Patent Application
20230027989
The present disclosure relates to an analysis method, analysis apparatus, analysis system, and analysis program, each analyzing chemical reaction from a starting materials.
In order to efficiently obtain a desired product from a starting material by chemical reaction, researchers have employed a method in which experiments on chemical reactions are repeatedly conducted by changing reaction conditions based on researchers' own empirical rules by using a reaction device to search for more efficient reaction conditions. Such a method requires researchers to determine reaction conditions on their own, and further, it is time-consuming because the operation of the reaction device and the analysis of data are conducted by a person.
In contrast, PTL 1 discloses repeating a cycle comprising analyzing a chemical reaction of a subset of compounds selected from a library of compounds based on specific selection criteria and selecting a subset of compounds from a library of compounds based on the selection criteria generated from the above analysis, thereby generating chemical entities with defined physical, chemical, and/or bioactive properties. PTL 2 discloses a reaction mechanism generation method comprising the step of performing a molecular dynamics calculation at each time step of atoms that constitute each molecule in a reaction system, the step, performed when a chemical reaction has occurred in the reaction system before and after the time step, of identifying a reacting molecule and a product molecule that contributed to the chemical reaction, the step of constructing, based on atomic relationships between the reacting molecule and the product molecule, an elementary reaction structured from the reacting molecule and the product molecule having the relationship, and the step of calculating a reaction rate constant for the constructed elementary reaction.
PTL 1: JP 2001-507675A
PTL 2: WO 0216/133002
In order to achieve the above object, the analysis method according to one embodiment of the present disclosure is a method for analyzing a chemical reaction from one or more starting materials, comprising preparing a reaction network diagram indicating the one or more starting materials, at least part of one or more intermediate products and one or more final products generated from the chemical reaction, and reaction pathways of the chemical reaction, and predicting the reaction rate in each reaction pathway using an artificial intelligence algorithm.
Preferable embodiments of the present disclosure are described below with reference to the attached drawings. The present disclosure is not limited to the following embodiments.
The chemical reaction may be a degradation reaction, synthesis reaction, or polymerization reaction; however, it is more preferably a degradation reaction. Examples of the degradation reaction include degradation of a pharmaceutical agent (degradation relating to metabolism in the body), ozone (layer) depletion by CFCs, degradation and detoxification of NOx in exhaust gas, analysis of degradation mechanisms of chemical products, and etching processing.
Steps S1 to S5 can be performed by the analysis system 100 of the present embodiment. The details of processing of steps S1 to S5 are described below with reference to
The analysis apparatus 1 is an apparatus for analyzing a chemical reaction from a starting material. The hardware configuration of the analysis apparatus 1 is not limited, and the analysis apparatus 1 can be formed of a general computer, a portable computer such as a tablet or smartphone, or a quantum computer.
The analysis apparatus 1 includes, as functional blocks, a preparation unit 11, a prediction unit 12, an updating unit 13, and a reaction condition search unit 14. When the analysis apparatus 1 is formed of a general computer, the units 11 to 14 are functioned when the GPU or CPU in the analysis apparatus 1 reads out the analysis program of this embodiment into the main memory to perform the analysis program. The analysis program may be downloaded to the analysis apparatus 1 via a communication network, such as the Internet, or the analysis program may be recorded on a non-transient computer-readable recording medium such as a CD-ROM, and installed in the analysis apparatus 1 via the recording medium. The functions of units 11 to 14 are described below.
The PLC 2 is a control device for controlling the reaction device 3. The analysis apparatus 1 may have a function of PLC 2.
The reaction device 3 is a device for performing a chemical reaction such as a decomposition reaction or an additional reaction. The reaction device 3 comprises a reactor 31 for storing a chemical material, and allows a chemical reaction to consecutively proceed in the reactor 31 under the predetermined reaction conditions (e.g., temperature and residence time).
The analysis device 4 is a device that analyzes the components of a material in the reactor 31. The reaction device 3 may have a function of the analysis device 4.
The following section describes the details of processing when each of the steps S1 to S5 in the method for analyzing a chemical reaction from a starting material, which is shown in
The preparation step S1 is performed by the preparation unit 11 in the analysis apparatus 1. The preparation unit 11 has a function of preparing a reaction network diagram indicating the one or more starting materials, at least part of one or more intermediate products and one or more final products generated from the chemical reaction, and reaction pathways of the chemical reaction.
The starting material A, intermediate products B to V, and final product W can be gases, liquids, or solids, and are preferably gases. The starting material A, intermediate products B to V, and final product W may be low-molecular-weight compounds or high-molecular-weight compounds, and are preferably low-molecular-weight compounds. The number of intermediate products B to V and final product W, and the number of reaction pathways are not limited; however, they are determined by considering the calculation amount in the prediction step S2, updating step S4, and reaction condition search step S5 described below. The reaction network diagram ND may not include the final product.
