Machine learning based predictive methodology for the development of composite materials for tire tread compounds

Abstract

The present invention relates to a machine learning based predictive method for the development of composite materials for tire tread compounds.

Description

[Technical field]

[0001] The present invention relates to a method for predicting the viscoelastic or processability properties of rubber compounds before, during and after vulcanization, which method is based on machine learning and is therefore implemented by a computer for the development of composite materials for tire tread compounds. [Background technology]

[0002] The present invention relates to the field of tire manufacturing, and in particular to the determination of the composition of the rubber compounds used in the manufacture of tire treads.

[0003] The RPA (rubber process analyzer) is an advanced dynamic mechanical rheology testing instrument commonly available in every mill.

[0004] RPA (Rubber Process Analyzer), as an advanced dynamic mechanical rheological testing instrument, is generally available in every plant to monitor the composite manufacturing parameters at each step of the process. In fact, the workability of a composite is determined by the specific ranges of the rheometric curves and shear modulus descriptors before and after curing defined during the development stage (e.g. ML and MH torque, T10, T50 and T90, scorch time, vulcanized and unvulcanized shear modulus G' and tand at constant applied stress conditions).

[0005] These properties are ensured by the characteristics of the recipe used in the composite material, in particular the components, their amounts and the particular synergies established between two or more of them. Summary of the Invention [Problem to be solved by the invention]

[0006] Typically, the correct formulation of the recipe used for a composite material must first undergo several validation steps in the laboratory to find the appropriate technological package, and then optimize the formulation by step-by-step fine-tuning until the objective is fully achieved.

[0007] Each such iterative experimental campaign increases product development lead times and costs (time to market) from a product perspective, and from a data perspective produces a database with inherent variability due to random noise in the measurements made during the various test campaigns.

[0008] Predicting product performance under these conditions typically requires extensive laboratory testing to achieve compound validation, which is time-consuming and resource-intensive.

[0009] It is therefore an object of the present invention to solve these problems left unsolved by the prior art, by providing a process as defined in claim 1.

[0010] In particular, it is an object of the present invention to simulate laboratory tests to provide accurate estimations of some of the important viscoelastic and processability properties of composite materials for the manufacture of rubber compounds for tires, without the need to perform any physical tests.

[0011] A further object of the invention is a composite material analyzing apparatus (RPA) as claimed in claim 8.

[0012] Further features of the invention are defined in the corresponding dependent claims.

[0013] Using software tools that can predict composite material behavior and therefore tire performance, it is possible to: · Significant reduction in operating costs (raw materials, labor costs, etc.). Optimizing laboratory testing capacity and quality (allowing personnel to be allocated to other activities). · Reduced time to market for new products. · Improved forecast accuracy for known methodologies.

[0014] Other distinct advantages over the prior art, together with the characteristics and applications of the present invention, will become apparent from the following detailed description of preferred embodiments thereof, given purely as non-limiting examples. [Brief description of the drawings]

[0015] With reference to the drawings in the accompanying drawings: [Figure 1A] FIG. 1 is a schematic block diagram of the training steps of the machine learning and trait prediction algorithm according to the present invention. [Figure 1B] FIG. 1 is a schematic block diagram of the training steps of the machine learning and trait prediction algorithm according to the present invention. [Diagram 2] 1 shows a graph useful for verifying the time required to reach a given increase in vulcanization torque.

[0016] [Theoretical background] Polymer matrix composites are unique materials that exhibit characteristic properties of both elastic and viscous responses when subjected to stress.

[0017] Prediction of the rheometric curve of a composite based on one of its basic parameters (e.g. torques ML and MH, T10, T50 and T90, scorch time, vulcanized and unvulcanized shear moduli G' and tand at fixed shear conditions) is fundamental to determine and evaluate the workability of the composite from mixing to extrusion and vulcanization steps and to avoid problems such as mixer downtimes, defects in the extruded products, clogging of the press during vulcanization or under / over curing.

