Patent Application
20220027759
The invention is directed at a method for improving prediction relating to the production of a polymeric product.
There is a wide range of different applications for polymeric products. Consequently, there are also wide-ranging and strongly varying different product property specifications that are prescribed for those polymeric products. These different specifications determine desired values or value ranges for weight, rigidity, emissions, temperature durability and many other features of the respective polymeric product. Further, the polymeric product properties do not only depend on the formulation, which specifies the material ingredients of the polymeric product, used for the production of the polymeric product, but also on different process parameters applied in the production process of the polymeric product, including properties of production tools used. The polymeric product properties may also depend on production parameters ambient variables during the production process of the polymeric product.
The usual process involves proceeding from a known formulation and known process parameters resulting in a product with known properties and then making adjustments based on prior experience and general considerations. Based on these adjustments, a number of sample products are produced and their relevant properties experimentally determined. From such observations, there may also be dependencies derived and expressed in terms of formulas. This then permits a prediction of product properties based on the formulation data or vice versa. However, the expression of such formulas presupposes that the person analyzing the dependencies thinks of the correct formula or types of formula—i.e. a hypothesis—to apply and test, i.e. to verify the hypothesis. It may happen that there are functional relationships which remain substantially elusive to such an approach because the proper hypothesis is not thought of.
WO 2005/019948 A1, from which the present invention proceeds, discloses molding and process control techniques for manufacturing products. Computer-aided modeling techniques are described that allow the manufacturer to predict a profile for a multivariate output that is necessary to achieve a target performance property for a manufactured product. Desired performance properties for a product are selected and based on the selected performance properties, product control center invokes the reverse chemometric model to predict the required multivariate output. Product control center determines the necessary operating parameters to achieve the predicted multivariate output and the product is produced according to the operating parameters.
Consequently, the object of the invention is to provide a method for improving prediction relating to the production of a polymeric product which is able to identify dependencies which are difficult or costly to find based on a purely analytical approach.
The object of the invention is achieved by a method for improving prediction relating to the production of a polymeric product with the features of claim 1.
The invention is based on the realization that reinforcement learning may be used to develop and improve a prediction model which, at least for some cases, may be more accurate and explorative than a prediction model generated either by purely analytic considerations or by other forms of machine learning such as supervised learning. In particular for complex, subtle and potentially counterintuitive correlations between product properties on the one hand and production parameters and formulation data on the other hand, such an evolutionary approach may be the most efficient way to achieve the desired prediction model. The resulting improvement in the prediction may then be used both for determining product properties based on production parameters and formulation data and for determining production parameters and formulation data based on desired product properties.
The method according to the invention is for improving prediction relating to the production of a polymeric product.
The polymeric product may be a polyurethane product or a polyisocyanurate product. Thus, the polymeric product properties may be polyurethane or polyisocyanurate product properties, the polymeric production may be polyurethane or polyisocyanurate production and the polymeric formulation may be polyurethane or polyisocyanurate formulation. Alternatively or in addition, the polymeric product may be a foam product. Thus alternatively or in addition, the polymeric product properties may be foam product properties, the polymeric production may be foam production and the polymeric formulation may be foam formulation.
Preferably, the polymeric product is a polycarbonate product. Accordingly, the polymeric product properties may be polycarbonate product properties, the polymeric production may be polycarbonate production and the polymeric formulation may be polycarbonate formulation.
In the method according to the invention, a prediction model is provided for describing a functional relationship between production parameters, which production parameters comprise formulation portions data specifying raw material portions used for the production of a respective polymeric product and comprise processing parameters data specifying machine process properties during the production of that polymeric product, and product properties data associated with that polymeric product on a computer system. In other words, the prediction model is either a computer program or a parameter set for a computer program, which computer program is executed on a computer device and which provides a way to determine at least some production parameters based and/or some product properties data based on given other production parameters and/or other product properties data. Thus for example, the prediction model may take at least the product properties data as input and provide at least the production parameters as output. It may also be the other way around. Any electronic device with a microprocessor presents a computer device in the present sense. The prediction model constitutes a part of what is understood as policy in the context of reinforcement learning
In the method according to the invention, the production parameters and product properties data form data entry properties for a respective polymeric product. The data entry properties may in addition to the production parameters and the product properties data comprise also further variables or data. In the present sense, a machine may be any kind of apparatus and in particular a lab apparatus for any or all steps of polymeric material production. In other words, the formulation portions specify what raw material is used and to what proportion for producing the polymeric product. This also relates to material such as blowing agents which may be used in the production of the polymeric product but substantially are not part of the polymeric product. The processing parameters may comprise settings for the machine or group of machines that processes the raw material in order to obtain the polymeric product. The processing parameters may also comprise constant characteristics of the machine or group of machines. These may include geometrical dimensions, maximum power etc. of the machine or group of machines.
