EP4334949A1

EP4334949A1 - Method for constructing a model simulating a chemical reaction

Info

Publication number: EP4334949A1
Application number: EP22724743.4A
Authority: EP
Inventors: Cedric MEHL; Thibault FANEY
Original assignee: IFP Energies Nouvelles IFPEN
Current assignee: IFP Energies Nouvelles IFPEN
Priority date: 2021-05-05
Filing date: 2022-04-22
Publication date: 2024-03-13
Also published as: US20240212794A1; CA3213684A1; WO2022233602A1; FR3122767A1; JP2024517767A

Abstract

The present invention relates to a method for constructing a model (MSI) simulating at least one chemical reaction, which constructs an intermediate model by means of an artificial neural network (RNA) and of a training database (BAP), and that then constructs the simulating model by adding an additional mass-conservation layer (RNC) to the artificial neural network (RNA). Furthermore, the invention relates to a method for simulating (SIM) a chemical reaction that implements the constructed simulating model (MSI).

Description

METHOD FOR CONSTRUCTING A CHEMICAL REACTION SIMULATION MODEL

Technical area

The present invention relates to the field of the simulation of chemical kinetics, with the aim of predicting a mixture of chemical species at the output of a reactive system.

The prediction of the behavior of chemical reactive systems is important for the design of many industrial systems, such as processes or energy conversion systems in transport with chemical propulsion, as well as in the transport issues of these chemical compounds. In these systems, a fluid containing a mixture of chemical species undergoes various transformations, called chemical reactions. For example, it may involve the prediction of the composition of chemical species within combustion, in particular in a combustion engine, chemical reactions within an electric battery, within a fuel cell , within a reactive flow, within a catalysis process, etc.

This modeling consists of predicting the evolution of chemical reagents (chemical species) in a reactive system between an input and an output. In this patent application, the notion of input and output can correspond to one of the following definitions: the input and the output can correspond to a physical input and output of a system (before and after a reaction reaction), as well as the evolution of the system between two time steps during a numerical simulation (before or after a time step of a chemical reaction).

Prior technique

To simulate chemical reactions, so-called “data-based” approaches can be used. Such approaches make it possible to predict the behavior of the chemical reaction in the reactive system from data resulting from simulations or prior experiments, and by means of machine learning algorithms. Generally, these approaches not directly related to the physics, do not make it possible to take into account all the physical phenomena implemented in chemical reactions, in particular the conservation of the mass of each atom present in the mixture of chemical species. Therefore, such approaches remain imprecise.

To take into account the conservation of mass, some approaches have provided an additional step in which the conservation of mass is verified. However, this additional step is performed outside the model, which limits the optimization of the prediction.

Patent application WO2020176914 describes a prediction of gas concentrations from neural networks. The method is applied to the post-treatment of engine exhaust gases. The direct calculation of chemical kinetics is expensive, hence the use of neural networks. Nevertheless, for this method the conservation of mass is not taken into account by the neural networks, which does not allow an accurate prediction of the concentrations of the exhaust gases.

The patent application CN109918702 envisages the prediction of chemical states by neural networks, taking into account the conservation of the mass of the atoms by means of a constraint. However, this constraint is not included in the neural network.

The document: "Hendricks et al, LINEARLY CONSTRAINED NEURAL NETWORKS, arXiv: 2002.01600, 2020" relates to the field of fluid mechanics. In this document, an artificial neural network is proposed to guarantee zero divergence of a velocity field in the case of an incompressible fluid. This document does not take into account the conservation of mass for a chemical reaction.

The document “Mohan et al., EMBEDDING HARD PHYSICAL CONSTRAINTS IN NEURAL NETWORK COARSE-GRAINING OF 3D TURBULENCE, arXiv:2002.00021, 2020) also concerns the field of fluid mechanics. In this document, an artificial neural network is also proposed to guarantee zero divergence of a velocity field in the case of an incompressible fluid. This document does not take into account the conservation of mass for a chemical reaction.