[0018] The processability properties are evaluated by carrying out rheometric tests in several steps. Some of the process parameters tested in these factories have recently been linked to performance parameters by specific evaluations, so that their prediction becomes even more important in order to estimate the variability of the factory performance. Such evaluations require several laboratory tests to arrive at the validation of the composite, which requires time and resources.

[0019] On the other hand, using a digital predictor makes it possible to: · Reduction of operating costs (raw materials, labor costs, etc.). Optimization of the test laboratory workload and quality (allowing personnel to focus on other activities). · Reduced time to market for new products.

[0020] Thus, the potential end users are all engineers and laboratory professionals who could benefit from this device.

[0021] As anticipated in the above paragraph, processability testing equipment is also available in factories to monitor composite materials to meet quality standards. The present invention may also be extended and released to factories as end users, allowing factory technical services to evaluate changes in compound properties during compound development in a much shorter time to resolve potential problems in factories with limited production downtime / loss, or simply to improve processability response in R&D.

[0022] The Rubber Process Analyzer (RPA) is a valuable instrument designed to measure the viscoelastic or processability properties of polymers and composites before, during and after vulcanization. Vulcanization properties can be determined by measuring the properties as a function of time and temperature. Tests can be performed at different conditions depending on the test method required, and measurements of G' and tand can be recorded continuously as a function of time and / or strain applied by cyclic torque at different shear rates.

[0023] Some of the outputs from each test were selected in ranges based on their cardinality in the data set and their importance to the engineers in assessing manufacturability. 1.ML, low torque of the vulcanization curve, modulus of elasticity of the green (unvulcanized) compound. 2. MH, the maximum torque of the vulcanization curve or the torque when the curve rises to a plateau, the elastic modulus of the vulcanized compound. 3. T10, T50, and T90 are the times to reach the homolog percentage of ML+ΔT, where ΔT is the torque difference between MH and ML. 4. Ts, i.e. scorch time, the time to reach the vulcanization setting + 1 dNm torque increment (see Figure 2). 5. Strain before vulcanization G'@100%, which is a parameter related to the viscosity of the green (unvulcanized) composite. 6. Strain after vulcanization G'@1%, strain tand@15%, which are related to the dynamic properties of the vulcanized compound used to predict tire tread performance. 7. Strain G'@50% after vulcanization, related to high strain static properties. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] The present invention will now be described with reference to the above figures.

[0025] Thus, a methodology is described for predicting the viscoelastic or processability properties (e.g., ML and MH torque, T10, T50 and T90, scorch time (Ts), vulcanized and unvulcanized strain modulus G' and tand under applied strain) of composite materials that can be used in the production of rubber compounds for tires.

[0026] In general terms, a methodology for predicting processability properties such as ML and MH torque, T10, T50 and T90 cure times, scorch time, shear modulus, and tand under applied strain for new composite materials that have never been previously mixed is described according to the following procedure. · Create a database by collecting existing recipes and their corresponding viscoelastic or processability properties (i.e., ML, MH, T10, T50, T90, TS, G'@100%, G'@1%, G'@50%) to obtain a primary data set. · The procedure of integrating the primary dataset through data augmentation to obtain an augmented dataset. · The procedure of transforming the augmented data to obtain a transformed data set. Training algorithms based on machine learning (e.g. linear mixed models) with the transformed dataset. · Prediction of viscoelastic or processability properties of new recipes to be tested using trained algorithms.

[0027] In particular, after a dedicated "step of training the model" of a machine learning algorithm is performed on the augmented and transformed dataset, it is possible to predict processability properties with greater accuracy than by directly applying the algorithm to the original dataset.

[0028] In fact, this method can significantly reduce the impact of random noise in the database on prediction accuracy.

[0029] In particular, after the step of training the model using the data set contained in the expanded and transformed database as described above, it is possible to predict the viscoelastic and processability properties of the composite material with greater accuracy than by applying the algorithm directly to the raw data of the database.

[0030] In fact, doing so can significantly reduce the impact of database noise and inherent variability in the data on prediction accuracy.

