Method for producing at least one electrode of a battery cell

Abstract

A method for producing at least one electrode for a battery cell, the method comprising at least the following steps: a) providing a suspension for producing the at least one electrode, wherein at least one target process parameter (1) can be specified for providing the suspension; b) detecting at least one actual process parameter (2) during the provision in step a); c) performing a prediction of at least one property (4) of the suspension with the aid of a machine learning prediction algorithm (5), the machine learning prediction algorithm estimating the at least one property (4) of the suspension based on the at least one actual process parameter (2) and taking into account information about previously provided suspensions; d) specifying at least one target process parameter (1) for the provision in accordance with step a) based on the prediction result in step c).

Description

Technical Field

[0001] The present invention relates to a method for producing at least one electrode for a battery cell, a method for training a prediction algorithm with machine learning capabilities, a computer program for executing the corresponding method, and a machine-readable storage medium on which the corresponding computer program is stored. The present invention can be used, in particular, for (online) quality prediction during the production of battery suspensions, for example, for electrodes of lithium-ion batteries. Background Art

[0002] This quality, which can be described, for example, by properties such as the viscosity and / or particle size distribution of the battery suspension used to produce cathodes and / or anodes for lithium-ion battery cells, has a significant impact on the quality of the (final) product. To avoid costly waste and / or to improve process understanding, it is necessary to provide a method by which the most reliable possible information about the quality of the resulting suspension and / or even the finished product produced therefrom can be obtained even during the ongoing production process. In particular, efforts are being made to achieve inline-capable predictive capabilities during the extrusion process.

[0003] However, predictions based on physical models have proven to be too imprecise, particularly due to the high complexity of the extrusion process and / or the large number of process parameters and potentially unknown interfering variables. Expert predictions have been the most commonly used option to date. However, these predictions are subject to individual subjective perceptions and are therefore also unreliable. Furthermore, a disadvantage is that the process understanding built up over many years may no longer be available when the relevant experts leave the company.

[0004] Various methods for producing at least one electrode of a battery cell are known from US Pat. No. 9,806,326 B2, US Pat. No. 1,018,1587 B2, and WO 2016 / 073438 A1. However, the known methods do not allow, in particular, to predict the quality of the battery suspension. Summary of the Invention

[0005] Based on this, the object of the present invention is to at least partially solve the problems described in conjunction with the prior art. In particular, the invention is to specify a method for producing at least one electrode of a battery cell and a method for training a prediction algorithm with machine learning capabilities, which allow for a machine-based, inline, and as reliable a prediction as possible of at least one property of a suspension used to produce the electrode. Furthermore, the prediction should be possible with minimal time expenditure and / or low cost.

[0006] These objects are achieved by the features of the independent patent claims. Further advantageous embodiments of the proposed solution are described in the dependent patent claims. It should be noted that the features listed individually in these dependent patent claims can be combined with one another in any technically reasonable manner and define further embodiments of the invention. Furthermore, the features described in the patent claims are explained and illustrated in greater detail in the specification, which presents further preferred embodiments of the invention.

[0007] A method for producing at least one electrode of a battery cell contributes to this goal, the method comprising at least the following steps:

[0008] a) providing a suspension for producing the at least one electrode, wherein at least one target process parameter can be specified for providing the suspension;

[0009] b) detecting at least one actual process parameter during the provision in step a);

[0010] c) performing a prediction of at least one property of the suspension by means of a machine learning prediction algorithm, which estimates the at least one property of the suspension as a function of the at least one actual process parameter and taking into account previously provided information about the suspension;

[0011] d) specifying at least one target process parameter for providing according to step a) based on the prediction result in step c).

[0012] To perform the method, steps a) to d) can be performed at least once, for example, in the order described. Furthermore, steps a) to d) can be repeated (multiple times), or the method can be repeated (in a loop) starting with step a). At least some of steps a) to d), in particular at least some of steps a) and / or b) and / or c), can be performed at least partially in parallel or simultaneously. The method can be performed using a prediction algorithm that has been trained using the training method also described herein.

