimum in the scoring function Includes at least one of the following.

[0026] Some of the above modifications may represent "crossover" and "mutation" in a genetic algorithm, and thus allow the top scoring recipe to be further optimized. In some embodiments, the modifications to the recipe may be (pseudo) random, while in other embodiments, the modifications may be intentionally directed toward a local minimum in the scoring function. For example, gradient descent techniques may be used, which may include evaluating the scoring function of the recipe and / or potential modifications to the recipe to determine the direction in which the recipe will be modified. Another example is that the recipe may be modified in a direction that follows the direction of the modification in a previous iteration. For example, if an increase in the amount of one of the ingredients resulted in an improvement with the scoring function in a previous iteration, the amount may be further increased in a subsequent iteration. The type of modification to the recipe may also vary in time, or may be alternated. For example, in one or more initial iterations, the recipe may be randomly (pseudo) modified, while in subsequent iterations, the recipe may be intentionally modified, for example, using gradient descent, or by continuing the modification in a particular direction.

[0027] Optionally, providing a random set of recipes comprises randomly selecting relative contributions of ingredients. The random recipes may be generated by randomly selecting the relative contributions of these ingredients.

[0028] Optionally, randomly selecting the relative contributions of the components includes randomly setting the component contributions to values ​​selected from a range including zero as a lower limit and the maximum relative contribution of the component in the known recipe as an upper limit. Thus, the contribution of an individual component, e.g., its mass fraction, may be limited to a maximum value from known recipes. In this way, it is possible to avoid generating recipes that are significantly outside the data points represented by existing recipes, whereby the predictive model may not be able to accurately predict compound material properties of such recipes.

[0029] Optionally, the step of randomly selecting the relative contributions of the components includes randomly setting the contributions of the components to zero. Although the number of possible components may be large, a compound may actually have relatively few components, because such few components may be sufficient to meet the target specifications and / or the fewer components may contribute positively to the manufacturability of the compound. Thus, a recipe with fewer components may constitute a more realistic starting point for the recipe. By setting the material contributions to zero, such components may be omitted from the recipe. In some examples, the contributions of a majority of all possible components, e.g., 50%, 60%, 70%, 80%, 85% or 90% of all possible components, may be set to zero. For example, a randomly generated recipe may include all possible components with different, possibly small, relative contributions, e.g., all of the available components in total, 50 or 100, but a random recipe with, e.g., 10 or fewer components may be generated, thereby serving as a more realistic starting point for the recipe.

[0030] Optionally, generating candidate recipes for the compound includes receiving a starting recipe as input and generating candidate recipes based on the starting recipe. In some embodiments, recipe generation by the systems and methods is constrained to generate as output a recipe that minimally includes the components and amounts of components of the starting recipe. Thus, in such embodiments, the generated recipe may include only additional components and / or additional amounts of components. In this manner, a method can be determined to achieve or approximate a target specification of the thermoplastic compound by providing one or more additives.

[0031] Optionally, the scoring function is: Recipes with fewer ingredients; and The set of ingredients in a recipe that have a low relative contribution compared to the base component of the compound The device is further configured to provide a reward to at least one of the

[0032] The scoring function can not only quantify the agreement between the target value on the one hand and the predicted or actual measured value of the compound on the other hand, but can also embody other preferences for the recipe. For example, in general, recipes with fewer components may be preferred, for example, to limit the possibility of unexpected interactions between components, and because fewer components contribute to the manufacturability of the compound (e.g., by having to supply and store fewer components, etc.). Thus, in addition to expressing the correspondence between the material properties and the target specifications, the scoring function can prioritize recipes with fewer components, for example, by using the number of components as a negative factor in the scoring function. Similarly, for some components, such as additives, it may be preferred to use as few components as possible relative to the base component of the compound. This desire can also be expressed in the scoring function, for example, by using the amount of the component relative to the base component as a negative factor in the scoring function.

[0033] As explained above, the compound may comprise as a base constituent a thermoplastic polymer, preferably a thermoplastic polymer having a melting temperature of at least 200°C, or 150°C, or 100°C.

[0034] As explained above, the recipe can define additives to be added to the base polymer. Optionally, the additives include at least one colorant. Optionally, the at least one colorant includes at least one dye or at least one pigment.

[0035] Optionally, as part of the compound data, measurements and recipes of compounds containing only one type of colorant are provided. Such compounds may therefore represent "pure" colored compounds in that they can only be colored by one type of colorant, e.g., one type of pigment. Such compound data of this type may be beneficial in training predictive models, since measurements of such compounds may show the effect of a single colorant on the color of the compound, effectively representing data points on the boundaries of all recipes within the data space of the compound data.

[0036] Optionally, the one or more compound material properties include: the color of the compound, preferably defined as a color value in a perceptually uniform color space such as CIELAB and / or as a reflectance spectrum; One or more mechanical properties of the compound, preferably one or more of elongation at break, tensile strength and tensile modulus. Includes.

[0037] Advantageously, by training a predictive model to predict reflectance spectra, so-called metamerism can be avoided or at least predicted, whereby compounds having the same color in terms of color values ​​may appear to have the same color under one type of light source, e.g. a light source with a high color rendering index, but a different color under another type of light source, e.g. a light source with a low color rendering index. By predicting the reflectance curve, the appearance of the compound can be computationally evaluated for such different types of illumination, e.g. without the need to perform physical experiments. Advantageously, if the target specification provides a target reflectance curve, a recipe can be generated such that the compound approximates the target reflectance curve.

[0038] Optionally, the machine-learnable model includes a neural network or a Gaussian process. For example, in some embodiments, the machine-learnable model may be composed of one model capable of predicting several material properties. In other embodiments, the machine-learnable model may be composed of several individual sub-models, each capable of predicting a separate material property or a subset of material properties, and the sub-models together represent the machine-learnable model.

[0039] Optionally, the machine learning enabled model is configured to provide an uncertainty quantification of the selected recipe, the uncertainty quantification being output together with the selected recipe. Such uncertainty quantification may allow a user to obtain feedback regarding the certainty with which the thermoplastic compound is predicted to meet the target specifications. In this way, the user can obtain feedback regarding the probability that the compound produced according to the selected recipe will meet the target specifications. Furthermore, if the uncertainty is very high, this may indicate that more training data may be required for the predictive model to improve predictions in this portion of the solution space.

[0040] Optionally, a user interface subsystem can be provided to allow a user to interact with the processor system, with the selected recipe being output via the user interface subsystem and the sample measurements being received via the user interface subsystem.

[0041] In a further aspect of the invention, there is provided a method of generating a recipe for a compound, comprising executing on a processor system any one of the computer-implemented methods described herein, Producing a sample of the compound; and Measuring one or more material properties of the specimen to obtain sample measurements of the one or more material properties of the specimen. A method is provided, further comprising:

[0042] Thus, the above method may include performing the physical steps of generating a sample of the compound and measuring one or more material properties, in addition to using a processor system to perform any one of the computer-implemented methods described herein.

[0043] Optionally, generating a sample of the compound includes using at least one of a combination of extrusion and injection molding, and injection molding without extrusion.

[0044] Optionally, generating the recipe comprises at least two iterations of performing steps i) to v); In a first iteration, generating the sample includes using injection molding without extrusion; In the second and subsequent iterations, generating the samples includes using a combination of extrusion and injection molding.