Preferably, the production parameters comprise ambient parameters describing the environment during the production of the polymeric product. Such ambient parameters may be descriptive of the temperature, pressure, humidity, sunshine intensity and/or of other physical parameters of the environment.
In addition, the production parameters may also comprise formulation description data for describing dynamic behavior of raw material portions. For example, if the polymeric production is a foam production and the foam is created by the reaction of the raw materials according to the raw material portions, then the formulation description data may describe properties relating to that reaction. Thus, the formulation description data may specify starting time, rising time and/or setting time.
The product properties data may in principle relate to any physical or chemical property of the polymeric product. In particular, the product properties data may comprise density, compression characteristics, restoring characteristics, compression hardness, thermal conductivity, compressive strength, torsional stiffness and/or flame resistance.
Before the first pass of the prediction model, the prediction model may be empty in the sense that only a trivial or zero-order approximation relationship is provided. Alternatively, parameters of the prediction model are randomly set or populated. With successive iterations of the method, the prediction model is incrementally built up.
In the method according to the invention, user input is provided comprising user product targets specifying only a part of the data entry properties. In other words, the user input provides the above-mentioned user product targets, which may either describe at least some desired product properties data of the polymeric product and—as an alternative or in addition—describe at least some formulation portions data and processing parameters data. Thus, the user may specify product properties data and look for production parameters in order to produce a polymeric product with this product properties data. The user may also specify production parameters and want to determine the resulting product properties data. Still further, the user may specify some production parameters and some product properties data and obtain as output additional production parameters as well as additional product properties data for the corresponding polymeric product thereby defined.
The user product targets may also refer to value brackets instead of specific individual values. The user product targets may also be indirectly given, e.g. by reference to known product properties data of a known polymeric product.
In the method according to the invention, a new data entry is generated by the computer system for realizing the user product targets by applying the user product targets to the prediction model. The generation of the new data entry is at least partially based on applying the user product targets to the prediction model, i.e. by providing the user product targets as input to the prediction model. The above means that the resultant new data entry, also comprising data entry properties which in turn comprise product properties data and production parameters, is to be such that the user product targets are to be at least approximated by a polymeric product produced in accordance with data entry properties of the new data entry according to the prediction model. The degree to which those user product targets are actually realized may be different for different cases.
Applying the user product targets to the prediction model may also comprise the prediction model determining and in particular calculating derived or intermediate values based on the user product targets. These derived or intermediate values may then be used to calculate the new data entry according to the prediction model.
In the method according to the invention, for the new data entry the specified data entry properties are determined In the parlance of reinforcement learning, any new data entry with the data entry properties constitutes a state with associated values. This determination presents a verification of the projection of the prediction model and may be done in different ways, which shall be discussed in more detail below. In any case, this determination occurs substantially independent from the prediction model, since the goal is to verify and check the prediction model.
Further in the method according to the invention, a reward metric is determined by the computer system based on the determined data entry properties. Here, reward metric denotes a variable indicative of the extent to which the determined data entry corresponds to new data that is to be encouraged in some sense, i.e. which presents a step in the right direction. How such a right direction is defined is, in principle, arbitrary and several possibilities exist. Properties correspond to the user product targets and to which extent the prediction of the prediction model has been verified by the determination of the data entry properties. In the context of reinforcement, this is the reward, which may be positive or negative depending on the success of the prediction model.
In the method according to the invention, the prediction model is updated based upon the reward metric by the computer system. In other words, in the case of a high reward it is fed back to the prediction model that the generated new data entry was “right” as quantified by the reward. it has at least predicted in the correct direction. In the case of a low reward of compliance, the information of a generated new data entry in a false direction is fed back. This feedback results in a corresponding change and further refinement in the prediction model.
The computer system may comprise one or more computers of any kind which may be interconnected by any kind of network. The computer system may also be fully or partially implemented by a cloud computing environment.
In the method according to the invention, the new data entry is entered into a formulation database of data entries with data entry properties.
