A basin hopping algorithm based on machine learning recommendation system and application thereof

Abstract

A basin hopping algorithm based on machine learning recommendation system, comprising the following steps: obtaining an initial random occupation configuration for occupation configuration optimization; determining real-time machine learning force field calculation parameters; determining basin hopping parameters; introducing real-time machine learning force field as a recommendation system into the basin hopping algorithm to globally optimize the system occupation configuration; and performing density functional theory calculation verification on the obtained optimized energy minimum occupation configuration. The present application introduces a machine learning recommendation system into the conventional basin hopping algorithm framework, improves the probability of the traditional basin hopping algorithm searching for the nearby lowest local minimum, and avoids exploring a large number of unnecessary high-energy occupation configurations through relatively expensive DFT calculation. The present application can more efficiently find a lower-energy occupation configuration and promote the understanding of the structure-activity relationship of materials.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence structural optimization technology, and in particular to a basin jumping algorithm based on machine learning recommendation systems and its applications. Background Technology

[0002] Hydrogen is a green and clean energy source. Electrolysis of water is a clean method for producing hydrogen, but the oxygen evolution reaction (OER) at the anode is slow. An effective approach is to use a suitable catalyst to lower the OER reaction energy barrier and increase the reaction rate. Currently, commercially available catalysts are mainly noble metal oxide catalysts, which possess relatively good stability and catalytic activity, such as IrO2 and RuO2. However, the high cost of noble metals limits their widespread application; therefore, there is a need to find a low-cost catalyst with high catalytic performance.

[0003] Perovskite oxides have recently become a research hotspot due to their low cost, diverse chemical composition, vast configurational space, and high catalytic activity. The chemical formula of perovskite is ABX3. Here, A and B cover almost all metal cations in the periodic table, and X is typically a chalcogenide or a halogen. Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 O3 is a recently discovered OER catalytic material with high catalytic activity. However, this multi-cation perovskite is characterized by having multiple different cations in a sublattice, leading to disorder in the atomic arrangement or site configuration within the structure. This makes determining the site configuration difficult and hinders the understanding of the material's structure-property relationship.

[0004] Computational studies of placeholder configurations include special quasi-random structure methods and cluster expansion methods. Placeholder configuration optimization is a subfield of structure optimization, so structure optimization algorithms can also be used for placeholder configuration optimization, such as particle swarm optimization, evolutionary algorithms, simulated annealing, and basin jump algorithms. These structure optimization algorithms often require expensive DFT computations. The emergence of artificial intelligence has accelerated structure search, for example, by using machine learning potential functions to replace expensive DFT computations. However, there is currently limited research on incorporating machine learning into the basin jump global optimization algorithm. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, this invention provides a basin-hopping algorithm based on machine learning recommendation systems and its application, which solves at least some of the problems in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A basin-hopping algorithm based on machine learning recommendation systems includes the following steps:

[0008] Step 1: Obtain the initial random placeholder configuration for placeholder configuration optimization;

[0009] Step 2: Determine the parameters for real-time machine learning force field calculation;

[0010] Step 3: Determine the basin jump parameters, including temperature and jump rules;

[0011] Step 4: Using real-time machine learning force fields as a recommendation system, introduce the basin jump algorithm to apply the Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 Global optimization of the O3 spacer configuration;

[0012] Step 5: Validate the optimized minimum energy occupancy configuration using density functional theory (DFT) calculations.

[0013] In a further embodiment, step one, the initial random occupancy configuration includes: structural model size, number of atoms, lattice parameters, and atomic category information in the sublattice.

[0014] In a further embodiment, in step one, the initial random occupancy configuration is a 3×4×5 supercell structural model; the initial random occupancy configuration contains 300 atoms, and its lattice parameters are... Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 The A-site sublattice of O3 is occupied by Ba and Sr ions, the B-site sublattice is occupied by Co and Fe ions, and O acts as an "observer" in the anion sublattice.