[0045] Preferably, a so-called "short feedback loop" (or simply "short loop") can be used first, where a sample is generated using injection molding, but without extrusion. By omitting the extrusion step, the sample can be generated more efficiently, e.g., requiring less time and effort (hence the adjective "short"). Nevertheless, a sample generated by injection molding may be sufficient to initially verify whether the selected recipe at least reasonably approximates the target specification. After this first iteration, or after several such first iterations, the recipe may be closer to the target specification and a more accurate evaluation of the recipe may be preferable. Thus, a sample can then be generated using a combination of extrusion and injection molding ("long loop"), which may more closely resemble the real (mass-produced) product and thus result in more accurate measurements, at the expense of requiring more time and effort. Preferably, a long loop may only be performed when greater precision is required, i.e., when approaching the target specification, otherwise a short loop may be performed to save time and effort.

[0046] Optionally, the base polymer can be a virgin polymer, recycled polymer, blended or scrap polymer, colored polymer, or a mixture of the aforementioned types of polymers. Thus, the recipe can define additives that are added to either the virgin polymer, recycled polymer, blended or scrap polymer, or a mixture thereof to meet target specifications, for example, in terms of color and / or mechanical properties. In some embodiments, the compound may include two or more base polymers, in which case each base polymer can be the same or different of the aforementioned types of polymers.

[0047] Optionally, the base polymer is a colored polymer. For example, the base polymer may be a recycled polymer with a particular color. The methods and systems described herein can be used to generate recipes for differently colored thermoplastic compounds using the colored base polymer as a starting point. In such embodiments, the recipes can define colorants that are added to obtain the desired color and, optionally, other desired properties, such as mechanical properties.

[0048] Optionally, one or more compound material properties define a target value for color, and the base polymer is selected to have a color closest to the target value. In this manner, a starting point can be selected with respect to a color closest to the target value. For example, if the target value is defined as a point in CIELAB color space, a base polymer can be selected, for example, from a database of known base polymers having a CIELAB color closest to the point. Similarly, if the target value is defined as a reflectance spectrum, a base polymer can be selected having a reflectance spectrum closest to the target reflectance spectrum. It will be appreciated that "closest" can be quantified using any suitable metric, for example as a Euclidean distance in 3D CIELAB color space, or using a spectral metric.

[0049] Optionally, training the machine learning enabled model on the compound data comprises: Compound data, selecting pairs of compounds from the compound data, each pair including a first compound produced using a first recipe and a second compound produced using a second recipe; and determining a third recipe for producing a second compound using the first compound as a base polymer, the third recipe defining one or more additives to be added to the first compound to approximate values ​​of one or more compound material properties of the second compound, the third recipe being determined based on a difference between the first recipe and the second recipe; and processing the data by training the machine learning enabled model to predict the one or more compound material properties of the second compound based on the one or more compound material properties of the first compound and the third recipe as inputs, thereby obtaining as a predictive model a model that predicts values ​​of the one or more compound material properties of a compound produced from a combination of a) the input recipe and b) the value of the one or more compound material properties of the base polymer; Includes.

[0050] In a further aspect of the invention, there is provided a computer implemented method of generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound comprising a base polymer, the recipe defining a set of ingredients and relative contributions of the ingredients for producing the compound, the set of ingredients comprising one or more additives to be added to the base polymer; acquiring compound data for compounds that are thermoplastic compounds, the compound data including, for each compound, a) a measurement of one or more compound material properties of each compound, b) a measurement of one or more compound material properties of each base polymer, and c) a recipe defining one or more additives to be added to each base polymer to obtain each compound; training a machine learning enabled model on the compound data to predict values ​​of the one or more compound material properties of the compound from the values ​​of the one or more compound material properties of the base polymer and the input recipe, thereby obtaining a predictive model; To approximate a target value of one or more compound material properties. i) generating candidate recipes for compounds and predicting values ​​of one or more compound material properties for each candidate recipe using a predictive model; ii) selecting a recipe from the candidate recipes, the recipe comprising evaluating a scoring function, the scoring function configured to reward correspondence between the predicted values ​​and target values ​​of one or more compound material properties of the candidate recipes; and iii) outputting the selected recipe. generating a recipe for producing a compound by A computer-implemented method is provided, comprising:

[0051] Optionally, the compound data is second compound data; accessing first compound data for compounds that are thermoplastic compounds produced using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties, and b) a recipe for each compound; The second compound data is selecting pairs of compounds from the first compound data, each pair including a first compound produced using a first recipe and a second compound produced using a second recipe; and determining a third recipe for producing a second compound using the first compound as a base polymer, the third recipe defining one or more additives to be added to the first compound to approximate values ​​of one or more compound material properties of the second compound, the third recipe being determined based on a difference between the first recipe and the second recipe; The generating step and Further comprising: The second compound data includes, for each compound, a) measured values ​​of one or more compound material properties of each compound, b) measured values ​​of one or more compound material properties of each base polymer, and c) a recipe defining one or more additives to be added to each base polymer to obtain each compound.

[0052] In a further aspect of the invention, there is provided a method of producing a compound using a recipe generated by any one of the methods described herein.

[0053] In a further aspect of the invention there are provided compounds obtainable by the processes for making the compounds described herein.

[0054] In a further aspect of the invention, there is provided a computer implemented method for visualizing a parameter space of a recipe for a compound to be manufactured according to a target specification, the compound being a thermoplastic compound, the target specification comprising target values ​​for one or more compound material properties of the compound; obtaining compound data for compounds that are thermoplastic compounds produced using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties, and b) a recipe for each compound, the recipe defining a set of ingredients and the relative contributions of the ingredients for producing the compound; training a machine learning enabled model on the compound data to predict values ​​of one or more compound material properties from an input recipe, thereby obtaining a predictive model; obtaining target values ​​for one or more compound material properties of the thermoplastic compound to be produced; using the predictive model to visualize a parameter space associated with one or more compound material properties, the visualization including indicating one or more regions in the parameter space where a recipe results in or approximates a target value for the one or more compound material properties; A computer-implemented method is provided, comprising:

[0055] For example, the visualization can be generated by inputting recipes into a predictive model and determining which recipes result in or approximate a target value. Another example is that the visualization can be generated using a genetic algorithm to find recipes for compound material properties that result in or approximate a target value.

[0056] It will be appreciated by those skilled in the art that two or more of the above-described embodiments, implementations and / or optional aspects of the invention may be combined in any way deemed useful.

[0057] Modifications and variations of any processor system, any method, computer-implemented or otherwise, or any computer-readable medium corresponding to the described modifications and variations of another one of the entities may be made by one skilled in the art based on this specification.

[0058] These and other aspects of the invention will be apparent and elucidated further with reference to the embodiments described by way of example in the following description and with reference to the accompanying drawings, in which: [Brief description of the drawings]

[0059] [Figure 1] A method for generating a recipe for a compound is presented, the method including several computer-implemented steps, generating a sample based on a selected recipe, and measuring one or more material properties of the sample. [Diagram 2] A step of generating candidate recipes for a compound is shown, the generating step including providing a set of random recipes and iteratively modifying the set of random recipes to obtain improved scores by a scoring function. [Figure 3A] 13 illustrates a user interface element that allows a user to specify values ​​of target material properties for compounds for which a recipe is desired. [Figure 3B] 2 illustrates a user interface element showing a prediction of material properties for a recipe generated using the method of FIG. 1. [Figure 3C] 13 shows user interface elements that allow a user to input measurements of material properties of a sample produced based on a recipe. [Figure 4] 1 shows a processor system for generating a compound recipe. [Diagram 5] FIG. 1 illustrates a computer-readable medium containing data. It should be noted that the drawings are purely schematic and are not drawn to scale. In the drawings, elements corresponding to elements already described may have the same reference numbers.