Summary of the invention

The aim of the invention is to construct a simulation model of at least one chemical reaction making it possible to predict in an accurate and reliable manner the chemical compositions during the at least one chemical reaction. For this, the present invention relates to a method building a simulation model of at least one chemical reaction, which builds an intermediate model by means of an artificial neural network and a learning base, then which builds the simulation model by adding an additional mass conservation layer to the artificial neural network.

Furthermore, the invention relates to a method for simulating a chemical reaction, implementing the built simulation model.

The invention relates to a method for constructing a simulation model of at least one chemical reaction between several chemical species within a reactive system, by means of a learning base, said learning base comprising learning input data and learning output data of said chemical reaction, said learning input and learning output data being representative of the mass fractions of said chemical species respectively at the input and at the output of said chemical reaction. For this method, the following steps are carried out: a. An intermediate model of said simulation model is constructed by means of an artificial neural network, which comprises an input layer, at least one hidden layer and an output layer, the construction of said intermediate model being implemented by means of said learning base; and B. Said simulation model is constructed by means of said intermediate model and an additional layer implemented after said output layer of said neural network, said additional layer being formed by a linear operator which applies a linear correction to the output data of said output layer of said network of artificial neurons to ensure the conservation of the mass of the chemical elements forming said chemical species used in said chemical reaction.

According to one embodiment, said additional layer is constructed by means of a linear correction calculation step and a step of applying said calculated linear correction.

Advantageously, the linear correction is calculated for a number of chemical species of said chemical reaction corresponding to the number of chemical elements of said chemical reaction. According to one implementation, the synaptic weights of the artificial neural network are optimized after the addition of the additional layer by means of the input and output data of the learning base.

According to one aspect, said input and output data are determined respectively before and after said chemical reaction, or respectively before and after a time step of said chemical reaction.

According to one implementation option, said input data are mass fractions of said chemical species.

According to one characteristic, said output data are mass fractions of said chemical species or variations of the mass fractions of said chemical species.

Furthermore, the invention relates to a method for simulating at least one chemical reaction between several chemical species within a reactive system. For this method, the following steps are carried out: a. A simulation model is constructed by means of the method for constructing a simulation model according to one of the preceding characteristics; and B. Said simulation model is applied to simulation input data, said simulation input data being representative of said mass fractions of said chemical species as input to said chemical reaction.

According to one embodiment, said simulation model is constructed for a time step, and the step of applying said simulation model is reiterated for a plurality of successive time steps, and at each reiteration, said data of simulation input are the output data of the previous time step.

According to one implementation, said chemical reaction is a chemical reaction within an energy conversion system, for example a combustion within an engine or a chemical reaction within an electric battery, or a chemical reaction within within a fuel cell, or a chemical reaction within a system for transporting said chemical species.

Other characteristics and advantages of the method according to the invention will appear on reading the following description of non-limiting examples of embodiments, with reference to the appended figures and described below. List of Figures

FIG. 1 illustrates the steps of the method for constructing a simulation model of a chemical reaction according to a first embodiment of the invention.

FIG. 2 illustrates the steps of the method for constructing a simulation model of a chemical reaction according to a second embodiment of the invention.

FIG. 3 illustrates the steps of the method for simulating a chemical reaction according to a first embodiment of the invention.

FIG. 4 illustrates the steps of the method for simulating a chemical reaction according to a second embodiment of the invention.

FIG. 5 illustrates the steps of the method for simulating a chemical reaction according to a third embodiment of the invention.

FIG. 6 illustrates a simulation model of a chemical reaction according to one embodiment of the invention.

FIG. 7 illustrates a simulation model of a chemical reaction according to one embodiment of the invention.

FIG. 8 illustrates the curves of the mass fractions of four chemical elements as a function of time for a simulation by means of a method according to the prior art.

FIG. 9 illustrates the curves of the mass fractions of four chemical elements as a function of time for a simulation by means of a method according to an embodiment of the invention.

Description of embodiments

The present invention relates to a method for constructing a simulation model of at least one chemical reaction between several chemical species within a reactive system. In other words, the present invention relates to the prediction of chemical kinetics; that is to say the prediction of the chemical species during a chemical reaction, or at a given moment of the chemical reaction.