[0015] In a further embodiment, in step two, the real-time machine learning force field parameters include: the training weight ratio of atomic forces and system energy, the atomic distribution of the broadened descriptor, and the number of basis functions of the broadened descriptor.

[0016] In a further scheme, in step three, during the process of determining the basin jump parameters, the temperature K is set to 2500K, and the jump rule is set to perform only one exchange between different cations in the same sublattice for each jump.

[0017] In a further embodiment, in step four, the action on Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 The O3 placeholder configuration is globally optimized, including the following steps:

[0018] (1) Implementation of the basin jump algorithm:

[0019] During each jump, an exchange is performed between two different cations in a sublattice of the occupied configuration, and the new configuration is either structurally relaxed or locally minimized.

[0020] If the Boltzmann factor exp(-ΔE / kT) is greater than a random number between (0, 1), then the new placeholder configuration will be accepted;

[0021] If the conditions are not met, the new placeholder configuration is rejected and the system returns to the previous local minimum state, repeating the jumping process until the preset maximum number of steps is reached.

[0022] in,

[0023] k represents the Boltzmann constant; T represents the temperature; ΔE is the difference between the current local minimum energy E and the previous local minimum energy.

[0024] (2) Relaxation / Local Minimum Process:

[0025] In the energy minimization process of each configuration, the structural relaxation of the first two steps is performed by DFT calculation, and the real-time machine learning force field (on-the-fly MLFF) is updated based on the obtained DFT data; then the remaining 15 relaxation steps are quickly completed using on-the-fly MLFF; this relaxation based on on-the-fly MLFF is effective because the local environment of the atom changes little.

[0026] (3) The recommendation system determines the jump direction:

[0027] Each time a "jump" occurs, 100 random pre-jumps are performed to obtain a new perturbation configuration. The energy of the perturbation configuration is then predicted using on-the-fly MLFF, and the configuration with the lowest predicted energy is taken as the direction of the next actual "jump".

[0028] In a further scheme, step five involves performing DFT calculations on the optimized lowest energy occupancy configuration and visualizing it. This helps in analyzing the relationship between structure and material properties.

[0029] On the other hand, the aforementioned basin-hopping algorithm based on a machine learning recommendation system is used in perovskite catalysts Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 Application in O3 occupancy configuration optimization.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. This invention introduces machine learning recommendation systems into the framework of conventional basin jump algorithms, which improves the probability of traditional basin jump algorithms searching for the nearest local minimum and avoids exploring a large number of unnecessary high-energy occupancy configurations through relatively expensive DFT calculations.

[0032] 2. The method of the present invention can find lower energy occupancy configurations more efficiently, and promote the understanding of the structure-property relationship of materials. Attached Figure Description

[0033] Figure 1 This is a flowchart of the basin jump global optimization algorithm based on machine learning recommendation system in Embodiment 1 of the present invention.

[0034] Figure 2 This is a comparison diagram of the novel method in Embodiment 1 of the present invention with the traditional basin jump method, special quasi-random structure method and random search method.

[0035] Figure 3 This is the RBH running result under the condition of changing the dataset size of the machine learning model in Embodiment 1 of the present invention.

[0036] Figure 4 This is a comparison between the prediction and DFT calculation results of the relaxation part machine learning system in Embodiment 1 of the present invention.

[0037] Figure 5 This is a comparison of the prediction energy and relaxation energy of the machine learning recommendation system in Embodiment 1 of the present invention.

[0038] Figure 6 This is a comparison diagram of the initial configuration of Embodiment 1 of the present invention and the lowest energy occupancy configuration obtained after optimization and verified by DFT calculation. Detailed Implementation

[0039] The following non-limiting embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way. The following description is merely an exemplary illustration of the scope of protection of the present invention, and those skilled in the art can make various changes and modifications to the invention based on the disclosed content, which should also fall within the scope of protection of the present invention.