[0060] [Explanation of reference numbers and abbreviations] The following list of reference numbers is provided to facilitate interpretation of the drawings and should not be construed as limiting the scope or provisions of the claims.

[0061] 20 Past Compound Data 40 Sample compound data 100 Computer-implemented method for recipe generation 110 Steps for training a machine learning capable model 120 Steps to generate candidate recipes 122 Steps to provide a set of random recipes 124 Steps to predict compound material properties 126 Steps to Evaluate Scoring Functions 128 Select the top N 130 Steps to change mass fraction 131 Recipe Mixing Steps 132 Steps to omit components 133 Step of adding ingredients 134 Steps to keep recipes 136 Steps to get a new set of recipes Steps repeated 138 M times 140 Steps for selecting recipes based on a scoring function 150 Steps to output selected recipes 160 receiving the measurement data and determining deviation from a target 170 Determining whether the recipe is acceptable 200 Steps for generating samples 210 Measuring the sample 300 processor system 310 Processing Subsystem 320 Data Storage Interface 330 Data Storage Unit 340 User Interface Subsystem 350 Display 360 User Input Devices 370 Network Interface 380 Network 400 Non-transitory computer-readable medium 410 Data

[0062] Detailed Description of the Embodiments In the following, definitions used in the present specification are introduced, and then a method for generating a compound recipe is described with reference to Figures 1 to 2, a user interface that allows a user to interact with the method or a processor system for generating a compound recipe is described with reference to Figures 3A to 3C, a processor system for generating a compound recipe is described with reference to Figure 4, and a computer-readable medium that stores data of the embodiments described in the present specification is described with reference to Figure 5.

[0063] definition The following definitions are used in the context of the present invention.

[0064] "Compound": The term "compound" may refer to a mixture of components, such as a mixture of a base polymer and an additive, and "base" may refer to the polymer that constitutes the main component of the compound. The polymer may preferably be a thermoplastic polymer. The thermoplastic polymer may have, for example, a melting temperature of at least 200°C, or at least 150°C, or at least 100°C.

[0065] "Ingredient": The term "ingredient" can refer to a constituent of a compound, the type and amount of which may be dictated by a recipe. Such ingredients can include base components, such as a base polymer, as well as additives that may be added to the base component, such as glass fibers, stabilizers, flame retardants, colorants (i.e., colorants such as pigments or dyes), processing aids, flow improvers, etc.

[0066] "Recipe": A recipe can define a set of ingredients and the relative contributions of the ingredients for producing a compound. The relative contributions can be expressed, for example, as mass fractions or percentages. A recipe can be provided, for example, in a computer-readable format, for example, as a text file or a CSV file, or in a markup language such as XML, or as one or more entries in a database, such as a SQL-based database.

[0067] "Preparation of a sample of a compound": can include any known manufacturing technique, including but not limited to injection molding, in combination with or without extrusion, blow molding and film extrusion. Such film extrusion can be used, for example, to prepare a sample of a compound by producing a film of the compound, for which material properties can be evaluated and measured from the film.

[0068] "Material properties" can include mechanical properties of a compound, such as physical properties such as elongation at break, tensile strength, tensile modulus, etc., but can also include non-mechanical properties such as UV stability performance, heat aging performance, viscosity, CTI (Comparative Tracking Index) value, flammability, etc.

[0069] FIG. 1 illustrates a method for generating a recipe for a compound, the method including several computer-implemented steps 100, which together with several physical steps 200, 210 may be referred to as method 100. Recipe generation according to the method may be described as follows. First, compound data 20 may be obtained or generated. The compound data 20 may be referred to elsewhere as a "dataset" and may include, for each or at least a subset of compounds, a) measurements of one or more compound material properties for the respective compound, and b) a recipe. The recipe and its measurement data may be referred to elsewhere as a "data point" in the dataset. Both the measurements and the recipe may be provided in a computer-readable format, for example, as a text or CSV file, or in a markup language such as XML, or as entries in a database, such as a SQL-based database.

[0070] As an example, in one implemented embodiment of the present invention, the ingredients and CIELAB color values ​​of over 2000 historical recipes were obtained and structured into a database. For many of these recipes, the physical properties of the compounds, such as elongation at break, tensile strength, and tensile modulus, were measured. For approximately 1400 recipes, the total reflectance curves of the compounds were measured.

[0071] In some embodiments, recipes included in the compound data can be selected according to one or more criteria. For example, recipes that contain obsolete ingredients, e.g. ingredients that are no longer available or ingredients that should not be used, e.g. for regulatory reasons, may be omitted. As another example, recipes with obvious errors, e.g. recipes with mass fractions far from 1.0 in total, or recipes with unknown / unknown ingredients may be omitted. In some embodiments, measurements and recipes of compounds that contain only one type of additive, such as a colorant, can be added as part of the compound data. For example, recipes and measurements of a single pigment can be added to the data set. In the implemented embodiment described above, reflectance curves and physical properties were measured for 20 different pigments and dyes at four different concentrations, resulting in 80 (additional) data points in the data set. In some embodiments, for example, in embodiments where the compound includes a thermoplastic polymer, such as a high temperature polymer (e.g., having a melting point above 100° C. or 150° C. or 200° C.), the compound data can also specify a grade type, which can indicate the color variation of the compound. For example, in the implemented embodiment described above, the compound data specified which of six possible grades the thermoplastic polymer of the compound belongs to. In embodiments where the components include additives, compounds without additives can also be added to the data set. For example, in the embodiment implemented above, natural polymers (e.g., without pigments and dyes) were added to the data set in six different grades.

[0072] After obtaining compound data 20, for example by (re)structuring existing data or new measurements, a machine-learnable model can be trained 110 on the compound data to predict values ​​of one or more compound material properties from an input recipe. Thus, the compound data 20 can constitute training data for the machine-learnable model. For example, a neural network (NN) or a Gaussian process (GP) can be trained on the compound data 20 as training data to obtain a predictive model for predicting material property values. In a particular example, the compound data can include measurements of compound material properties, such as compound color and one or more mechanical properties of the compound, and the machine-learnable model can be trained to predict values ​​of such material properties based on an input recipe. For example, in the compound data, color may be defined as color values ​​and / or reflectance spectra in a perceptually uniform color space such as CIELAB, and the compound mechanical properties may be one or more of elongation at break, tensile strength, and tensile modulus. Thus, the machine-learnable model can be trained to predict color and mechanical properties from a compound recipe.