[0031] This device is characterized by the implementation of the following actions and algorithms: 1. Collection of laboratory data representative of the composite recipe to be tested. 2. Data Augmentation Procedure: This procedure allows enhancing the predictive capabilities of machine learning algorithms by utilizing various techniques for introducing new characteristic parameters. 3. Augmented Data Transformation Step: This step can enhance the predictive capabilities of machine learning algorithms by leveraging various transformation techniques. 4. Machine learning algorithms (e.g., linear mixed model algorithms): These algorithms aim to predict the processability properties of rubber compounds. The steps of the method according to the present invention will now be described with reference to the exemplary diagrams of FIGS. 1A and 1B.

[0032] This process includes the following steps following the generation of the primary data set, as previously described. 1. Integration of data from the primary dataset (data augmentation): This step aims to develop new characteristic parameters that must be added to the primary dataset (including, for example, the list of ingredients and their quantities for each recipe) to generate an augmented dataset. This means that for each recipe, a set of new characteristic parameters is estimated in order to improve the predictive ability of the machine learning model. The new characteristic parameters result from the application of a set of non-homogeneous techniques, such as clustering algorithms, non-linear operators applied to the input set, and dedicated and customized artificial neural networks, as described below.

[0033] 2. Transformation of the data contained in the extended dataset: Similar to the previous step, this one aims to improve the forecasting performance. The difference is that this step does not focus on adding new characteristic parameters, but on their transformation. The adopted data transformation algorithm is based on spline functions and Box-Cox transformation, as detailed below.

[0034] 3. Training step of the machine learning algorithm: The data of the augmented and transformed dataset (the compound recipes and the corresponding characteristic parameters) are provided as input for a predictive model implemented by a machine learning technique (e.g., a mixed linear model). The internal parameters of the model are then adjusted to fit the data (training step).

[0035] 4. Prediction step: Once the training is complete, the algorithm is able to generalize the results so that it is suitable to predict the processability property values ​​of composites for new recipes of composites that have never been mixed before (prediction step).

[0036] The original dataset to which the dataset integration (data augmentation) and transformation (dataset transformation) procedures have been applied is defined as the preprocessed dataset.

[0037] (Data integration procedure for primary data sets) Additional features have been introduced to improve the prediction of composite processability properties. In fact, it has been observed that the information available from the composite formulation is insufficient to achieve the goal of property prediction.

[0038] For example, the mixture terms introduced as characteristic parameters have been observed to play a key role and be very informative, and therefore have been used to augment the original dataset.

[0039] On the other hand, models formed with too many parameters are characterized by a lack of generalization in predictions: they may predict correctly on the training dataset, but not on new data, i.e. on the operation of a new test dataset. This phenomenon is called overfitting, and when it occurs, it becomes impossible to generate a model that can be run in the production step.

[0040] For this reason, a technique called "data augmentation" has been developed to improve predictive performance while avoiding overfitting.

[0041] The data augmentation procedure involves estimating and integrating one or more of the following characteristic parameters into the primary data set: <Estimation of the Gini coefficient for each recipe>: Also known as the Gini index or Gini ratio, it is a measure of statistical dispersion and is applied directly to each recipe. This characteristic parameter was added to provide a description of the hidden structure of the recipe.

[0042] <Define the mixing category for each recipe>: The category of the method of mixing the recipe of the composite material was introduced as a new characteristic parameter of type category. This characteristic parameter was added to provide information about the mixing conditions of the components (e.g. geometric parameters of the mixer itself, temperature, time and rotor speed) on which the predicted physical property influence is directed. While a more comprehensive way to provide this information could be achieved by introducing all the mixing parameters directly into the dataset, our strategy is to use only the parameters of the category type in order to provide a lean solution and avoid overfitting. In doing so, all the mixing parameters are summarized by a single parameter of type category indicating the type and thus the mixing category type applied, e.g. tangential or interpenetration screw mixer.

[0043] <Define composite application type for each recipe>: The application of the composite (e.g., the category of tire components or vehicle on which the tire will be mounted) has been introduced as a new categorical characteristic parameter. This characteristic parameter was added to provide information about the final application of the composite and to provide a categorical preview of the expected macro requirements when the composite is developed. All expected macro requirements are then summarized by a category type parameter that indicates the type of application for which the composite will ultimately be used. In fact, various applications require different macro requirements. A typology of application of a composite material could be, for example, for automobiles or commercial vehicles, which have different conditions of pressure and temperature during vulcanization, etc.