[0013] This method advantageously enables a machine-based, in-line (usable in mass production and, in particular, during an ongoing process) and highly reliable prediction of at least one property of a suspension used to produce an electrode. Furthermore, the machine-based prediction can be performed with minimal time expenditure and / or low cost. The use of this machine-learning prediction algorithm advantageously allows for a high level of information to be taken into account, including historical information about previously provided suspensions and / or information about previously occurring prediction results for the ongoing process. Furthermore, the method can advantageously facilitate in-line improvements to (target) process parameters, for example by verifying them using suspension properties measured during the ongoing process.

[0014] In step a), a suspension for producing the at least one electrode is provided, wherein at least one target process parameter can be specified for providing the suspension. This provision can be achieved, for example, by extruding the suspension. The at least one target process parameter can be set or specified, for example, via a human-machine interface of an extrusion system or production system. The at least one target process parameter can be, for example, one or more of the following parameters: dosage, screw speed of the extruder screw, and / or temperature in the extruder.

[0015] In step b), at least one actual process parameter (and optionally at least one ambient condition) is detected during the provision in step a). This detection can be performed, for example, by sensor measurement. The at least one actual process parameter generally describes the actually achieved value of the target process parameter. The at least one ambient condition can be, for example, one or more of the following: ambient temperature, ambient pressure, (air) humidity, and / or time of day.

[0016] Furthermore, at least one target process parameter and / or at least one raw material property of the raw material used to produce the suspension can also be detected in step b). In particular, information about a temporally prior predicted outcome of the ongoing process can also be detected in step b). For verification purposes, the measured property of the suspension can also be detected (measured) or read in in step b).

[0017] In step c), a prediction of at least one property of the suspension is performed using a machine learning prediction algorithm, which estimates the at least one property of the suspension as a function of the at least one actual process parameter (and, if necessary, the at least one environmental condition) and taking into account previously provided information about the suspension. The at least one property of the suspension estimated using the prediction algorithm may, for example, be one or more of the following: viscosity, density, shear rate, and / or homogeneity.

[0018] If corresponding information is detected, it can also be provided that the prediction algorithm (additionally) also estimates at least one property of the suspension based on one or more of the following information: at least one target process parameter, at least one raw material property of the raw material used to provide the suspension and / or at least one information about a temporally preceding prediction result of an ongoing process.

[0019] Information about the previously provided suspension can be learned, in particular, during an (initial) training phase. In particular, the information learned during the (initial) training phase can be represented, for example, by correspondingly designing (or adapting) and / or linking elements of the algorithm. These elements can be, for example, model parameters of the algorithm, such as weights, functions, thresholds, and so on. The algorithm can be implemented via and / or within an (artificial intelligence or KI) model. Furthermore, the algorithm can also include multiple parts or sub-algorithms, which can, for example, operate in parallel on a single level and / or on top of each other across multiple levels and / or one after another over multiple time steps.

[0020] For example, the algorithm can be configured to map a set of input data (or input data) to at least one output or at least one set of output data (or output data). The input data to the algorithm typically includes at least one (time-varying or time-dependent) actual process parameter and, optionally, at least one (time-varying or time-dependent) environmental condition. Optionally, these input data may also include: at least one (time-varying or time-dependent) target process parameter; and / or at least one (possibly time-invariant or time-independent) raw material property; and / or at least one piece of information about a previously provided suspension; and / or information about a previous time step, in particular, a previous prediction result of the algorithm over time; and / or a measured property of the suspension (for validation purposes). The at least one output typically includes a prediction result, i.e., at least one (predicted and, in particular, time-varying or time-dependent) property of the suspension and / or at least one (predicted) final property. In particular, during an ongoing process (i.e., while providing, for example, extruding, a suspension), a machine prediction of at least one property of the (just provided) suspension and / or, if appropriate, even of the finished suspension is performed at fixed time intervals and / or at at least one specific (future or subsequent) point in time (e.g., as a prediction of the final property that will occur when the ongoing process ends, or when a possible container is filled to a predeterminable level and / or when the suspension cools). In other words, this can also be described as allowing the prediction to be performed such that at least one property of the finished suspension or at least one (discrete) product property at the final point in time can be estimated. In particular, the prediction can be performed based on the process time. For example, the prediction can be performed for the (immediately following) next time step and / or for all time steps until the end of the process. If a set of output data is output, various (physical) properties of the suspension can be output, such as, for example, its viscosity and shear rate. For example, the data set can be provided in the form of vectors, such as at least one input vector and at least one output vector.