By way of illustration only, the various terms will be illustrated in the case of combustion of methane within a combustion chamber. However, the process according to the invention can be applied to other chemical reactions. The different possible applications of the method according to the invention are described at the end of this application.

We call chemical reaction: the exchanges of matter between various chemical species in a reactive system. Generally, the chemical reaction can be written in the form of a balance equation. For the example of methane combustion (without taking into account the presence of nitrogen in the air), the chemical reaction can be written: CH ₄ + 2O ₂ ® CO ₂ + 2 H ₂ O.

The reactive system is the system in which the chemical reaction takes place. For the example of the combustion of methane, the reactive system is the combustion chamber. The molecules used in the chemical reaction are called chemical species. For the example of the combustion of methane, the chemical species are methane CH ₄ , dioxygen O ₂ , carbon dioxide CO ₂ and water H ₂ O. works in the chemical reaction. For the example of the combustion of methane, the chemical elements are carbon C, hydrogen H, oxygen O.

The present invention implements a learning base. The learning base comprises input data representative of the mass fractions of the chemical species entering the reactive system. The learning base further comprises output data representative of the mass fractions of the chemical species at the output of the reactive system. The mass fraction of a chemical species is defined by the ratio of the mass of the chemical species to the total mass of the mixture in the reactive system.

The notion of input and output can correspond to one of the following definitions: this notion of input and output can correspond to a physical input and output of a system (before and after a chemical reaction), or to the evolution of the system between two time steps during a numerical simulation (before or after a time step of a chemical reaction).

Advantageously, the input data of the learning base can correspond to the mass fractions of the chemical species.

According to one embodiment, the output data from the learning base can correspond to the mass fractions of the chemical species.

As a variant, the output data from the learning base can correspond to the variations of mass fractions of the chemical species, these variations being those produced during the chemical reaction. The input and output data of the learning base can come from experimentally measured data and/or from simulation data. In accordance with one implementation of the invention, the method may include a prior step of constructing the learning base, from data measured experimentally and/or data from simulation.

According to the invention, the method for constructing a simulation model comprises the following steps:

1. Construction of an intermediate model (artificial neural network)

2. Construction of the simulation model

These steps can be implemented by computer means, comprising in particular a computer or a server. These steps will be detailed in the remainder of the description.

Figure 1 illustrates, schematically and in a non-limiting manner, the steps of the method according to a first embodiment of the invention. From a BAP learning base, an intermediate model is trained by means of an RNA artificial neural network and an OPT (learning) optimization of the RNA artificial neural network. This optimization step makes it possible to determine the synaptic weights of the RNA artificial neural network according to the data from the BAP learning base. Then, the MSI simulation model is built, using the optimized (learned) RNA artificial neuron model and an additional layer, called the additional RNC conservation layer.

According to one embodiment, the synaptic weights of the artificial neural network can be optimized after the addition of the additional conservation layer by means of the data from the learning base. In other words, learning is applied to the entire model formed of the artificial neural network and the additional conservation layer. Thus, the simulation model can be optimized by taking into account data from the learning base.

Figure 2 illustrates, schematically and in a non-limiting manner, the steps of the method of this embodiment. We build an artificial neural network RNA, and the additional conservation layer RNC. Then, from a BAP learning base, the artificial neural network RNA is optimized OPT (learning) simultaneously with the additional layer of RNC conservation in order to build the MSI simulation model. Thus, during this optimization step, an update of the synaptic weights of the artificial neural network with the additional RNC conservation layer is implemented from the learning database BAP.

Furthermore, the invention relates to a method for numerical simulation of at least one chemical reaction between several chemical species within a reactive system. Such a numerical simulation method corresponds to a method for predicting chemical species during a chemical reaction within a reactive system. The simulation method implements the model construction method according to any one of the variants or combinations of variants described in the present application, and a step of applying the simulation model to simulation input data. The numerical simulation then generates simulation output data.

The simulation input data and the simulation output data are of the same nature as the input and output data of the learning base. In other words, the simulation input and output data are representative of the mass fractions of the chemical species respectively at the input and at the output of the simulated reactive system.