[0040] Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2O3 is a recently discovered OER catalytic material with high catalytic activity. However, this multi-cation perovskite is characterized by multiple different cations in a sublattice, leading to disorder in the atomic arrangement or site configuration within the structure. This makes determining the site configuration difficult and hinders the understanding of the structure-property relationship of the material. Computational studies of site configurations include special quasi-random structure methods and cluster unrolling methods. Structure optimization algorithms such as particle swarm optimization, evolutionary algorithms, simulated annealing, and basin jump algorithms can also be used for site configuration optimization. These structure optimization algorithms often require expensive DFT calculations. The emergence of artificial intelligence has accelerated structure search, for example, by using machine learning potential functions to replace expensive DFT calculations.

[0041] This invention introduces a machine learning recommendation system into the framework of the conventional basin-hopping algorithm, abbreviated as RBH (recommendation-based basin-hopping). Unlike the traditional basin-hopping algorithm (BH), RBH accelerates structural relaxation or local minima by combining density functional theory (DFT) calculations with machine learning. Furthermore, the recommendation system determines the direction of the "jump." This recommendation system increases the probability of the BH algorithm searching for the nearest local minima and avoids exploring a large number of unnecessary high-energy occupied configurations through relatively expensive DFT calculations. This method can more efficiently find lower-energy occupied configurations, promoting a better understanding of the structure-property relationship of materials.

[0042] To address the above problems, this invention provides a basin-hopping algorithm based on a machine learning recommendation system, the flowchart of which is shown below. Figure 1 As shown in the figure, the black curve represents the original potential energy surface; the blue curve represents the transformed potential energy surface; the white circle represents the perturbation-occupied configuration; the green dot represents the energy local minimization configuration under the accepting state; the red dot represents the energy local minimization configuration under the rejecting state; the black dashed line represents the process of accelerating local minimization / relaxation by combining DFT calculation and on-the-fly MLFF; and the green arrow represents the machine learning recommendation system used to specify the direction of the next "jump".

[0043] A basin-hopping algorithm based on machine learning recommendation systems specifically includes the following steps:

[0044] Step 1: Obtain the initial random placeholder configuration for placeholder configuration optimization;

[0045] Step 2: Determine the calculation parameters for the real-time machine learning force field (on-the-fly MLFF);

[0046] Step 3: Determine the basin jump parameters, including temperature and jump rules;

[0047] Step 4: Using real-time machine learning force fields as a recommendation system, introduce the basin jump algorithm to apply the Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 Global optimization of the O3 spacer configuration;

[0048] Step 5: Perform DFT calculations to verify the optimized minimum energy occupancy configuration.

[0049] In a further embodiment, step one, the initial random occupancy configuration includes: structural model size, number of atoms, lattice parameters, and atomic category information in the sublattice.

[0050] In a further embodiment, in step one, the initial random occupancy configuration is a 3×4×5 supercell structural model; the initial random occupancy configuration contains 300 atoms, and its lattice parameters are... Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 The A-site sublattice of O3 is occupied by Ba and Sr ions, the B-site sublattice is occupied by Co and Fe ions, and O acts as an "observer" in the anion sublattice.

[0051] In a further embodiment, in step two, the real-time machine learning force field parameters include: the training weight ratio of atomic forces and system energy, the atomic distribution of the broadened descriptor, and the number of basis functions of the broadened descriptor.

[0052] In a further scheme, in step two, the training data for on-the-fly MLFF comes from the DFT data of structural relaxation during RBH operation. That is, DFT calculation, potential function construction and update and application are nested with each other, and structural search and potential function update are performed synchronously.

[0053] In a further scheme, in step two, the parameters for the on-the-fly MLFF include: "ML_IWEIGHT=3" for energy, force, and stress normalization; "ML_WTOTEN=100", "ML_WTIFOR=4", and "ML_WTSIF=1" correspond to the weights applied to energy, force, and stress during the on-the-fly MLFF training process, respectively. Since this paper focuses on the energy of the configuration, the weight for energy fitting accuracy is specifically increased; "ML_SION1=0.3" and "ML_SION2=0.3" are used to broaden the atomic distribution of the descriptor; and "ML_MRB1=12" and "ML_MRB2=12" are used to expand the number of basis functions of the descriptor.