[0021] This prediction algorithm can be formed, for example, in the form of a so-called machine learning model. It can include, for example, modeling the production process using a so-called gated recurrent unit. Alternatively or cumulatively, it can include modeling the process using a so-called multilayer perceptron. Here, a feature vector or input vector can, for example, contain all time-varying (input) signals up to a specific point in time. If this point in time is reached during the process, the result of a future measurement can be predicted (once).

[0022] In step d), at least one target process parameter is specified based on the prediction result of step c) for provision according to step a). In principle, this specification can be performed manually, for example, by a person responsible for the process. Advantageously, this specification can be performed automatically, for example, by a process control system that has access to and / or includes the prediction algorithm. In particular, if the prediction result indicates a violation of a tolerance limit, the at least one target process parameter can be adapted or modified.

[0023] According to an advantageous embodiment, it is provided that the method is carried out to produce at least one anode or cathode of a lithium-ion battery. In this respect, the suspension can, for example, include one or more of the following raw materials: storage time, particle size distribution, and storage conditions (such as temperature and humidity).

[0024] According to another advantageous embodiment, the suspension is provided by extruding it into a container. The container can, for example, have an internal shape that corresponds to the external shape of the electrode to be produced. Alternatively, the container can simply be used for transport to a downstream production step (foil coating). This provision can, for example, continue until a predeterminable filling level of the container is reached. The final properties of the suspension typically appear when or after the predeterminable filling level of the container is reached. If the final properties appear after the predeterminable filling level is reached, this particularly relates to the properties after a (predeterminable) cooling process.

[0025] According to another advantageous embodiment, the machine learning prediction algorithm is formed using an artificial neural network. This network typically includes elements or model parameters that map input data to output data. The corresponding elements or model parameters may include, for example, nodes, weights, links, thresholds, and so on. A so-called recurrent neural network (RNN) is an example of an artificial neural network.

[0026] According to another advantageous embodiment, the machine learning prediction algorithm is implemented using a long short-term memory network (LSTM network, or LSTM for short). The basic structure of such networks is well known from other application areas, particularly from natural language processing, so it will not be discussed in detail here. LSTM networks are particularly suitable for mapping time-series-based algorithms. LSTM networks are particularly favored over alternative algorithms due to the following advantages: They can better model time-varying (longer) signals than gated recurrent units or recurrent neural networks, particularly due to faster convergence in such cases thanks to improvements in the gradient descent method. In contrast to LSTM networks, multilayer perceptrons generally lack information about previous inputs and outputs. Furthermore, LSTM networks can particularly advantageously (and relevant to the described use case) model the time-varying nature of continuous process parameters, environmental conditions, and time-invariant material properties.

[0027] According to another advantageous embodiment, at least one property of the suspension is measured in order to verify the performed prediction. Furthermore, the measured (time-varying or time-dependent) property of the suspension can be used to further train or improve the prediction algorithm ("inline," i.e., during normal or ongoing operation of the prediction algorithm).

[0028] According to another aspect, a training method for a prediction algorithm with machine learning capabilities is proposed, the training method comprising at least the following steps:

[0029] i) reading in training input data for the prediction algorithm, which training input data include actual process parameters and optionally environmental conditions of a plurality of provision processes for providing a respective suspension for producing electrodes for battery cells;

[0030] ii) reading in training output data for the prediction algorithm, which training output data include the properties of the suspension provided by means of a corresponding provision process;

[0031] The elements, preferably the weights, of the prediction algorithm are adapted in order to map the incoming training input data onto the incoming training output data as precisely as possible.