The notion of simulation input and output can correspond to one of the following definitions: on the one hand to a physical input and output of a system (before and after a chemical reaction), on the other hand to the evolution of the system between two time steps during a numerical simulation (before or after a time step of a chemical reaction).

Advantageously, the simulation input data can correspond to the mass fractions of the chemical species.

According to one embodiment, the simulation output data may correspond to the mass fractions of the chemical species.

Alternatively, the simulation output data may correspond to variations in mass fractions of the chemical species.

Simulation input data may be obtained experimentally, or may be obtained by simulation.

Thus, the simulation process includes the following steps:

1. Construction of an intermediate model (artificial neural network) 2. Construction of the simulation model

3. Application of the simulation model

These steps can be implemented by computer means, comprising in particular a computer or a server. Preferably, steps 1 and 2 can be implemented beforehand only once, and step 3 can be repeated with the simulation model. These steps are detailed in the remainder of the description.

FIG. 3 illustrates, schematically and in a non-limiting manner, the steps of the digital simulation method according to a first embodiment of the invention. From a BAP learning base, an intermediate model is trained by means of an RNA artificial neural network and an OPT optimization of the RNA artificial neural network. This optimization step makes it possible to determine the synaptic weights of the RNA artificial neural network according to the data from the BAP learning base. Then, the MSI simulation model is built, using the optimized RNA artificial neuron model and an additional RNC conservation layer. Next, the simulation model MSI is applied SIM to simulation input data DES to determine simulation output data DSS.

FIG. 4 illustrates, schematically and in a non-limiting manner, the steps of the digital simulation method according to a second embodiment of the invention. This second embodiment of the digital simulation method implements the second embodiment of the method for constructing the simulation model of FIG. 2. An artificial neural network RNA, and the additional conservation layer RNC, are constructed. Then, from a BAP learning base, the artificial neural network RNA is optimized OPT simultaneously with the additional conservation layer RNC so as to build the simulation model MSI. Thus, during this optimization step, an update of the synaptic weights of the artificial neural network with the additional RNC conservation layer is implemented from the learning database BAP. Next, the simulation model MSI is applied SIM to simulation input data DES to determine simulation output data DSS.

In accordance with an implementation of the invention, for which a simulation is carried out for a plurality of successive time steps (that is to say for several intermediate stages of the chemical reactions), it is possible to repeat the stage of the application of the simulation model, and, at each reiteration, the input data of simulation can be the simulation output data of the previous time step. Thus, a chemical reaction can be simulated step by step, in a precise and robust manner.

FIG. 5 illustrates, schematically and in a non-limiting manner, the steps of the digital simulation method according to a third embodiment of the invention. This third embodiment of the digital simulation method implements the second embodiment of the method for constructing the simulation model of FIG. 2. An artificial neural network RNA, and the additional conservation layer RNC, are constructed. Then, from a BAP learning base, the artificial neural network RNA is optimized OPT simultaneously with the additional conservation layer RNC so as to build the simulation model MSI. Thus, during this optimization step, an update of the synaptic weights of the artificial neural network with the additional RNC conservation layer is implemented from the learning database BAP. Next, the simulation model MSI is applied SIM to simulation input data DES to determine simulation output data DSS. Then, the simulation step is repeated for new simulation input data, which correspond to the simulation output data of the previous time step.