[0073] In some embodiments, the machine-learnable model may also provide a confidence interval of its prediction, e.g., of each compound material property, or of its overall prediction. In some embodiments, the machine-learnable model may be a single model trained to predict several material properties together. Such a model may be referred to as a "combination" model in that it may predict a combination of several material properties, e.g., a combination of color values ​​(L*, a*, b* at the same time) or a combination of mechanical properties (tensile strength, tensile modulus, Charpy impact test score, etc.) or a combination of color values ​​and mechanical properties. In other embodiments, the machine-learnable model may be composed of a set of models, e.g., individual models each capable of predicting an individual material property, such as a color or mechanical property, or a set of combination models, etc. Such a machine-learnable model may be referred to as a "composite model." In addition to predicting one or a few values ​​of a property, the machine-learnable model may also predict a large number of values, e.g., an entire reflectance curve, which may allow, for example, the calculation of CIELAB color values ​​for any type of light source. Such a machine-learnable model may also be referred to as a "curve model" since it may be trained to predict a reflectance curve or other type of measurement curve. The machine-learnable model may generally be optimized for the material property to be predicted. For example, a NN model can be optimized with respect to the number of layers, the number of nodes and the overfitting risk. A GP model can include a combination of two kernels and an appropriate scaling of both the input and output data. Such types of optimization of model architectures and parameters are known per se.

[0074] After training 110 the machine learning enabled model on the compound data 20, recipes for the compounds can be generated to approximate the target specifications, i.e., to approximate the target values ​​of one or more compound material properties. To that end, candidate recipes for the compounds can be generated 120. Such a step of generating candidate recipes 120 can include predicting values ​​of one or more compound material properties for each candidate recipe using a predictive model. The generation of the candidate recipes can be followed by a step of selecting a recipe 140. Both steps can be further explained with reference to FIG. 2, which shows exemplary steps 122-138 for generating candidate recipes, followed by a recipe selection step 140. The steps 122-138 for generating candidate recipes can be related to the following: Such a step of generating candidate recipes 120 can be based on random recipes, in that a set of random recipes is obtained and can be iteratively modified to obtain improved scores according to a scoring function. Such iterative modification can include, for example, iteratively modifying the set of random recipes using a genetic algorithm to obtain improved scores according to a scoring function. When using such a genetic algorithm, and also in general, the material properties of the candidate recipes can be predicted by a predictive model and the predicted material properties can be evaluated by a scoring function. This quantification can make it possible to create improved variants of recipes, especially those recipes that have already been shown to be promising candidates for a compound by their scores by the scoring function. It is advantageous to be able to generate variants of the most promising recipes. These steps 122 to 138 can then be repeated until a recipe is selected 140 whose deviations from the target specifications are considered to be acceptable.

[0075] With continued reference to the scoring function, such a scoring function may be provided to evaluate the compatibility of the predicted material properties against the target specifications. In particular, the scoring function may be configured to reward correspondence between the predicted and target values ​​of one or more compound material properties of the candidate recipe. Such a scoring function may take various forms. As an example, in the above-described implemented embodiment, a scoring function may be provided that may be composed of four parts, each of which may be viewed as a sub-score. That is, the total score may be composed of a color score, a physical property score, a component number score, and a mass fraction sum score. Mathematically, the total score may be determined as a weighted combination of these scores, for example, as follows:

number

[0076] Here, the parameters a1, a2, a3, a4 can be optimized to give more importance to any individual score, if desired. For example, the parameters a3 and a4 can be selected as negative parameters, so that the scoring function rewards recipes with fewer ingredients (i.e., the number of ingredients can represent a penalty in the score calculation) and rewards recipes with a smaller total mass fraction of pigment relative to the mass of base polymer (i.e., the total mass fraction of pigment can represent a penalty). In this example, the scoring function can provide higher scores for recipes that may be more suitable, i.e., closer to the target specification, and provide lower scores for recipes that may be less suitable, i.e., significantly further from the target specification. However, without being limited thereto, for some types of scoring functions, lower scores may be preferred.

[0077] With continuing reference to color scores, such color scores may be expressed, for example, as a distance in 3D CIELAB color space and may be calculated as follows:

number

[0078] where L, a and b can represent individual color coordinates in the CIELAB color space, "model" color coordinates can represent color coordinates predicted by the predictive model, and "target" can represent color coordinates of the target specification. If color distance is used as the color score as in the example above, the parameter a1 can be chosen to reward smaller distances and / or negative to penalize larger distances.

[0079] For physical properties such as mechanical properties that should exceed a required minimum specification, a so-called s-curve can be used, with a sharp increase in score at the required minimum value. In this way, scores below the minimum specification can be severely penalized, while scores above the minimum specification can still be incentivized. An example score for a physical property (Charpy impact test score) with a target value of 10 may be the following, where the score may be defined as a function: [Table 1]

[0080] 2, in step 122, some random recipes may be provided. For example, such random recipes may be provided by randomly selecting the ingredients and / or relative contributions of the ingredients of the recipe. In some embodiments, such relative contributions may be limited to a range, for example, to define an acceptable bandwidth of the mass fraction of the ingredients. This range may extend, for example, from zero as a lower limit to a maximum value. This maximum value may be, for example, user-defined, or may be automatically selected, for example, based on the maximum relative contributions of the ingredients in known recipes. Thus, such a range may represent an acceptable bandwidth of the relative contributions per ingredient. In some embodiments, the user may define the maximum value to 0 to exclude one ingredient. In some embodiments, any such limitations on the relative contributions may be applied to the recipe after the recipe is randomly generated. With continued reference to step 122, the relative contributions of ingredients, for example, as initially randomly generated, may also be randomly set to zero. That is, since most real recipes have only a small number of ingredients, for example, after the recipe is randomly generated, the contributions of most ingredients may be randomly set to zero. This allows a recipe to always be defined by the same data structure, e.g. the same vector defining the relative contribution of each possible ingredient, e.g. a 50, 75 or 100 dimensional vector, but limited to having only a small number of ingredients, e.g. a maximum of 10 or 5 ingredients.

[0081] In step 124, the predictive model may be used to predict values ​​of material properties for each randomly generated recipe. In step 126, a scoring function may be evaluated to determine a (numeric) score for each recipe. In steps 128-134, a genetic algorithm may be used to improve the recipes. That is, in step 128, the best recipes may be selected, e.g., as the top N of all M recipes. In steps 130-133, various variants of these best recipes may be generated, e.g., by step 130, which may include randomly changing the relative contributions (e.g., mass fractions of ingredients in the recipes), by step 131, which may include blending two or more recipes, by step 132, which may include randomly omitting ("dropping") ingredients from the recipe, and by step 133, which may include randomly adding ingredients to the recipe. In step 134, the best recipes themselves may also be retained, in that the new recipes obtained in step 136 may be composed of the best recipes (e.g., the top N) and their variants generated by steps 130-133. In a particular example, 1000 recipes may be randomly generated initially. Of the 1000 recipes, only the 200 best recipes (with the highest scores) may be kept, and 800 new recipes may be created by generating variations of those 200 recipes. This may include, for example, 200 random changes in mass fractions, 200 blends of two recipes, 200 random drops of ingredients, and 200 random additions of ingredients. This process of improving recipes using a genetic algorithm may be repeated 138 times, for example 50 or 100 times.

[0082] 1 and 2 together, from the candidate recipes generated by steps 122-138, a best recipe may be selected (140), for example, by applying the predictive model to the latest set of candidate recipes last generated by the iteration of steps 122-138, and selecting the best recipe according to a scoring function. The selected recipe may then be output, for example, in a human readable format, using an output device such as a display, in step 150. Based on the output recipe, a sample may be generated, in step 200, using a manufacturing step, for example, a combination of extrusion and injection molding or injection molding without extrusion. In step 210, the sample may be measured, in that one or more material properties of the sample may be measured. As a result, sample measurements of one or more material properties of the sample may be obtained. In step 160, these measurements may again be received by the method 100, for example, by a user inputting values ​​using a graphical user interface. Also, in step 160, the sample measurements may be compared to target values ​​to determine deviations from the target values. Such deviations may be expressed in many ways, for example, as a sum of absolute differences or a mean squared error. The deviations can then be evaluated against the acceptability criteria in step 170, and if the deviations meet the acceptability criteria, the selected recipe can be output as a recipe for the compound, for example using the display described above or by writing the recipe as data in a data store. An example of an acceptability criterion is a threshold, for example an absolute threshold or a relative threshold. In another example, the deviations can be presented to a user, for example using a graphical user interface on a display, and the user can accept or reject them.