[0054] In a further scheme, in step three, during the process of determining the basin jump parameters, the temperature K is set to 2500K, and the jump rule is set to perform only one exchange between different cations in the same sublattice for each jump.

[0055] In a further embodiment, in step four, the action on Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 The O3 placeholder configuration is globally optimized, including the following steps:

[0056] (1) Implementation of the basin jump algorithm:

[0057] Each "jump" involves exchanging two different cations in a sublattice of the occupier configuration, resulting in structural relaxation or local minimization of the new configuration. If the Boltzmann factor exp(-ΔE / kT) is greater than a random number between (0, 1), the new occupier configuration is accepted. If the condition is not met, the new occupier configuration is rejected and the system returns to the previous local minima. The "jump" operation is iterated until the preset maximum number of steps is reached.

[0058] In the Boltzmann factor, k represents the Boltzmann constant; T represents the temperature; and ΔE is the difference between the current local minimum energy E and the previous local minimum energy.

[0059] (2) Relaxation / Local Minimum Process:

[0060] In the energy minimization process for each configuration, the structural relaxation in the first two steps is performed by DFT calculation, and the on-the-fly MLFF is updated based on the obtained DFT data. Then, the remaining 15 relaxation steps are quickly completed using the on-the-fly MLFF. This relaxation based on the on-the-fly MLFF is effective because the local environment of the atom changes little.

[0061] (3) The recommendation system determines the jump direction:

[0062] Each "jump" involves 100 random pre-jumps to obtain new perturbation configurations. The energy of these new configurations is predicted using on-the-fly MLFF, and the configuration with the lowest predicted energy is chosen as the direction for the next actual "jump." This recommendation system improves the probability of the BH algorithm searching for the nearest local minimum and avoids exploring a large number of unnecessary high-energy occupancy configurations through relatively expensive DFT computations.

[0063] In a further scheme, step five involves performing DFT calculations on the optimized lowest energy occupancy configuration and visualizing it. This helps in analyzing the relationship between structure and material properties.

[0064] The following embodiments are illustrated in conjunction with the accompanying drawings and Ba. 0.5 Sr 0.5 Co 0.8 Fe 0.2 O3 introduces the basin jump algorithm based on machine learning recommendation system of the present invention.

[0065] Example 1

[0066] A basin-hopping algorithm based on machine learning recommendation systems includes the following steps:

[0067] Step 1: Obtain the initial random placeholder configuration for placeholder configuration optimization.

[0068] Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 O3 has a cubic structure with an initial occupancy configuration of a 3×4×5 supercell containing 300 atoms: 30 Ba and Sr ions; 48 Co ions; 12 Fe ions; and 180 oxygen ions. The lattice parameters of a single unit cell are... There are two sublattices, with sublattice A occupied by Ba and Sr ions, and sublattice B occupied by Co and Fe ions. Oxygen ions are "bystanders". The ions in the initial occupied configuration are randomly arranged.

[0069] Step 2: Determine the computational parameters for the on-the-fly MLFF (Machine Learning Force Field), including the training weights of atomic forces and system energy, the atomic distribution of the broadened descriptor, and the number of basis functions for the broadened descriptor.

[0070] The training data for on-the-fly MLFF comes from the DFT data of the structure relaxation during the operation of RBH or BH. That is, DFT calculation, function construction and update and application are nested with each other, and the structure search and potential function update are performed synchronously.

[0071] The parameters for on-the-fly MLFF include: "ML_IWEIGHT=3" for energy, force, and stress normalization; "ML_WTOTEN=100", "ML_WTIFOR=4", and "ML_WTSIF=1" correspond to the weights applied to energy, force, and stress during on-the-fly MLFF training, respectively. Since this paper focuses on the energy of the configuration, the weight for energy fitting accuracy is intentionally increased; "ML_SION1=0.3" and "ML_SION2=0.3" are used to broaden the atomic distribution of the descriptor; and "ML_MRB1=12" and "ML_MRB2=12" are used to expand the number of basis functions of the descriptor.