1 . Construction of the intermediate model

During this step, an intermediate model is constructed using an artificial neural network. The model is said to be intermediate, because it is not the simulation model obtained by the process: this intermediate model is modified in step 2. The artificial neural network makes it possible to obtain a model allowing rapid evaluation of the data of output data, and with a limited computer memory requirement, in particular faster than a complex numerical model representative of the chemical reaction. In addition, the artificial neural network is suitable for complex chemical reactions, in particular involving a large number of chemical species. Indeed, the simulations of these complex systems can require significant computer resources, in particular in terms of memory, and high simulation times. An artificial neural network is an algorithm whose parameters are optimized using a so-called training database derived from the learning base, which contains the learning input/output pairs. Once this network is trained on data, it can then be used to make inferences about inputs never seen before. The artificial neural network includes an input layer, at least one hidden layer and an output layer. The input layer and the output layer have a number of neurons which corresponds to the number of chemical species in the chemical reaction. The construction of the artificial neural network is implemented by means of the learning base. In other words, the synaptic weights of the neural network are learned in a supervised way to match the input and output data of the learning base. Thus formed, the intermediate model makes it possible to determine output data from input data, while being representative of the data of the learning base. The inputs of the intermediate model are of the same nature as the input data of the learning base, and the outputs of the intermediate model are of the same nature as the output data of the learning base.

The artificial neural network can be of any form, for example a deep neural network, a recurrent neural network, a multi-layer Perceptron neural network, etc.

The construction of the intermediate model may include a validation step of the artificial neural network, in particular to avoid problems of overlearning. The validation can be implemented by means of part of the data from the learning base.

2. Construction of the simulation model

During this step, the simulation model is constructed by means of the intermediate model constructed in step 1 and an additional layer, called additional conservation layer. The additional conservation layer is added after the output layer of the artificial neural network. The additional conservation layer is formed by a linear operator for the conservation of the mass of the chemical elements present in the chemical reaction. In other words, the simulation model comprises the input layer, the at least one hidden layer and the output layer of the artificial neural network, the additional conservation layer. Thus, the additional conservation layer makes it possible to verify the conservation of the mass of the chemical elements within the simulation model itself, which makes the simulation model more precise. Moreover, the linearity of the additional conservation layer makes it possible to retain the advantages of the artificial neural network: simulation of complex systems with limited computer memory and limited calculation time. The additional conservation layer comprises a number of inputs and a number of outputs which correspond to the number of chemical species of the chemical reaction. At the output of the additional conservation layer, and consequently at the output of the simulation model, data representative of the mass fractions of the chemical species of the chemical reaction are determined. Once determined, the additional layer of conservation is static and constant. In other words, the additional preservation layer is not modified.

According to one aspect of the invention, the additional conservation layer can be constructed by means of a step for calculating the linear correction and a step for applying the calculated linear correction.

For this embodiment, in the case for which the number of chemical species is greater than the number of chemical elements, the correction step can be applied for a number of chemical species which corresponds to the number of chemical elements. Preferably, the chemical species chosen can contain at least once each atom present in the mixture. Thus, an underdetermination of the problem can be avoided.

According to an embodiment for which the input data and the output data are normalized, these two steps (correction and application of the correction) can be preceded by a normalization inversion step, and can be followed by a standardization step.

These two steps (correction and application of the correction) are detailed in the remainder of the description.

In the remainder of the description, a mixture of chemical species numbered 1, ..., N _s and described by the mass fraction Y _k of each of these species is considered. The mass fraction is defined by Y _k = m _k /m where m _k is the mass of species k in the mixture and m the total mass of the mixture. It is considered that chemical species react with each other, excluding nuclear-type reactions. Therefore, the mass of each atom composing the molecules of the mixture is conserved between the input state and the output state. For the rest of the description, an input chemical state is denoted in and the corresponding output out. The corresponding mass fractions are denoted and respectively. The passage between entry and exit involves a number of chemical reactions between these species.

We note j any atom (a chemical element) present in the mixture (for example: j = 0,H, C,N in the case of a system involving combustion reactions). It is assumed that there are N _a distinct atoms in the mixture. The fraction Y _j corresponds to the mass of atom j relative to the total mass of the mixture. This mass fraction can be written from the mass fractions of species as follows: Where M _k corresponds to the molar mass of the chemical species k, is the number of atoms j present in the molecule of chemical species k and M _j corresponds to the molar mass of atom j. This relation can be written in a matrix way by setting the vector containing the mass fractions of atoms and the vector of the mass fractions chemical species. We can then write:

With: M = (m _jk ) such that

The conservation of the mass of the atoms between the input and the output gives the following property, which can be written for any atom j present in the mixture: Either in matrix version:

We note g the output vector of the artificial neural network. This embodiment consists in correcting the vector g in order to satisfy the previous relationship. The correction can be written by adding an a priori term to be determined, denoted α:

With ^s a vector allowing to correct the vector g. The previous equation becomes:

This equation is a linear system with unknowns . Insofar as , this system contains N _s unknowns and N _a equations. Typically, the number of atoms is much less than the number of species (N _a « N _s ), and the system is therefore underdetermined. In order to remove the underdetermination, a set of N _a chemical species among the N _s present in the mixture can be extracted. It is preferable that these species contain at least once each atom present in the mixture. Note the indices of these species. We assume that and a _k = 0 otherwise (in other words, for this implementation, a correction can be made for the chemical species selected, and no correction is made for the other species of the chemical reaction). A new correction vector can be defined as follows:

The system of equations then becomes:

Where M' corresponds to the sub-matrix of M containing only the columns corresponding to the species . We then have Mr. Moreover, if each atom is represented in the selected species, it is possible to show that the system is invertible. The solution is then noted:

Thus, by this calculation, the correction is determined. Then, by applying this correction to the selected species, the exact conservation of the atoms is guaranteed. This application of the correction can be written for each chemical species k:

In accordance with an implementation of the invention (corresponding to FIG. 2), a modification of the neural network, in this case a modification of the synaptic weights, can be carried out after the application of the correction. Thus, this correction being applied inside the neural network, the parameters of the at least one upstream hidden layer can be adjusted so that the output is well predicted, and this despite the fact that an arbitrarily chosen set of species is selected to correct the mass of the atoms.

FIG. 6 illustrates, schematically and in a non-limiting manner, the construction of the simulation model according to one embodiment of the invention. First, we build the artificial neural network RNA. The RNA artificial neural network includes an input layer CE, for which the mass fractions of the species chemicals are the input data. The RNA artificial neural network includes at least one CC hidden layer. The artificial neural network RNA further comprises an output layer CS, for which mass fractions of the chemical species g ₁ , ..., g _Ns constitute the output data. The layers of the artificial neural network RNA comprise a plurality of artificial neurons NA, represented by circles. Each artificial neuron NA is associated with a synaptic weight. In addition, the RNC simulation model further includes an additional layer of conservation L _c which determines the linear correction of the output data g ₁ ,...,g _Ns of the artificial neural network RNA to determine the mass fractions of the chemical species at the outlet, while ensuring the conservation of the mass of the chemical elements. For the embodiment of FIG. 6, for which the output data are mass fractions, the additional conservation layer L _c takes the input data into account.

Figure 7 is a figure similar to figure 6, in the case where the output data are variations of the mass fraction of the chemical species w ₁ ..., w _Ns with . The elements identical to FIG. 6 are not detailed again. For this embodiment, the additional conservation layer L _c does not need to take into account the input data

3. Application of the simulation model

This step relates only to the method of digital simulation of at least one chemical reaction. During this step, the simulation model constructed in step 2 is applied to simulation input data. In other words, the chemical kinetics of the chemical reaction are simulated from simulation input data and from the simulation model. It can be a simulation for the entire chemical reaction or for a time step of the chemical reaction.

In accordance with an implementation of the invention, for which a simulation is carried out for a plurality of successive time steps (that is to say for several intermediate stages of the chemical reaction), it is possible to repeat the step of the application of the simulation model, and, at each reiteration, the simulation input data can be the simulation output data of the previous time step. Thus, a chemical reaction can be simulated step by step, in a precise and robust manner. This implementation corresponds to that illustrated in figure 5. The chemical reaction may be a chemical reaction from the field of reactive fluid mechanics. In this case, the chemical species can be liquids or gases.

According to a first example of implementation, the invention may relate to a chemical reaction of combustion, in particular within a combustion engine, a turbine, or any similar system. Indeed, the methods according to the invention can make it possible to model complex systems with numerous chemical species.

According to a second example of implementation, the invention may relate to a chemical reaction within an energy storage and supply system, in particular within an electric battery or a fuel cell. Indeed, the methods according to the invention can make it possible to model complex systems with numerous chemical species.