[0083] In some embodiments, if the deviations do not meet the acceptability criteria, the sample measurements and sample recipes can be used to retrain the predictive model 110, which is illustrated in FIG. 1 as training 110 being performed again on the sample data 40, including the aforementioned sample measurements and sample recipes. In other embodiments, particularly those in which the sample data 40 is used by the predictive model during inference, the sample measurements and sample recipes can simply be added to the sample data 40 instead of retraining the predictive model. In both types of embodiments, the method 100 can be repeated, for example, until a recipe is obtained that is deemed acceptable according to the acceptability criteria.

[0084] 3A-3C show a portion of a user interface for allowing a user to interact with the processor system described with reference to FIG. 4 using the method of FIGS. 1-2. Here, FIG. 3A shows a user interface element that allows a user to specify a value of a target material property for a compound for which a recipe is sought. Such a value may be entered numerically, for example, using a keyboard, or may be selected graphically, for example, using a mouse. Furthermore, the target value may be specified as a value to be reached, for example, using an equal operator ("="), or as a minimum value, for example, using a greater than operator (">"), or as a maximum value, for example, using a less than operator ("<"). Although not shown in FIG. 3A itself, the target material property itself may be preselected by the user, for example, from several available target material properties, or may be predefined in a graphical user interface. In some embodiments, an importance may be selected for each target material property, which importance may be used to adjust the scoring function, for example, by adjusting the parameters a1, a2, a3, a4 described above.

[0085] Figure 3B illustrates a user interface element showing predictions of material properties for recipes generated using the method of Figure 1. Such a user interface element may be used, for example, to output a selected recipe to enable production of a sample in step 150 and / or to verify the acceptability of a selected recipe in step 170. As seen in Figure 3B, the user interface element may display predicted values, including accuracy, of at least some of the material properties of Figure 3A.

[0086] 3C shows a user interface element that allows a user to input measurements of material properties of a sample produced based on a recipe. In this example, the CIELAB values ​​L*=55.8, a*=-0.5, and b*=-11.2 were entered. In some embodiments, several sets of values ​​may be entered, e.g., representing several measurements of the same sample, e.g., to account for measurement variability.

[0087] 3A-3C, the user interface can in some embodiments present the nearest recipes for any given target specification, where the nearest recipes are past recipes from the compound data whose scores according to the scoring function indicate the best approximation of the target specification, or at least a portion of the target specification, such as color. Such nearest recipes can be used to determine improvement of recipes generated by the genetic algorithm.

[0088] The following embodiments relate to a machine-learnable model that is trained to further use the material properties of the base polymer as inputs in training and inference. In such embodiments, training the machine-learnable model on the compound data can include first processing the compound data. For example, the compound data can be processed by selecting pairs of compounds from the compound data, each pair including a first compound produced using a first recipe and a second compound produced using a second recipe, and then determining a third recipe for producing the second compound using the first compound as a base polymer, the third recipe defining one or more additives to be added to the first compound to approximate the value of one or more compound material properties of the second compound, the third recipe being determined based on the difference between the first and second recipes. The machine-learnable model can then be trained to predict one or more compound material properties of the second compound based on the one or more compound material properties of the first compound and the third recipe as inputs, thereby obtaining a model that predicts the value of one or more compound material properties of a compound produced from a combination of the input recipe and the value of one or more compound material properties of the base polymer as a predictive model.

[0089] In a related embodiment, a computer-implemented method for generating a recipe for a thermoplastic compound can be provided. The thermoplastic compound can be a compound that includes a base polymer, and the recipe can define a set of ingredients and the relative contributions of the ingredients to produce the compound, the set of ingredients including one or more additives to be added to the base polymer. The method can include obtaining compound data for the compounds that are thermoplastic compounds, the compound data including, for each compound, a) measurements of one or more compound material properties of the respective compound, b) measurements of one or more compound material properties of the respective base polymer, and c) a recipe that defines one or more additives to be added to the respective base polymer to obtain the respective compound. A machine-learning capable model can be trained on the compound data to predict values ​​of the one or more compound material properties of the compound from the input recipe and the one or more compound material properties of the base polymer, thereby obtaining a predictive model.

[0090] As described elsewhere, a recipe for a compound to be produced may be generated by generating candidate recipes for the compound, predicting values ​​of the one or more compound material properties for each candidate recipe using a predictive model to approximate a target value of the one or more compound material properties, and selecting a recipe from the candidate recipes, the recipe selection including evaluating a scoring function configured to reward correspondence between the predicted values ​​and target values ​​of the one or more compound material properties of the candidate recipes, and outputting the selected recipe. For example, a genetic algorithm may be used, as described elsewhere.

[0091] According to the above embodiment regarding the machine-learnable model trained to further use the material properties of the base polymer as input, the following is noted. The machine-learnable model can be trained to predict one or more compound material properties from a combination of an input recipe and one or more compound material properties of the base polymer to which the input recipe is associated. That is, the input recipe can define one or more additives to be added to the base polymer. The machine-learnable model can receive the input recipe as well as the compound material properties of the base polymer as input and can be trained to predict the compound material properties of the compound produced using the input recipe. This allows the machine-learnable model to better take into account the starting polymer, e.g., the base polymer to which the additive is added, which is advantageous since there may be different starting polymers available with different properties. For example, virgin polymers may differ in terms of their material properties. Thus, if the compound to be produced has a target color, the color of the starting polymer may be taken into account when determining how to best reach the target color. It is noted that in some embodiments, the compound material properties of the base polymer may be explicitly or implicitly defined as part of the recipe. An example of the latter is that the recipe may specify a particular base polymer whose material properties are known and well-defined. In such cases, it may not be necessary to provide the material properties of the base polymer as input to the machine-learnable model, but rather it may be sufficient to provide an identifier of the base polymer to the machine-learnable model.

[0092] In some embodiments, again relating to the machine-learnable model trained to further use the material properties of the base polymer as input, the compound data on which the machine-learnable model is trained is the second compound data. In such embodiments, a first compound data of compounds that are thermoplastic compounds produced using a known recipe can be accessed, the compound data including, for each compound, a) measurements of one or more compound material properties of each compound, and b) a recipe. The second compound data can then be generated by selecting pairs of compounds from the first compound data, each pair including a first compound produced using the first recipe and a second compound produced using the second recipe, and determining a third recipe for producing the second compound using the first compound as the base polymer, the third recipe defining one or more additives to be added to the first compound to approximate the value of one or more compound material properties of the second compound, the third recipe being determined based on the difference between the first recipe and the second recipe. The second compound data then includes, for each compound, a) measurements of one or more compound material properties of each compound, b) measurements of one or more compound material properties of each base polymer, and c) a recipe that defines one or more additives to be added to each base polymer to obtain each compound. In this way, training data can be obtained for training a machine-learning enabled model that is trained to predict one or more compound material properties from a combination of an input recipe and one or more compound material properties of the base polymer to which the input recipe is associated. Obtaining such training data can be particularly important for the following reasons.