[0072] Step 3: Determine the basin jump parameters. Set the temperature K to 2500K and the jump rule to allow only one exchange between different cations in the same sublattice for each jump.

[0073] Step 4: Using real-time machine learning force fields as a recommendation system, introduce the basin jump algorithm to apply the Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 The O3 placeholder configuration is globally optimized.

[0074] (1) Implementation of the basin jump algorithm:

[0075] Each "jump" involves exchanging two different cations in a sublattice of the occupier configuration, resulting in structural relaxation or local minimization of the new configuration. If the Boltzmann factor exp(-ΔE / kT) is greater than a random number between (0, 1), the new occupier configuration is accepted. If the condition is not met, the new occupier configuration is rejected and the system returns to the previous local minima. The "jump" operation is iterated until the preset maximum number of steps is reached.

[0076] In the Boltzmann factor, k represents the Boltzmann constant; T represents the temperature; and ΔE is the difference between the current local minimum energy E and the previous local minimum energy.

[0077] (2) Relaxation / Local Minimum Process:

[0078] In the energy minimization process for each configuration, the structural relaxation in the first two steps is performed using DFT calculations, and the on-the-fly MLFF is updated based on the obtained DFT data. Then, the remaining 15 relaxation steps are quickly completed using the on-the-fly MLFF. In this way, the energy minimization process is accelerated with the assistance of machine learning. The maximum number of relaxation steps allowed for each energy minimization is set to 15. It has been verified that the convergence of most perturbation configurations is within 15 steps.

[0079] (3) The recommendation system determines the jump direction:

[0080] The direction of the "jump" is specified by a recommendation system. Each "jump" involves 100 random pre-jumps to obtain new perturbation configurations, predicting their energy using on-the-fly MLFF, and selecting the configuration with the lowest predicted energy as the actual "jump" direction for the next time. This recommendation system improves the probability of the BH algorithm searching for the nearest local minimum and avoids exploring a large number of unnecessary high-energy occupancy configurations through relatively expensive DFT computations.

[0081] Step 5: Perform DFT calculations to verify the optimized minimum energy occupancy configuration.

[0082] The results of RBH are compared with those of traditional BH, special quasi-random structures, and cluster expansion methods, such as... Figure 2 As shown. Figure 2 (a) is the energy distribution of the configurations generated by random search and SQS. Figure 2 (b) is the traditional BH method for Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 Three runs of O3 occupancy configuration optimization Figure 2 (c) shows three runs of the RBH method for site configuration optimization. The figure shows that the configuration energy distribution curve of the new method is significantly lower than that of other traditional methods. This demonstrates the effectiveness of RBH.

[0083] Previously, the RBH method used a dataset with a maximum of 30 configurations for training on-the-fly MLFF. This invention expands this to 60 configurations, and the RBH results are as follows: Figure 3 As shown in the figure. The results show that the RBH results are similar to those with 30 configuration data points. Whether with 30 or 60 configuration data points, RBH can effectively optimize the placeholder configuration. This indicates that even with small datasets, on-the-fly MLFF can still achieve high accuracy and effectively determine the jump direction.

[0084] Proof of the validity of on-the-fly MLFF in the RBH method. For example... Figure 4 As shown, the predictions of the machine learning system and the DFT calculation results were compared in the relaxation part, and the prediction error was found to be 0.31 eV, or 1 meV / atom, indicating high accuracy. The predictions and relaxation energy of the machine learning recommendation system were compared in the "jump" part. Figure 5 As shown, more than one-third of the data is within the 0.3eV error range, which is acceptable for a recommendation system.