In the parlance of reinforcement learning, the generation of the new data entry is called an action.
In the method according to the invention, the new data entry is generated by selecting a base data entry from the formulation database and modifying at least one data entry property of the base data entry based on applying the user product targets to the prediction model. Thus, the new data entry—i.e. the new state—is a modification of an existing data entry, namely the selected data entry. Preferably, the base data entry from the formulation database is selected based on a similarity of the data entry properties of the base data entry with the data entry properties specified by the user product targets.
Preferably, modifying the at least one data entry property comprises changing the at least one data entry property by a modification value. This may be addition, subtraction, multiplication or division of the corresponding original value of the data entry property by the modification value.
In principle, the above modification may be done in a substantially deterministic manner. In a further preferred embodiment of the method, generating the new data entry comprises a pseudo-random or evolutionary determination. Such pseudo-randomness is particularly suitable for reinforcement learning. Here it is further preferred that selecting the base data entry and/or determining the modification value and/or choosing the at least one data entry property for modification comprises a pseudo-random or evolutionary determination.
In principle, the modification value may have an arbitrary value and the base entry may be arbitrarily selected. However, there may be limitations and boundaries on what states are permitted and—accordingly—what kind of actions are possible to obtain a new state. Any such rules, limitations or boundaries also form part of the policy in the parlance of reinforcement learning According to a preferred embodiment of the method, modification boundary parameters are provided, preferably by a user, that limit selection of the base entry. Thus, it may be that the base entry is restricted to a subset of the data entries of the formulation database. Alternatively or in addition, the modification boundary parameters limit which at least one data entry property can be chosen for modification. Thus, not all data entry properties may be available for modification. Such a limitation is particularly useful when it is known that the respective property is irrelevant for the matter at hand. Alternatively or in addition, the modification boundary parameters provide a numerical limit for the modification value. Thus, the modification may be restricted to small degrees if it is known that there exists great numerical sensitivity. In particular, they may provide a numerical limit for the modification value depending on the at least one data entry property chosen for modification. Thus, the potentially different sensitivity of different data entry properties may be considered.
In the method according to the invention the reward metric is a proximity metric based on compliance between the determined data entry properties and the user product targets and , the prediction model is updated such that for a proximity metric indicating a higher compliance, in particular with the criterion underlying the reward metric, the modification of the modified at least one data entry property is reinforced and that for a proximity metric indicating a lower compliance, the modification of the modified at least one data property is weakened.
In a further preferred embodiment of the method, the user product targets specify product properties data. Thus, production parameters for producing a polymeric product with the user product targets is sought. Here it is preferred, that the data entry is modified by modifying production parameters to generate the new data entry.
According to a preferred embodiment of the method, the user product targets specify production parameters. Then, the resulting product properties data of a corresponding polymeric product is sought. Preferably then the data entry is modified by modifying product properties data to generate the new data entry.
One preferred way of determining the specified data entry properties for the new data entry is by way of an experiment. Thus, according to a further preferred embodiment of the invention, for the new data entry the specified data entry properties are determined by producing a polymeric product according to the production parameters of the new data entry and by measuring at least some of the specified data entry properties. It is further preferred that the production parameters of the new data entry are applied to provide raw materials according to the formulation portions to a machine for polymeric production. It may in particular be that the machine process properties comprise user-settable machine process settings and that the production parameters of the new data entry are applied to select machine process settings in a machine for polymeric production, such that a polymeric product is produced by the machine from the raw materials. Thus, in this preferred embodiment the method culminates in the production of an actual polymeric product, which may then form the basis for reliable feedback to the prediction model.
An alternative for such measurement on a polymeric product are calculations based on computational model, which in turn may be based on known formulas. Thus in a preferred embodiment of the method, for the new data entry the specified data entry properties are determined by applying a computational model to at least some of the data entry properties of the new data entry. It is preferred that the computational model is a physical model. In addition or alternatively, it could also be a chemical model.
In general for reinforcement learning, a single iteration is insufficient for arriving at a sufficiently precise prediction model. Therefore in a further preferred embodiment of the method, a cycle of generating the new data entry, determining the specified data entry properties for the new data entry, determining the reward metric and updating the prediction model is repeated until a new data entry is generated for which the reward metric exceeds a predetermined limit. For example, when the reward metric is a proximity metric, the cycle is repeated until the prediction is sufficiently precise.