[0093] The historical data for a compound may include measurements of the compound material properties of the compound and recipes for making the respective compounds, but may not include measurements of the compound material properties of the base polymers used in these compounds. Rather, the recipes may simply identify the base polymers, which historically may typically be virgin polymers. However, non-virgin polymers are gaining importance for sustainability reasons, such as reclaimed polymers, blended polymers or scrap polymers. These polymers may have significantly different material properties, for example, with respect to color, mechanical properties, etc. To be able to generate recipes using such base polymers as a starting point, a large amount of new measurements may be required, which is a complex, expensive and time-consuming endeavor.

[0094] The above means avoids the need to obtain such new measurement data by using existing past compound data and processing the past compound data as follows: That is, pairs of compounds, for example pairs of compounds that are similar in terms of material properties, are identified as compound data. One recipe of a compound, sometimes called a "second" compound, can include an additional additive or a further amount of an existing additive compared to the first compound. This can mean that the second compound can be made by using the first compound as a starting point and adding an additional or further amount of additive. In other words, the first compound can be considered as a base polymer for the second compound. The recipe that defines how the second compound is made from this base polymer can be determined by comparing the recipes of the first compound and the second compound. In a particular example, if the recipes are defined as vectors, the vector that defines the recipe of the first compound can be subtracted from the vector that defines the recipe of the second compound to obtain a vector for a recipe for obtaining the second compound using the first compound as a base polymer. In this way, it can be avoided to have to perform a large amount of new measurements, which can save costs, reduce complexity, and save time.

[0095] FIG. 4 illustrates a processor system 300 for generating a recipe for a compound. The processor system 300 can include a data storage interface 320 for reading and / or writing any type of data described herein from and / or to the data storage 330. The data storage 330 can store output data generated by the processor system 300, such as, for example, the compound data 20, a computer-readable version of the machine learning enabled model, and data representing the selected recipe. The data storage 330 can take various forms, such as a hard drive or array of hard drives, a solid state drive or array of solid state drives, memory, and the like. By way of example, FIG. 4 illustrates the data storage 330 as an external data storage, but the data storage 330 can also be an internal component of the processor system 300. The processor system 300 can further include a network interface 370 to a network 380, such as a local area network (LAN) or a wide area network (WAN) such as the Internet. Network interface 370 may be, for example, a wired communication interface, such as an Ethernet or fiber optic based interface, or a wireless communication interface, for example based on 5G, Wi-Fi, Bluetooth, ZigBee, etc. In yet another example, network interface 370 may be a virtual, i.e., software-defined, network interface.

[0096] The processor system 300 may further comprise a processing subsystem 310, which may be configured, for example, by hardware design or software, to perform the operations described herein insofar as they relate to generating a recipe for a compound. In general, the processing subsystem 310 may be embodied by a single CPU, such as an x86 or ARM-based CPU, but may also be embodied by a combination or system of such CPUs and / or other types of processing units, such as GPUs, NPUs, etc. In embodiments in which the processor system 300 is distributed across different entities, e.g., different servers, the processing subsystem 310 may also be distributed, for example, across the respective CPUs.

[0097] The processor system 300 may further comprise a user interface subsystem 340 that may be configured to enable a user to interact with the processor system 300 during operation of the processor system 300, for example, using a graphical user interface. In particular, the graphical user interface may enable a user to read or interpret the output of the processor system 300 and provide input, for example, to specify target values ​​for material properties, to input measurements, to approve or reject recipes, etc. For that purpose, and for other purposes, the user interface subsystem 340 may comprise a user input interface (not explicitly shown in FIG. 4 ) configured to receive user input data from a user input device 360 ​​operable by a user. The user input device 360 ​​may take a variety of forms, including, but not limited to, a computer mouse, a touch screen, a computer keyboard, a microphone, and the like. FIG. 4 illustrates user input devices that are a computer keyboard and a computer mouse 360. In general, the user input interface may be of a type that corresponds to the type of the user input device 360, i.e., a corresponding type of user device interface. The user interface subsystem 180 may further include a display output interface (not explicitly shown in FIG. 4) configured to provide display data to a display 350 for visualizing the output of the processor system 300. In the example of FIG. 1, the display is an external display 350. Alternatively, the display may be an internal display of the processor system 300.

[0098] In some embodiments, the processing system 300 may be a server capable of operating with a client (not shown in FIG. 4 ) according to a client-server model. In such embodiments, a graphical user interface may be presented on the client, and information shown to a user using the graphical user interface, and input provided by a user using the graphical user interface, may be transmitted from and to the processing system 300. For example, the processing system 300 may be configured to establish a web-accessible graphical user interface that may be accessed by a client via the Internet or an intranet, or to interface with a web server to establish such a graphical user interface for access by a client.

[0099] In general, each entity described herein may be embodied as or in a device or apparatus. The device or apparatus may comprise one or more (micro)processors running appropriate software. The processor of each entity may be embodied by one or more of these (micro)processors. Software implementing the functionality of each entity may be downloaded and / or stored in a corresponding memory or memories, e.g. volatile memory such as RAM, or non-volatile memory such as Flash. Alternatively, the processor of each entity may be implemented in the device or apparatus in the form of programmable logic, e.g. as a Field Programmable Gate Array (FPGA). Any input and / or output interfaces may be implemented by respective interfaces of the device or apparatus. Each functional unit of each entity may be implemented in the form of a circuit or circuitry. Each entity may also be implemented in a distributed manner, e.g. including different devices or apparatus. An example of a device or apparatus is a computer, such as a workstation or a server.

[0100] It should be noted that any of the methods described herein, e.g., in any of the claims or clauses, may be implemented on a computer as a computer-implemented method, as dedicated hardware, or as a combination of both. Instructions for a computer, e.g., executable code, may be stored in a computer-readable medium 400, e.g., in the form of a series of machine-readable physical marks 410 and / or as a series of elements having different electrical, e.g., magnetic or optical properties or values, as shown in FIG. 5. The executable code may be stored temporarily or non-temporarily. Examples of computer-readable media include memory devices, optical storage devices, integrated circuits, etc. FIG. 5 shows an optical storage device 400 as an example.

[0101] The following clauses define embodiments of the computer-implemented method and processor system for generating recipes for compounds, methods for manufacturing compounds, and compounds, which may be claimed separately and / or combined with other embodiments described herein. Clause 1. A computer-implemented method for generating a recipe for a compound, the recipe defining a set of ingredients and the relative contributions of the ingredients for producing the compound; acquiring compound data for compounds produced using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties, and b) a recipe for each compound; training a machine learning enabled model on the compound data to predict values ​​of one or more compound material properties from an input recipe, thereby obtaining a predictive model; To approximate a target value of one or more compound material properties. i) generating candidate recipes for compounds and predicting values ​​of one or more compound material properties for each candidate recipe using a predictive model; ii) selecting a recipe from the candidate recipes, the recipe comprising evaluating a scoring function, the scoring function configured to reward correspondence between the predicted values ​​and target values ​​of one or more compound material properties of the candidate recipes; iii) outputting the selected recipe to enable production of a sample of the compound and measurement of one or more material properties of the sample; iv) receiving a sample measurement value of one of the sample or material properties and comparing the sample measurement value to a target value to determine a deviation from the target value; and v) outputting the selected recipe as the recipe for the compound if the deviation meets the acceptability criteria. generating a recipe for the compound by 4. A computer-implemented method comprising:

[0102] Clause 2. The computer-implemented method of clause 1, further comprising the steps of retraining the predictive model using the sample measurements and the sample recipe if the deviations do not meet the acceptability criteria, and repeating steps i) to v) using the retrained predictive model.