[0032] To perform the method, steps i) to iii) can be performed at least once, for example, in the order described. Furthermore, steps i) to iii) can be repeated (multiple times), or the method can be repeated (in a loop) starting with step i). At least some of steps i) and ii) can be performed at least partially in parallel or simultaneously. This method can be used for machine learning of the prediction algorithm also described herein.

[0033] According to one advantageous embodiment, it is proposed that a gradient descent method be used or executed to adapt the elements of the prediction algorithm. For example, during training, the elements of the algorithm or the model parameters can be adapted by means of the gradient descent method until the prediction error is minimized across all training examples.

[0034] According to another aspect, a computer program for executing the method described herein is also provided. In other words, this relates in particular to a computer program (product) comprising instructions which, when executed by a computer, cause the computer to carry out the method described herein.

[0035] According to another aspect, a machine-readable storage medium is also provided, on which the computer program is stored. The machine-readable storage medium is typically a computer-readable data carrier.

[0036] Furthermore, a production system for producing at least one electrode of a battery cell may be described, which is configured and set up to carry out the method described herein. For this purpose, the production system may, for example, include a controllable extruder, a (possibly controllable) raw material feed, sensors for detecting at least some of the aforementioned input data, and / or a container into which the suspension can be extruded. The production system particularly includes a control device on which (electronic or digital) control for carrying out the method described herein is implemented. The control device may, for example, include a computing unit (controller) that can access the storage medium and / or execute the computer program to carry out the method described herein. The storage medium and / or the computer program may also be components of the control device or be connected to it via signal technology.

[0037] The details, features, and advantageous embodiments discussed in conjunction with the production method may also be present in the training method, the computer program, the storage medium, and / or the production system presented herein, and vice versa. In this respect, reference is made in full to the statements there for a more detailed description of these features. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The solution presented here and its technical environment will be explained in more detail below with reference to the drawings. It should be noted that the present invention is not limited to the exemplary embodiments shown. In particular, unless explicitly stated otherwise, it is also possible to extract partial aspects of the facts illustrated in or in conjunction with the drawings and combine them with other components and / or knowledge from other drawings and / or this description. Among these:

[0039] Figure 1 A flow chart is schematically shown for illustrating an embodiment of the method described herein;

[0040] Figure 2 Schematic illustration of an embodiment of the prediction algorithm described herein;

[0041] Figure 3 schematically illustrates an illustration of possible prediction results of a prediction algorithm; and

[0042] Figure 4 A flow chart is schematically shown for illustrating an exemplary embodiment of the training method described here. DETAILED DESCRIPTION

[0043] Figure 1 A flow chart is schematically shown for illustrating an exemplary embodiment of the method described herein for producing at least one electrode of a battery cell. The sequence of steps a), b), c), and d) presented using blocks 110 , 120 , 130 , and 140 is exemplary and may occur, for example, in the normal course of the method.

[0044] In block 110, according to step a), a suspension for producing the at least one electrode is provided, wherein at least one target process parameter 1 can be specified for providing the suspension. In block 120, according to step b), at least one actual process parameter 2 and, exemplarily, at least one ambient condition 3 are detected during the provision in step a). In block 130, according to step c), a prediction of at least one property 4 of the suspension is performed using a machine learning prediction algorithm 5, which estimates the at least one property 4 of the suspension based on the at least one actual process parameter 2 and, exemplarily, the at least one ambient condition 3, and taking into account information about the previously provided suspension. In block 140, according to step d), at least one target process parameter 1 is specified for the provision according to step a) based on the prediction result in step c).