The predetermined limit may be comprised by the user input. In this way, the user can at least indirectly influence the precision to which the model is to be developed.
In the method according to the invention, the user input comprises a user selection of raw materials from a list of raw materials predefined in the computer system, thereby defining combinations of the raw materials for a polymeric formulation, and that for the new data entry the formulation portions specify raw material portions from the user selection of raw materials. This may be useful when the use of some raw materials is not possible or expeditious for economical, logistical or other reasons or the use of some other raw materials is especially preferred.
In principle, any raw material used for polymeric production may be specified by the formulation portions, even if it is not present in the finished polymeric product. According to a further preferred embodiment of the method, the user-selected raw materials comprise an isocyanate and a polyol. The user-selected raw materials may also comprise a plurality of isocyanates and/or a plurality of polyols. The user-selected raw materials may also comprise a blowing agent. Preferably, the user-selected raw materials further comprise a chain extender, a cross linker, a catalyst for accelerating the formation of polyurethane, a flame retardant, a pigment, at least one filler and/or a surfactant.
The machine process properties may in principle relate to any setting applied to a device, machine or plant involved in the production of a polymeric product or any constant characteristic describing such a device, machine or plant. In particular, the user-settable machine process settings may comprise machine process settings for variably adjusting the operation of a machine in the production of a polymeric product. According to a further preferred embodiment of the method, the machine process properties, preferably the user-settable machine process settings, comprise a component temperature, a mixing time, a mixing proportion, a tool temperature, a discharge capacity and/or a line speed. All the described machine process properties are in particular for the production of the polymeric product.
The formulation database may either be a single database or may be a system of several databases, with different kinds of information stored in each of the several databases.
It is preferred that the formulation database is at least partially based on experimental results. Thus in a preferred embodiment, at least some of the data entries of the formulation database are based on polymeric products produced prior to the user input, such that the formulation portions data specifies raw material portions used for the production of the respective polymeric product and the processing parameters data specifies the machine process properties during the production of that polymeric product.
Preferred embodiments, features and advantages of the computer system according to the invention correspond to those of the method according to the invention and vice versa.
Further advantageous and preferred features are discussed in the following description with respect to the Figures. In the following it is shown in
The exemplary method illustrated in
In the present case, it may be desired to determine the production parameters for obtaining a polyurethane product for which a density of 50 kg/m{circumflex over ( )}3 is desired, which presents user product target as part of user input.
Beside the above-mentioned prediction model 1, the policy 4 also comprises modification boundary parameters 6. In the present example, these specify that proportional water content may be in the bracket between 0% and 2% and have a modification step of 0.1%, that proportional isocyanate content may have a modification step of 10 parts and that the temperature of the isocyanate during production may have a modification step of 2° C. and is in the bracket between 0 and 100 parts and that the temperature of the isocyanate during production may be between 20° C. and 50° C.
Then a new data entry 7, where each data entry comprises production parameters 2 and product properties data 3, is generated by the computer system 5 and more precisely by an agent 11, which agent 11 is also a software running on the computer system 5. The generation of the new data entry 7 is done by means of an action 9 and according to the policy 4, i.e. in particular based on the prediction model 1 and the modification boundary parameters 6. In the present example, the generation of the new data entry 7 occurs by modifying a default data entry 10 which is either pre-determined or given randomly. In particular, the proportional water content is increased by 0.1% proceeding from the value of the production parameters 2 of the default data entry 10.
The generated new data entry 7 comprises generated production parameters 2 such that based on the generated production parameters 2, a corresponding polyurethane product is produced and the specified data entry properties—i.e. in the present example the density—as well as other product properties 3 are determined In the present example, the measurement of the produced polyurethane product results in a density of 40 kg/m{circumflex over ( )}3.
A reward metric 8—which in the present case is a proximity metric—is determined by comparing this measurement to the user product target of 50 kg/{circumflex over ( )}3 and based on the closeness to the target, a higher or lower reward metric 8 generated.
This reward metric 8 is fed back to the policy 4 and in particular the prediction model 1, in the present example by means of the agent 11, to reinforce or weaken the modification that was underlying the generation of the new data entry 7.
Then, the cycle is repeated, and the next new data entry 7 is generated based on modification of the previously generated new data entry 7 until the reward metric 8 is so high that the target defined by the reward metric 8 is achieved.
In
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2019/073815 | 9/6/2019 | WO | 00 |