[0103] Clause 3. The computer-implemented method of clause 1 or 2, wherein generating candidate recipes for the compounds includes providing a set of random recipes and generating candidate recipes based on the set of random recipes.

[0104] Clause 4. The computer-implemented method of clause 3, wherein generating candidate recipes for the compounds includes using a genetic algorithm to iteratively modify a set of random recipes to obtain improved scores according to the scoring function.

[0105] Clause 5. The step of iteratively varying the set of random recipes comprises: Randomly varying the relative contributions of ingredients in a recipe; Mixing two or more recipes; Randomly omitting ingredients from the recipe; and Steps to randomly add ingredients to a recipe 5. The computer-implemented method of claim 4, comprising at least one of:

[0106] Clause 6. Providing a random set of recipes, randomly setting the contribution of the ingredient to a value selected from a range including zero as a lower limit and the maximum relative contribution of the ingredient in a known recipe as an upper limit; and Randomly setting component contributions to zero 6. The computer-implemented method of any one of clauses 1 to 5, comprising the step of randomly selecting the relative contributions of the components, preferably by at least one of

[0107] Clause 7. The scoring function is Recipes with fewer ingredients; and The set of ingredients in a recipe that have a low relative contribution compared to the base component of the compound 7. The computer-implemented method of any one of clauses 1-6, further configured to reward at least one of the

[0108] Clause 8. One or more compound material properties are the color of the compound, preferably defined as a color value in a perceptually uniform color space such as CIELAB and / or as a reflectance spectrum; One or more mechanical properties of the compound, preferably one or more of elongation at break, tensile strength and tensile modulus. 8. The computer-implemented method of any one of clauses 1 to 7, comprising:

[0109] Clause 9. A method for generating a recipe for a compound, comprising the step of executing the computer-implemented method according to any one of clauses 1 to 8 on a processor system, Producing a sample of the compound; and Measuring one or more material properties of the specimen to obtain sample measurements of the one or more material properties of the specimen. The method further comprising:

[0110] Clause 10. The step of generating a sample of a compound, Combining extrusion and injection molding; and Injection molding without extrusion 10. The method of claim 9, comprising using at least one of

[0111] Clause 11. The step (200) of generating a recipe includes at least two iterations of performing steps i) to v); In a first iteration, generating the sample includes using injection molding without extrusion; In the second and subsequent iterations, generating the specimen includes using a combination of extrusion and injection molding. 10. The method according to clause 9 or 10.

[0112] Clause 12. A method for producing a compound using a recipe generated by the method according to any one of clauses 1 to 11.

[0113] Clause 13. A compound obtainable by the method according to clause 12.

[0114] Clause 14. A transitory or non-transitory computer readable medium containing data representing a computer program, the computer program containing instructions for causing a processor system to carry out the method according to any one of clauses 1 to 11.

[0115] Clause 15. A processor system for generating a recipe for a compound, the recipe defining a set of ingredients and the relative contributions of the ingredients to a recipe for producing the compound; a data storage interface configured to access compound data for compounds generated using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties, and b) a recipe for each compound; training a machine learning enabled model on the compound data to predict values ​​of one or more compound material properties from an input recipe, thereby obtaining a predictive model; To approximate a target value of one or more compound material properties. i) generating candidate recipes for compounds and predicting values ​​of one or more compound material properties for each candidate recipe using a predictive model; ii) selecting a recipe from the candidate recipes, the recipe comprising evaluating a scoring function, the scoring function configured to reward correspondence between the predicted values ​​and target values ​​of one or more compound material properties of the candidate recipes; iii) outputting the selected recipe to enable production of a sample of the compound and measurement of one or more material properties of the sample; iv) receiving sample measurements of one or more material properties of the sample and comparing the sample measurements to target values ​​to determine deviations from the target values; and v) outputting the selected recipe as the recipe for the compound if the deviation meets the acceptability criteria. To generate compound recipes, The processing subsystem that is configured A processor system comprising:

[0116] No example, embodiment or optional feature, whether non-limiting or not, should be understood as limiting the invention as recited in the claims or defined by any of the clauses.

[0117] The mathematical symbols and notations are provided to facilitate interpretation of the invention and should not be construed as limiting the claims or provisions.

[0118] It should be noted that the above embodiments are illustrative rather than limiting of the invention, and that those skilled in the art can design many alternative embodiments without departing from the scope of the appended claims or clauses. In a claim or clause, any reference sign placed between parentheses shall not be construed as limiting the claim or clause. Use of the verb "comprise" and its conjugations does not exclude the presence of elements or steps other than those stated in the claim or clause. The article "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Expressions such as "at least one" when preceding a list or group of elements, denote the selection of all or any subset of the elements from the list or group. For example, the expression "at least one of A, B, and C" shall be understood as including A only, B only, C only, both A and B, both A and C, both B and C, or all of A, B, and C. The invention can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In a device or system claim or clause enumerating several means, several of these means can be embodied by one and the same item of hardware. The mere fact that certain means are recited in mutually different dependent claims or clauses does not indicate that a combination of these means cannot be used to advantage.

Claims

1. A computer-implemented method (100) for generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound that includes a base polymer, the recipe defining a set of ingredients and relative contributions of the ingredients for producing the compound, the set of ingredients including one or more additives that are added to the base polymer; acquiring compound data (20) for compounds that are thermoplastic compounds produced using a known recipe, said compound data including, for each compound, a) measurements of one or more compound material properties of said each compound, and b) a recipe; training 110 a machine learning enabled model on the compound data to predict values ​​of the one or more compound material properties from an input recipe, thereby obtaining a predictive model; to approximate a target value of the one or more compound material properties. i) generating candidate recipes for the compounds (120) and predicting values ​​of the one or more compound material properties for each candidate recipe using the predictive model; ii) selecting (140) a recipe from the candidate recipes, comprising evaluating a scoring function, the scoring function configured to reward correspondence between predicted and target values ​​of the one or more compound material properties of the candidate recipes; iii) outputting (150) the selected recipe to enable the production of a sample of the compound and the measurement of one or more material properties of the sample; iv) receiving sample measurements of one or more material properties of the sample and comparing the sample measurements to the target values ​​to determine deviations from the target values ​​(160); and v) if the deviation meets the acceptability criteria, outputting (170) the selected recipe as a recipe for the compound, and if the deviation does not meet the acceptability criteria, retraining (110) or updating the predictive model using the sample measurements and the sample recipe, and repeating steps i) through v) (120 through 170) using the predictive model. generating a recipe for said compound by A computer-implemented method (100), comprising:

2. the predictive model comprises a neural network, and the neural network is retrained; and / or the predictive model includes a Gaussian process, the Gaussian process uses the compound data for inference, and the Gaussian process is updated by adding the sample measurements and the sample recipe to the compound data.