[0044] <Interaction between mixing method and composite application type for each recipe>: All possible combinations between the mixing category and the composite application are estimated and used to define a new characteristic parameter of category type for each recipe. This characteristic parameter of category type is introduced to explicitly describe the possible influence of the mixing conditions on the composite application, i.e. the macro-requirements assumed for the composite. In fact, the two variables of category type are connected, since the mixing category can affect the physical properties of the composite and the type of composite application provides a description of the macro-requirements of the composite. This connection is described through this new categorical variable that represents all possible combinations between the mixing category and the composite application.

[0045] <Estimated total amount of ingredients for each recipe>: This represents the sum of all the amounts of ingredients, estimated for each recipe. This way the amount of each ingredient is always related to the total amount, which varies from recipe to recipe.

[0046] <Estimation of ingredient ratios for each recipe>: An exhaustive search was conducted to determine how different ingredient ratios correlate with the predicted processability properties. Therefore, estimated ingredient ratios that were shown to correlate with the target properties were added as new property parameters. These property parameters were added to provide an explicit description of the nonlinear component terms that are beneficial in predicting the target properties.

[0047] · <Louvain grouping>: The method by the Louvain method was used to group recipes based on the co-occurrence of components. In fact, the Louvain unsupervised algorithm was attached to the recipes of the compounds that make up the available dataset. As an example, the recipes can be grouped according to the presence (co-occurrence) of synthetic rubber and silica rather than the co-occurrence of natural rubber and carbon black. Therefore, each recipe was grouped using the Louvain method, and as a result, new categorical characteristic parameters were estimated for the assigned clusters. This categorical parameter corresponds to the identifier of the grouping / cluster itself. This characterization parameter was introduced to provide an explanation for the non-uniformity of the primary dataset. In fact, the clustering algorithm can group different recipes based on the co-occurrence of components. This means that by integrating such grouping information into the dataset, it may be possible to explicitly provide insights into how components are used in different recipes.

[0048] · <Grouping by K-means method>: Each recipe was grouped using the method by the K-means method, and as a result, categorical type characteristic parameters were estimated for the assigned clusters. The categorical parameter corresponds to the identifier of the grouping itself. The grouping algorithm by the K-means method was attached to the recipes of the composite materials that make up the available dataset. Optimization by grid search was implemented to find the optimal hyperparameter optimization settings, that is, the number of clusters to identify or the optimization algorithm to implement. This characteristic parameter was introduced to provide an explanation for the non-uniformity of the primary dataset. In fact, the clustering algorithm can group different recipes according to the criterion of the distance in the space of components. That is, by introducing such grouping information into the primary dataset, it is possible to explicitly provide insights into how components are used in different recipes.

[0049] ·<Dimensionality reduction via autoencoders>: Autoencoders (AEs) are a special type of unsupervised artificial neural network trained to copy the input to the output. To achieve this, first the AE can map the input to a small latent space, then the AE encodes the latent representation to the output. As a result of this operation, the AE is trained to compress the data by reducing the reconstruction error. To find the optimal dimensionality of the compressed data, i.e. the dimensionality of the latent space, a grid search optimization was used. So-called dimensionality reduction algorithms using autoencoders were developed to process the set of data formed by the recipes of composite materials to create a dimensionally reduced, i.e. compressed, representation. This compressed data, which has reduced the dimensionality of the original recipe, is a very useful representation, so it was introduced as a new characteristic parameter.

[0050] According to a preferred embodiment, all of the above characteristic parameters are combined into a primary data set.

[0051] (Extended Data Conversion) Proper transformation operations before using it to train machine learning algorithms can significantly improve predictive accuracy.

[0052] The procedure for transforming the data present in the extended data set involves using one or more of the following transformation functions on the characterizing components and / or parameters: B-spline smoothing: B-spline functions were used to smooth the numerical data of the extended data set before the next training step. A grid search optimization algorithm was used to select the components and / or characteristic parameters to transform.