In the entire diagram formation step S11, when the starting material is determined by the user, the preparation unit 11 forms an exhaustive reaction network diagram (entire diagram) indicating the starting material(s), (almost) all of the intermediate product(s) and final product(s) resulting from the chemical reaction, and (almost) all of the reaction pathways of the chemical reaction. The entire diagram can be automatically formed according to a predefined reaction pattern. For example, if the chemical reaction is a degradation reaction, one or more intermediate products produced from the starting material(s) are determined according to the possible pattern of material(s) desorbed from the starting material(s). Similarly, intermediate products are additionally determined according to the possible pattern of material(s) desorbed from the intermediate product(s). This process is repeated to form the entire diagram.
The entire diagram does not necessarily include all of the intermediate products and final products. Similarly, the entire diagram does not necessarily include all the reaction pathways of the chemical reaction. The number of reaction pathways in the entire diagram is not limited. It is, for example, about 5000.
Although the entire diagram shows (almost) all of the reaction pathways, the use of the entire diagram in the prediction step S2, updating step S4, and reaction condition search step S5 results in a huge amount of calculation.
Accordingly, in this embodiment, the simplification step S12 is performed.
In the simplification step S12, the preparation unit 11 simplifies the entire diagram to form the reaction network diagram ND shown in
Thus, the preparation step S1 is completed.
The prediction step S2 is performed by the prediction unit 12 in the analysis apparatus 1. The prediction unit 12 has a function of predicting the reaction rate in each reaction pathway using an artificial intelligence algorithm. The artificial intelligence algorithm is not limited, and is preferably a gradient algorithm, a genetic algorithm, or a combination thereof.
In the rate constant calculation step S21, the prediction unit 12 calculates the rate constant showing the reaction rate in each reaction pathway. In this embodiment, the prediction unit 12 calculates the rate constant based on the Arrhenius equation. In the Arrhenius equation, the rate constant k is represented by the following formula:
wherein A is a frequency factor, Ea is an activation energy per mol, R is a gas constant, and T is an absolute temperature. Since the rate constant k varies according to the temperature and residence time in the chemical reaction, in the rate constant calculation step S21, the prediction unit 12 calculates the rate constants k1−0 to kn−0 (wherein n is the number of reaction pathways in the reaction network diagram ND) in individual reaction pathways under reaction conditions of temperature T0 and residence time t0 as the initial values of the rate constants. The temperature T0 and residence time t0 may be given by the user or randomly given by the prediction unit 12.
In the rate constant modification step S22, the prediction unit 12 modifies the rate constant by an artificial intelligence algorithm. In this embodiment, the rate constants k1−0 to kn−0 are modified by changing reaction conditions (temperature and residence time) using a gradient algorithm, a genetic algorithm, or a combination thereof as an artificial intelligence algorithm so that the value of the evaluation function (e.g., target chemical material yield and cost) increases. Thereby, reaction conditions under which the evaluation function is maximized are searched for, and the rate constants are converged to rate constants k1−1 to kn−1 in which the evaluation function is maximized in theory. The rate constants k1−1 to kn−1 are reaction rates predicted in the prediction step S2. The reaction conditions in this case are temperature T1 and residence time t1. The reaction conditions are not limited to temperature and residence time, and it may be, for example, reaction time, reaction pressure, raw material concentration, or raw material introduction flow rate.
It is more preferable to use a genetic algorithm or an algorithm involving the combination of a gradient algorithm and a genetic algorithm because the rate constant may fall into a local solution if the gradient algorithm alone is used.
Thus, the prediction step S2 is completed.
In the performing step S3, the chemical reaction is performed more than once using the reaction device 3 under different reaction conditions. As described below, since the performing step S3 to the reaction condition search step S5 are performed more than once, in the single performing step S2, the chemical reaction is performed once under specific reaction conditions.
In this embodiment, the starting material A is subjected to chemical reaction in the reactor 31 of the reaction device 3 by changing reaction conditions within the range of temperature T2 to T3 and residence time t2 to t3. In this embodiment, the temperatures T2 and T3 satisfy T2<T3 and T1=(T2+T3)/2; however, the temperatures T2 and T3 are not limited as long as they satisfy T2<T1<T3, and are the temperatures at which the reaction device 3 can be operated. Similarly, the residence times t2 and t3 in this embodiment satisfy t2<t3 and t1=(t2+t3)/2; however, the residence times t2 and t3 are not limited as long as they satisfy t2<t1<t3. The reaction conditions of the chemical reaction are determined by the reaction condition search unit 14, which is described below; however, the reaction conditions of the chemical reaction may be set by the user if the reaction condition search step S5 is omitted, as described below. The PLC 2 controls the reaction device 3 according to the determined reaction conditions.
When the chemical reaction in the reaction device 3 is completed, the analysis device 4 analyzes the concentration of each chemical material in the reactor 31, and the yield of the target compound. Furthermore, the analysis device 4 provides feedback on the analyzed concentration and yield to the analysis apparatus 1 as the results of the performing step.