[0045] For example, the method can be used to produce at least one anode or cathode of a lithium-ion battery. In this regard, provision can also be made for the suspension to be provided by extruding it into a container. The extruder typically produces the battery suspension continuously. During extrusion, a machine prediction of the properties 4 of the suspension just introduced into the container and / or, if necessary, even of the finished suspension can be made, for example, at fixed time intervals (e.g., as a prediction of the properties that will occur when the container is filled to a predeterminable level and / or when the suspension cools). In particular, the method allows for a highly accurate prediction of the properties 4 that the battery suspension will have or is expected to have at the end of the extrusion process (i.e., when the container is filled to a predeterminable level and / or when the suspension cools). This particularly advantageously enables (inline) adaptation of target process parameters during the ongoing extrusion process if the predicted properties 4 do not correspond to desired or predeterminable target values.

[0046] Figure 2 The following schematically illustrates an example of an embodiment of the prediction algorithm described herein. Reference numerals are used uniformly so that reference can be made to the previous explanations.

[0047] The machine learning prediction algorithm 5 can be formed, for example, by means of an artificial neural network. In this respect, Figure 2 This example illustrates a prediction algorithm using a long short-term memory network (LSTM network). The variable "x" represents the input data or input vector; the variable "h" represents the (possibly hidden) output data or (possibly hidden) output vector; the variable "c" represents the cell state of the LSTM network; the subscripts "t-1," "t," "t+1," and so on represent the different time steps; the subscripts "1 ... N" represent the different input data at the corresponding time; and the subscripts "1 ... M" represent the different output data at the corresponding time.

[0048] By way of example, the course of input data "x" is depicted by a time-varying signal. In this respect, the input data "x" at the respective point in time includes actual process parameters 2, optionally environmental conditions 3, and optionally (already specified) target process parameters 1 and / or optionally known raw material properties 8. In the illustrated embodiment, the relationship between the aforementioned input data and output data "h" is described using modeling using a long short-term memory (LSTM) network. In this respect, the output data "h" at the respective point in time includes the predicted properties 4 of the battery suspension (depending on the process time). In other words, this can also be described as the input vector of a feature vector "x" being fed into the LSTM network. This feature vector contains all (actual) process parameters and, if necessary, environmental conditions at a discrete point in time. The LSTM network is then configured to approximate the properties 4 of the battery suspension at that point in time, i.e., the target vector.

[0049] During production, the quality of the suspension or (actual) properties 6 of the suspension (for comparison with the predicted properties 4) can optionally and additionally also be measured manually, for example at fixed time intervals, in particular until the container is filled to a predeterminable filling level (or completely filled). Advantageously, these manually measured properties 6 can help to verify the predictions of the prediction algorithm 5. These properties 6 are Figure 2 This represents an example of how at least one property 6 of the suspension can be measured and how, if necessary, the at least one property of the suspension can be measured in order to verify the predictions performed.

[0050] If measured characteristics 6 are also present, an error "L" can be determined, which can be used (either inline or after actual or initial training) to further train or improve the prediction algorithm. For example, in this regard, an LSTM network can be unrolled over time. Furthermore, the prediction error "L" and the gradient can be calculated for each time step. Furthermore, these gradients can be averaged over all time steps. Element 7 of Algorithm 5 can then be adapted based on this. Alternatively or cumulatively, the so-called Truncated Backpropagation Through Time method (TBPTT) can also be used to improve the (already initially trained) network.

[0051] Figure 3 The diagram schematically illustrates an illustration of possible prediction results of the prediction algorithm 5. Reference numerals are used uniformly, so that reference can be made to the preceding explanations.

[0052] In particular, it can be seen here that the time-varying modeling of (continuous) process parameters 2 , optionally environmental conditions 3 and optionally time-invariant (raw) material properties 8 can be performed with the aid of an algorithm 5 in the form of a long short-term memory network.