10. The computer-implemented method of claim 1.

3. 3. The computer-implemented method of claim 1, wherein generating candidate recipes for the compound comprises providing a set of random recipes and generating the candidate recipes based on the set of random recipes.

4. 4. The computer-implemented method of claim 3, wherein generating candidate recipes for the compounds (120) comprises using a genetic algorithm to iteratively modify (130-133) the set of random recipes to obtain improved scores by the scoring function.

5. iteratively varying the set of random recipes, Randomly varying the relative contributions of the ingredients in the recipe (130); Blending two or more recipes (131); Randomly omitting ingredients from the recipe (132); Randomly adding ingredients to the recipe (133); modifying the recipe in a selected direction based on a previous modification direction in a previous iteration of the genetic algorithm; and Using gradient descent techniques to modify the recipe towards a local minimum of the scoring function. The computer-implemented method (100) of claim 4, comprising at least one of:

6. Providing a random set of recipes (122) randomly setting the contribution of an ingredient to a value selected from a range including zero as a lower limit and the maximum relative contribution of said ingredient in said known recipe as an upper limit; and Randomly setting component contributions to zero The computer-implemented method (100) of any one of claims 1 to 5, further comprising the step of randomly selecting the relative contributions of said components, preferably by at least one of:

7. The scoring function is Recipes with fewer ingredients; and A set of ingredients in a recipe that have a low relative contribution compared to the base ingredient of said compound The computer-implemented method (100) of any one of claims 1 to 6, further configured to provide a reward to at least one of the following:

8. The one or more compound material properties are the color of said compound, preferably defined as a color value in a perceptually uniform color space such as CIELAB and / or as a reflectance spectrum; One or more mechanical properties of said compound, preferably one or more of elongation at break, tensile strength and tensile modulus. The computer-implemented method (100) of any one of claims 1 to 7, comprising:

9. A method for generating a recipe for a compound, comprising executing on a processor system the computer-implemented method (100) of any one of claims 1 to 8, comprising: Producing a sample of said compound (200); and Measuring (210) one or more material properties of the specimen to obtain sample measurements of one or more material properties of the specimen. The method further comprising:

10. Producing a sample of the compound (200) A combination of extrusion and injection molding; and Injection molding without extrusion The method of claim 9, comprising using at least one of

11. generating the recipe (200) includes at least two iterations of performing steps i) through v) (120 through 170); In a first iteration, generating the sample includes using injection molding without extrusion; In the second and subsequent iterations, generating the sample includes using a combination of extrusion and injection molding.

11. The method according to claim 9 or 10.

12. 12. The method of any one of claims 1 to 11, wherein the machine learning enabled model is configured to provide an uncertainty quantification for the selected recipe, the method further comprising outputting the uncertainty quantification together with the selected recipe.

13. training the machine learning enabled model on the compound data, The compound data is selecting pairs of compounds from the compound data, each pair including a first compound produced using a first recipe and a second compound produced using a second recipe; and determining a third recipe for producing the second compound using the first compound as a base polymer, the third recipe defining one or more additives to be added to the first compound to approximate values ​​of one or more compound material properties of the second compound, the third recipe being determined based on a difference between the first recipe and the second recipe; and processing by training the machine learning enabled model to predict one or more compound material properties of the second compound based on one or more compound material properties of the first compound and the third recipe as inputs, thereby obtaining as the predictive model a model that predicts values ​​of one or more compound material properties of the compound produced from a combination of a) the input recipe and b) values ​​of one or more compound material properties of the base polymer; The method according to any one of claims 1 to 12, comprising:

14. The method of any one of claims 1 to 13, wherein the base polymer is a virgin polymer, a recycled polymer, a blended polymer, a colored polymer, or a scrap polymer.

15. The method of any one of claims 1 to 14, wherein the base polymer is a pigmented polymer.

16. A method for producing a thermoplastic compound using a recipe generated by the method of any one of claims 1 to 15 or 19 or 20.

17. A thermoplastic compound obtainable by the process according to claim 16.

18. A transitory or non-transitory computer readable medium (400) comprising data (410) representing a computer program, said computer program comprising instructions for causing a processor system to perform the method of any one of claims 1 to 15.

19. 1. A processor system (300) for generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound that includes a base polymer, the recipe defining a set of ingredients and relative contributions of the ingredients to the recipe for producing the compound, the set of ingredients including one or more additives that are added to the base polymer; a data storage interface (320) configured to access compound data (20) of compounds that are thermoplastic compounds produced using a known recipe, the compound data including, for each compound, a) measurements of one or more compound material properties of the respective compound, and b) a recipe; training a machine learning enabled model on the compound data to predict values ​​of the one or more compound material properties from an input recipe, thereby obtaining a predictive model; to approximate a target value of the one or more compound material properties. i) generating candidate recipes for the compounds and predicting values ​​of the one or more compound material properties for each candidate recipe using the predictive model; ii) selecting a recipe from the candidate recipes, the recipe comprising evaluating a scoring function, the scoring function configured to reward correspondence between predicted and target values ​​of the one or more compound material properties of the candidate recipes; iii) outputting the selected recipe to enable the production of a sample of the compound and the measurement of one or more material properties of the sample; iv) receiving sample measurements of one or more material properties of the sample and comparing the sample measurements to the target values ​​to determine deviations from the target values; and v) outputting the selected recipe as a recipe for the compound if the deviation meets the acceptability criteria, and updating the predictive model using the sample measurements and the sample recipe if the deviation does not meet the acceptability criteria, and repeating steps i) through v) using the updated predictive model. To generate compound recipes, A processing subsystem (310) configured A processor system (300) comprising:

20. A computer-implemented method (100) for generating a recipe for a thermoplastic compound, the thermoplastic compound being a compound that includes a base polymer, the recipe defining a set of ingredients and relative contributions of the ingredients for producing the compound, the set of ingredients including one or more additives that are added to the base polymer; acquiring compound data (20) for compounds that are thermoplastic compounds, said compound data including, for each compound, a) measurements of one or more compound material properties of said each compound, b) measurements of one or more compound material properties of a each base polymer, and c) a recipe defining one or more additives to be added to said each base polymer to obtain said each compound; training 110 a machine learning enabled model on the compound data to predict values ​​of one or more compound material properties of the compound from values ​​of one or more compound material properties of the base polymer and an input recipe, thereby obtaining a predictive model; to approximate a target value of the one or more compound material properties. i) generating candidate recipes for the compounds (120) and predicting values ​​of the one or more compound material properties for each candidate recipe using the predictive model; ii) selecting (140) a recipe from the candidate recipes, the selecting step including evaluating a scoring function, the scoring function configured to reward correspondence between predicted and target values ​​of the one or more compound material properties of the candidate recipes; and iii) outputting the selected recipe (150); generating a recipe for producing the compound by A computer-implemented method (100), comprising:

21. the compound data is second compound data, accessing first compound data for compounds that are thermoplastic compounds produced using a known recipe, said compound data including, for each compound, a) measurements of one or more compound material properties of said each compound, and b) a recipe; The second compound data, selecting pairs of compounds from the first compound data, each pair including a first compound produced using a first recipe and a second compound produced using a second recipe; and determining a third recipe for producing the second compound using the first compound as a base polymer, the third recipe defining one or more additives to be added to the first compound to approximate values ​​of one or more compound material properties of the second compound, the third recipe being determined based on a difference between the first recipe and the second recipe; The generating step and 21. The computer-implemented method of claim 20, further comprising:

Images