Subsequently, in the updating step S4, the updating unit 13 updates the reaction rates predicted in the prediction step S2, based on the results of the performing step S3. In this embodiment, the updating unit 13 updates the predicted reaction rate by an automatic differential method. Specifically, the updating unit 13 updates rate constants to k1−1 to kn−1 rate constants k1−2 to kn−2 in such a manner that the error is minimized by comparing the concentration obtained by the analysis device 4 and output results of the yield of the target compound with the concentration of a compound output from the reaction network diagram with the reaction rate group obtained in the immediately preceding prediction step S2 and output results of the yield of the target compound.
In the subsequent reaction condition search step S5, the reaction condition search unit 14 searches for optimal reaction conditions based on the reaction conditions and results of the performing step S3. In this embodiment, the reaction condition search unit 14 searches for the optimal reaction conditions by Bayesian optimization. In the first reaction condition search step S5, the reaction condition search unit 14 records reaction conditions and the results in the first performing step S3, as shown in
The number of times in which the performing step S3 to the reaction condition search step S5 are repeated is not limited; however, it is preferable that the steps be repeated (m times) until the rate constant is sufficiently converged in the updating step S4, and an appropriate reaction condition can be searched for (no possibility of falling into a local solution) in the reaction condition search step S5 (YES in step S6).
For example, in the i-th (1<i<m) performing step S3, the PLC 2 controls the reaction device 3 according to the reaction conditions determined by the reaction condition search unit 14 in the (i−1)-th reaction condition search step S5 to perform a chemical reaction. In the i-th updating step S4, the updating unit 13 updates the rate constant in the reaction network diagram ND based on the results of the first performing step S3 to i-th performing step S3. The more times the updating step S4 is performed, the smaller the latitude of the rate constant becomes, and so the rate constant is converged to a predetermined value. In the i-th reaction condition search step S5, the reaction condition search unit 14 records the reaction conditions and results in the i-th performing step S3, and determines reaction conditions for the (i+1)-th performing step S3 based on the optimization theory.
Based on the reaction network diagram ND in
As described above, in this embodiment, the reaction network diagram is prepared in the preparation step S1, and the reaction rates in the reaction network diagram are predicted in the prediction step S2 with an artificial intelligence algorithm such as a gradient method algorithm. Since the reaction network diagram is formed based on chemical knowledge, and the reaction rates are predicted with an artificial intelligence algorithm considering the reaction pathways shown in the reaction network diagram, excellent reaction conditions for chemical reaction can be efficiently searched for. In contrast, PTL 1 does not consider the reaction rate of the chemical reaction. In PTL 2, a reaction network diagram showing reaction pathways of chemical reaction generated from a starting material (initial molecule) is not used, nor does it disclose the prediction of the reaction rate with an artificial intelligence algorithm considering the reaction pathways.
Furthermore, in the performing step S3 to the reaction condition search step S5, since reaction conditions are inductively searched for by Bayesian optimization or the like based on the results of the chemical reaction using the reaction device, more excellent reaction conditions can be searched for in a shorter time. This reduces the work required for reaction condition search, and results in highly efficient search.
The embodiments are described above; however, it will be understood that various changes in forms and details can be made without departing from the spirit and scope of the claims.
For example, in the above embodiments, the reaction condition search step S5 is performed after the updating step S4; however, the updating step S4 may be performed after the reaction condition search step S5. Alternatively, the updating step S4 and the reaction condition search step S5 may be simultaneously performed.
The reaction condition search step S5 may be omitted. In such a case, the performing step S3 and the updating step S4 may both be repeated once under different reaction conditions, or the performing step S3 may be performed beforehand multiple times under different reaction conditions, followed by the updating step S4.
The preparation step S1 and the prediction step S2, or the preparation step S1 alone may be omitted. In such a case, the user may, for example, prepare a reaction network diagram based on the user's own empirical rules.
The performing step S3 to the reaction condition search step S5 may be omitted, and reaction pathways may be searched for based on the reaction network diagram ND obtained in the prediction step S2. The reaction rates predicted in the prediction step S2 are theoretical reaction rates predicted with an artificial intelligence algorithm such as a gradient method algorithm; however, it is possible to find near-optimal reaction pathways as long as a huge error, such as failing into a local solution, does not occur.
In the above embodiment, a gradient algorithm, genetic algorithm, or combination thereof was used as the artificial intelligence algorithm in the prediction step S2; however, the artificial intelligence algorithm in the present disclosure is not limited to these. Artificial intelligence algorithms, such as the Gauss-Newton method and EM algorithm, can also be used in the prediction step S2.
In the above embodiment, in the reaction condition search step S5, the optimal reaction conditions are searched for by Bayesian optimization; however, the method of searching for optimal reaction conditions is not limited to this. In the reaction condition search step S5, a method such as a genetic algorithm, ant colony optimization, and reinforcement learning can also be used.
In the above embodiment, the reaction conditions to be searched for are temperature and residence time; however, the type and number of reaction conditions are not limited. For example, the reaction conditions may include the partial pressure of the raw material.
The present disclosure can be applied for reaction condition search of chemical reaction for producing various chemical materials.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2020-067796 | Apr 2020 | JP | national |
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/JP2021/006568 | Feb 2021 | US |
| Child | 17956122 | US |