[0053] exist Figure 3 The upper left of the figure illustrates exemplary input data 2, 3, 8 and Figure 3 In the upper right corner of FIG, an exemplary illustration of output data 4 of algorithm 5 is illustrated. It can be seen that each time step is represented by a successive unit of algorithm 5. The algorithm 5 can usually read in input data 2, 3, 8 at the corresponding time step - these input data include, for example, actual process parameters 2, environmental conditions 3 and, if necessary, raw material properties 8 (and, if necessary, target process parameters 1, see Figure 2 )—and historical data or information about previously provided suspensions. The historical data or information about previously provided suspensions typically includes at least the information learned during a training phase from a training dataset, which is typically based on historical data or information about previously provided suspensions. This information learned during the (initial) training phase can be represented, for example, by corresponding design and / or linking of element 7 of algorithm 5.

[0054] exist Figure 3 It is also explained that historical data or information about the previously provided suspension can also include, for example, information from the (immediately) previous time step of the algorithm 5 (here, exemplarily, the upstream unit of the network). Furthermore, actual properties 6 of the suspension, measured for validation purposes in particular, can be fed into the algorithm 5 as input. These properties 6 can also contribute to further training and / or improvement of the algorithm 5 ("inline," i.e., during normal or ongoing operation of the algorithm).

[0055] exist Figure 3 The figure also shows that predictions can be made not only for a specific future time point (here, exemplarily, t_x+0.2), but also, if necessary, for multiple future time points, which can then be verified step by step using characteristics 6. This allows, for example, an estimate of a (continuous) product characteristic 4 at the respective time point t. In particular, a prediction for a final time point (here, exemplarily, "T") can also be made (as an alternative or cumulative addition to predictions for multiple future time points). In other words, this can also be described as allowing a prediction to be made that estimates at least one characteristic 4 of the finished suspension or at least one (discrete) product characteristic 4 at the final time point T.

[0056] This at least one property 4 of the finished suspension may also be referred to here as at least one "final" property 9 (symbol T) and may be determined, for example, according to Figure 3 The diagram in the lower middle section is used for monitoring. Here, the predicted final property 9 of the suspension at the corresponding time points or time steps during the process is plotted over time. This diagram can, for example, determine whether the predicted final property 9 of the suspension remains or has remained within a tolerance band 10 during the process or for which time intervals 11 the property 9 lies or has been outside of the tolerance band 10. For example, provision can be made for the target process parameters to be adapted, for example by the process manager or automatically, if a tolerance violation is predicted, based on the prediction and / or the resulting error.

[0057] Figure 4 Schematic diagram for illustrating the prediction algorithm 5 described here for machine learning capabilities (see Figure 2 and 3 ) is a flow chart of an embodiment of a training method for a training system. Reference numerals are used uniformly so that reference can be made to the previous explanations. The order in which steps i), ii), and iii) are presented using blocks 210, 220, and 230 is exemplary and may occur, for example, in the normal flow of the method.

[0058] In block 210, according to step i), training input data for the prediction algorithm 5 are read in. These training input data include actual process parameters 2 and, optionally, environmental conditions 3 of a plurality of preparation processes for preparing a respective suspension for producing an electrode for a battery cell. In block 220, according to step ii), training output data for the prediction algorithm 5 are read in. These training output data include properties 6 of the suspension provided by the respective preparation process. In block 230, according to step iii), elements 7 of the prediction algorithm 5 are adapted so that the read-in training input data are mapped as accurately as possible to the read-in training output data. For example, a gradient descent method can be used to adapt elements 7 of the prediction algorithm 5.

[0059] Advantageously, artificial intelligence (KI) methods can be used to predict the results of future measurements based on the historical (extrusion) process of the battery suspension and a data set from one or more measurements, preferably at multiple or possibly every time point in the (corresponding) process. In this regard, it is particularly advantageous to record and store the process parameters 1, 2, environmental conditions 3, and raw material properties 8 in a traceable manner. This database can serve as an exemplary training set for the (KI) algorithm 5.