According to a third example of implementation, the invention may relate to a chemical catalysis reaction, in particular a catalysis within an exhaust gas post-treatment system, or a catalysis implemented in the field of the production of hydrocarbons. Indeed, the methods according to the invention can make it possible to model complex systems with numerous chemical species.

According to a fourth example, the invention may relate to a chemical reaction specific to fluid flows in a porous medium, in particular a precipitation or dissolution reaction of one or more minerals. Indeed, the methods according to the invention can make it possible to model complex systems with numerous chemical species.

Furthermore, the invention relates to a digital chemical reaction simulator implementing the simulation model constructed by the method according to one of the variants or combinations of variants of the method according to the invention. The digital simulator uses simulation input data to determine simulation output data. The simulation input data and the simulation output data are representative of the mass fractions of the chemical species of the chemical reaction.

In addition, the invention relates to a method for designing a reactive system, in which at least one chemical reaction takes place, in which the following steps are implemented: - The chemical reaction is numerically simulated in the reactive system by means of the simulation method according to any one of the variants or combinations of variants previously described, and - The reactive system is designed, and the chemical reaction is implemented in the reactive system.

For this design process, a plurality of numerical simulations can be carried out, for a plurality of chemical reactions (for example by modifying the chemical species initially present) and/or for a plurality of reactive systems (for example for several dimensions or shapes of reactive systems). The comparison of the simulations makes it possible to determine the chemical reaction and/or the reactive system to be designed.

The present invention may also relate to a method for controlling a reactive system, in which at least one chemical reaction takes place, and in which the following steps are implemented: - the chemical reaction is numerically simulated in the reactive system by means of of the simulation method according to any one of the variants or combinations of variants previously described, and - the reactive system is controlled, preferably in real time, as a function of the simulated chemical reaction, in particular to avoid dangerous situations, or not suitable for the intended use.

Preferably, for the control of the reactive system, it is possible to measure data in real time, which serve as input data for the simulation.

For the example of a combustion, the control may consist in controlling the quantity of air or fuel in the combustion chamber. The purpose of this control may be to limit polluting emissions, or to avoid the phenomenon of knocking, etc.

For the example of an electric battery, the control may correspond to the control of the voltage or of the current, in particular to limit overheating or premature aging of the electric battery.

For the example of a fuel cell, the control may in particular correspond to the control of the air intake, to optimize the chemical kinetics within the fuel cell, so as to increase the performance thereof. The characteristics and advantages of the process according to the invention will appear more clearly on reading the comparative example below.

In order to illustrate the invention, an example is given in this part. We consider here the use of neural networks in a context of numerical simulation of a combustion. The idea is to replace an exact resolution algorithm with an equivalent neural network. In particular, the neural network takes as input mass fractions at a time t of the simulation (state in in the description of the method above) and evaluates the mass fractions at a time t+dt (state out). This is done iteratively for t _k = kdt, k ∈ [0, N _ite ] , where N _ite is the total number of simulation iterations.

The case considered here is a case of methane combustion at constant pressure, where the atoms generally present are carbon (C), hydrogen (H), oxygen (O) and nitrogen (N).

For this comparative example, a method according to the prior art implementing an artificial neural network (without additional conservation layer) and a method according to the invention implementing an artificial neural network with an additional conservation layer are compared. . For both simulations, the learning base is the same, and the number of layers as well as the number of neurons of the artificial neural network are the same. More particularly, the neural network selected is a network of multi-layer Perceptron type (from the English “Multi-Layer Perceptron”) which comprises two hidden layers with 64 units each. The database consists of simulations of the combustion of H ₂ /air mixtures for different initial temperatures T ₀ and richness Φ . One thousand simulations are selected from the learning base, using a Latin Hypercube Sampling algorithm. This sampling is carried out for values of T ₀ between 1600 K (approximately 1326°C) and 1800K (approximately 1526°C), and richness values between 0.7 and 1.5. The optimization of the network weights is performed using a gradient descent method.