[0053] · Box-Cox transformation: The Box-Cox transformation was applied to the numerical data in the augmented dataset to make the distribution closer to normal.

[0054] Scaling transformation: All numerical data in the augmented dataset was scaled to the same numerical range in order to improve the subsequent training steps of the machine learning model.

[0055] Applying one or more of the above transformation functions produces a transformed data set.

[0056] According to a preferred embodiment of the present invention, all of the above transformation functions are used to process the recipe of the primary data set.

[0057] Moreover, according to a further embodiment, the aforementioned transformation functions constituting the data transformation procedure should preferably be performed according to the proposed order.

[0058] According to these embodiments, all the described steps are introduced in the pre-processing pipeline because their synergistic interactions can maximize the predictive performance.

[0059] The transformed (pre-processed) data set thus obtained, after carrying out the data augmentation and transformation procedures, can be used to start the more general training step of the machine learning model. According to the embodiments described herein, a mixed linear model was implemented and trained to provide a predictor of processability properties. Nevertheless, as soon as the primary data set is processed as described, any other machine learning model may be used to make the final prediction.

[0060] Finally, the same pre-processing steps (data augmentation and data transformation) are applied to the data related to or representative of the recipe of the composite material to be tested (FIG. 1B) before being fed as input to an already trained machine learning algorithm to predict the viscoelastic or processability properties of the composite material to be tested.

[0061] The present invention has been described above with reference to its preferred embodiments. Purely by way of example, each of the technical features implemented in the preferred embodiments described herein can also be advantageously combined with other features in other ways than as described herein, to form other embodiments belonging to the same inventive concept, all of which are intended to fall within the scope of protection granted by the claims set forth below.

Claims

1. A method implemented using a computer to predict the viscoelastic or processability characteristics of a composite material tested for the manufacture of a tire tread compound, comprising: a) preparing a database of raw data to be used as a reference, i.e., a primary data set including existing recipes of composite materials and corresponding known viscoelastic or processability characteristics thereof; b) preprocessing the primary data set, comprising: i. integrating one or more characteristic parameters in the primary data set, the one or more characteristic parameters being selected from the following, namely: - the Dynkin coefficient of each recipe - the mixing category of each recipe - the type of application of each recipe - the total amount of materials of each recipe - the ratio of components of each recipe - grouping by the Louvain method in the recipe of the composite material - grouping by the K-means method in the recipe of the composite material, and - data with reduced dimensions through an autoencoder applied to the data set formed by the recipe of the composite material, thereby obtaining an extended data set; ii. a conversion step of the extended data set by applying one or more conversion functions to the components and / or numerical characteristic parameters of the extended data set, the one or more conversion functions being selected from the following, namely:

1. B-spline smoothing 2. Box-Cox transformation, and 3. scaling transformation thereby obtaining a converted data set; wherein the preprocessing step is performed by the above; c) training a machine learning-based algorithm using the data of the converted data set; d) applying the algorithm trained according to step c) to a series of data preprocessed according to step b) and representing the recipe of the composite material to be tested for predicting the viscoelastic or processability characteristics of the composite material to be tested. A method comprising the above steps.

2. The method according to claim 1, wherein the viscoelastic or processability characteristics include torque ML and MH, T10, T50 and T90, scorch time (Ts), vulcanized and unvulcanized shear modulus G', and tand under the imposed conditions.

3. The method according to claim 1, wherein the step of integrating one or more characteristic parameters in the primary data set includes the step of integrating all the parameters shown in step i.

4. The method according to claim 1, wherein the step of converting the extended data set includes the step of applying all the conversion functions shown in ii.

5. The method according to claim 4, wherein the conversion functions shown in ii. are executed sequentially.

6. The method according to claim 5, wherein the conversion functions shown in ii. are executed in the shown order (1., 2., 3.).

7. The method according to claim 1, wherein the algorithm based on machine learning is configured to apply a mixed linear model.

8. A composite material analysis apparatus (RPA) configured to perform the method according to any one of claims 1 to 7.

Images