[0060] During training, the algorithm 5 (here, exemplarily, an LSTM network) is fed with a number of training examples, typically consisting of a feature vector and a target vector. Furthermore, the randomly initialized model parameters of the LSTM network (here, exemplarily illustrated by element 7) can lead to inaccurate predictions at the beginning of training. During training, these model parameters can be adapted, for example, using gradient descent, until the prediction error is minimized across all training examples.

[0061] After training a model or algorithm for predicting the quality of a battery suspension, it can be used to make predictions during the (current) process for unknown combinations of process parameters 1, 2, optionally environmental conditions 3, and optionally material properties 8. Additional measurements of suspension properties 6 can help validate these predictions and / or improve the accuracy of predictions at later points in time. Due to the statistical significance of the large data volumes, the predictions of the algorithm 5 are advantageously more accurate than those of expert and physical models.

[0062] Thus, a method for producing at least one electrode for a battery cell and a method for training a prediction algorithm with machine learning capabilities are described, which at least partially solve the problems described in conjunction with the prior art. In particular, a method for producing at least one electrode for a battery cell and a method for training a prediction algorithm with machine learning capabilities are described, which enable a machine-based, inline-capable, and highly reliable prediction of at least one property of a suspension used to produce the electrode. Furthermore, the prediction can be performed with minimal time expenditure and / or at low cost.

[0063] List of Reference Numerals

[0064] 1 Target process parameters

[0065] 2 Actual process parameters

[0066] 3 Environmental conditions

[0067] 4 Features

[0068] 5 Prediction Algorithms

[0069] 6 Features

[0070] 7 components

[0071] 8 Raw material characteristics

[0072] 9 Features

[0073] 10 Tolerance band

[0074] 11 Time windows.

Claims

1. A method for producing at least one electrode of a battery cell, the method comprising at least the following steps: a) providing a suspension for producing the at least one electrode, wherein at least one target process parameter (1) can be specified for providing the suspension, wherein the at least one target process parameter is one or more of the following parameters: dosage, screw speed of the extruder screw and / or temperature in the extruder; b) detecting at least one actual process parameter (2) during the providing in step a), wherein the at least one actual process parameter describes the actually achieved value of the target process parameter; c) performing a prediction of at least one property (4) of the suspension by means of a machine learning prediction algorithm (5), said machine learning prediction algorithm estimating the at least one property (4) of the suspension as a function of the at least one actual process parameter (2) and taking into account previously provided information about the suspension, wherein the at least one property (4) comprises viscosity, density, shear rate and homogeneity; d) specifying at least one target process parameter (1) for providing according to step a) based on the prediction result in step c). 2 . The method according to claim 1 , wherein the method is performed to produce at least one anode or cathode of a lithium-ion battery.

3. The method according to claim 1 or 2, wherein said providing of said suspension is effected by extruding said suspension into a container.

4. The method according to claim 1 or 2, wherein the machine learning prediction algorithm is formed with the aid of an artificial neural network.

5. The method according to claim 1 or 2, wherein the machine learning prediction algorithm is formed with the aid of a long short-term memory network.

6. Method according to claim 1 or 2, wherein at least one property (6) of the suspension is measured in order to verify the prediction performed.

7. A training method for a machine learning prediction algorithm (5) according to any one of claims 1 to 6, the training method comprising at least the following steps: i) reading in training input data for the prediction algorithm (5), the training input data comprising actual process parameters (2) of a plurality of provision processes for providing a respective suspension for producing an electrode for a battery cell; ii) reading in training output data for the prediction algorithm (5), said training output data comprising the properties of the suspension provided by means of a corresponding provision process (6); iii) Adapting the elements (7) of the prediction algorithm (5) in order to map the read-in training input data onto the read-in training output data as accurately as possible.

8. Training method according to claim 7, wherein for adapting the elements (7) of the prediction algorithm (5) a gradient descent method is used. 9 . A computer program product comprising a computer program configured to execute the method according to claim 1 . 10 . A machine-readable storage medium having a computer program stored thereon, the computer program being configured to execute the method according to claim 1 .

Images