FIG. 8 illustrates the curves of the process according to the prior art, as a function of time t in ms, of the normalized mass fraction of carbon Y _c (top left), of the normalized mass fraction of hydrogen Y _H ( top right), the normalized mass fraction of oxygen Y _o (bottom left), and the normalized mass fraction of nitrogen Y _N (bottom right). Note for these figures that the curves of the mass fraction of carbon and hydrogen are not constant over time. By Consequently, the artificial neural network according to the prior art does not make it possible to guarantee the conservation of the mass of the carbon atoms and of the hydrogen atoms, whereas the conservation of the masses of atoms during the simulation is a point critical. Indeed, a deviation of the mass at a given moment can have negative repercussions on the rest of the simulation, and make the simulation imprecise.

FIG. 9 illustrates the curves of the process according to the invention (with the additional preservation layer), as a function of time t in ms, of the mass fraction of the normalized carbon Yc (top left), of the mass fraction of the normalized hydrogen Y _H (top right), normalized mass fraction of oxygen Y _O (bottom left), and normalized mass fraction of nitrogen Y _N (bottom right). Note that the four curves are constant over time. Consequently, the use of the layer L _c according to the invention allows the conservation of the mass of the atoms, and consequently better accuracy and better robustness of the simulation model.

Claims

1. Method for constructing a simulation model (MSI) of at least one chemical reaction between several chemical species within a reactive system, by means of a learning base (BAP), said base of training (BAP) comprising training input data and training output data of said chemical reaction, said training input and training output data being representative of the mass fractions of said chemical species respectively in input and output of said chemical reaction, characterized in that the following steps are implemented: a. An intermediate model of said simulation model is constructed by means of an artificial neural network (ANN), which comprises an input layer (CE), at least one hidden layer (CC) and an output layer (CS), the construction of said intermediate model being implemented by means of said learning base (BAP); and B. Said simulation model (MSI) is constructed by means of said intermediate model and an additional layer (RNC) implemented after said output layer of said neural network, said additional layer (RNC) being formed by a linear operator which applies a linear correction to the output data of said output layer (CS) of said artificial neural network (RNC) to ensure conservation of the mass of the chemical elements forming said chemical species used in said chemical reaction.

2. Method for constructing a simulation model according to claim 1, in which said additional layer (RNC) is constructed by means of a linear correction calculation step and a step of applying said calculated linear correction.

3. Method for constructing a simulation model according to claim 2, in which the linear correction is calculated for a number of chemical species of said chemical reaction corresponding to the number of chemical elements of said chemical reaction.

4. Method for constructing a simulation model according to one of the preceding claims, in which the synaptic weights of the artificial neural network are optimized (OPT) after the addition of the additional layer by means of the input data and output from the learning base.

5. Method for constructing a simulation model according to one of the preceding claims, in which said input and output data are determined respectively before and after said chemical reaction, or respectively before and after a time step of said chemical reaction.

6. A method of constructing a simulation model according to one of the preceding claims, wherein said input data are mass fractions of said chemical species.

7. Method for constructing a simulation model according to one of the preceding claims, in which said output data are mass fractions of said chemical species or variations of the mass fractions of said chemical species.

8. Method for simulating at least one chemical reaction between several chemical species within a reactive system, characterized in that the following steps are implemented: a. A simulation model (MSI) is constructed by means of the method for constructing a simulation model according to one of the preceding claims; and B. Said simulation model (SIM) is applied to simulation input data (DES), said simulation input data being representative of said mass fractions of said chemical species at the input of said chemical reaction.

9. Simulation method according to claim 8, in which said simulation model (MSI) is constructed for a time step, and in which the step of applying said simulation model (MSI) is repeated for a plurality of time steps. of successive times, and at each reiteration, said simulation input data (DES) are the simulation output data (DSS) of the previous time step.

10. Simulation method according to one of Claims 8 or 9, in which the said chemical reaction is a chemical reaction within an energy conversion system, for example a combustion within an engine or a chemical reaction within of an electric battery, or a chemical reaction within a fuel cell, or a chemical reaction within a system for transporting said chemical species.