[0085] The initial random occupancy configuration is compared with the lowest energy occupancy configuration confirmed by DFT. For example... Figure 6 As shown, (a) is the initial random site configuration, and (b) is the lowest energy site configuration. Cyan represents Ba ions; pink represents Sr ions; green represents Co ions; yellow represents Fe ions; and gray represents O ions. The results showed that Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2In O3, the Ba and Sr ions at the A site are arranged in a layered manner, which can reduce the bond stress of Ba-O and Sr-O; while the Co and Fe ions at the B site are arranged in a disordered manner, with configuration entropy being the main driving force.

[0086] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A basin-hopping method based on a machine learning recommendation system, characterized in that, Includes the following steps: Step 1: Obtain the initial random placeholder configuration for placeholder configuration optimization; Step 2: Determine the parameters for real-time machine learning force field calculation; Step 3: Determine the basin jump parameters, including temperature and jump rules; Step 4: Using a real-time machine learning force field as a recommendation system, the basin jump method is introduced to globally optimize the system's site-occupying configuration; in Step 4, the system in the global optimization of the system's site-occupying configuration using the basin jump method is Ba. 0.5 Sr 0.5 Co 0.8 Fe 0.2 O3; the Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 The initial random occupancy configuration of O3 is a 3×4×5 supercell structural model; the initial random occupancy configuration contains 300 atoms and its lattice parameter is 3.98 Å; Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 The A-site sublattice of O3 is occupied by Ba and Sr ions, the B-site sublattice is occupied by Co and Fe ions, and O acts as an "observer" in the anion sublattice. Step four, the step of globally optimizing the system's site-occupying configuration, includes: (1) Implementation of the basin jump method: During each jump, an exchange is performed between two different cations in a sublattice of the occupied configuration, and the new configuration is either structurally relaxed or locally minimized. If the Boltzmann factor is satisfied If the random number is greater than (0, 1), the new placeholder configuration will be accepted; If the conditions are not met, the new placeholder configuration is rejected and the system returns to the previous local minimum state, repeating the jumping process until the preset maximum number of steps is reached. in, k represents the Boltzmann constant; T represents the temperature; ΔE is the difference between the current local minimum energy E and the previous local minimum energy. (2) Relaxation / Local Minimum Process: In the energy minimization process of each configuration, the structural relaxation of the first two steps is performed by DFT calculation, and the real-time machine learning force field on-the-fly MLFF is updated according to the obtained DFT data; then the remaining 15 relaxation steps are quickly completed using on-the-fly MLFF; this relaxation based on on-the-fly MLFF is effective because the local environment of the atom changes little. (3) The recommendation system determines the jump direction: Each time a "jump" occurs, 100 random pre-jumps are performed to obtain a new perturbation configuration. The energy of the perturbation configuration is predicted using on-the-flyMLFF, and the position configuration with the lowest predicted energy is taken as the direction of the next actual "jump". Step 5: Perform density functional theory calculations to verify the optimized minimum energy occupancy configuration.

2. The basin-hopping method based on a machine learning recommendation system according to claim 1, characterized in that, In step one, the initial random occupancy configuration includes: structural model size, number of atoms, lattice parameters, and atomic category information in the sublattice.

3. The basin-hopping method based on a machine learning recommendation system according to claim 1, characterized in that, In step two, the real-time machine learning force field calculation parameters include: the training weight ratio of atomic forces and system energy, the atomic distribution of the broadened descriptor, and the number of basis functions of the broadened descriptor.

4. The basin-hopping method based on a machine learning recommendation system according to claim 1, characterized in that, In step three, during the process of determining the basin jump parameters, the temperature K is set to 2500K, and the jump rule is set to perform only one position exchange between different cations in the same sublattice for each jump.

5. The basin-hopping method based on a machine learning recommendation system according to claim 1, characterized in that, Step five includes: processing the obtained optimized Ba... 0.5 Sr 0.5 Co 0.8 Fe 0.2 Density functional theory calculations were performed on the lowest energy-occupying configuration of O3, and the lowest energy-occupying configuration was visualized.

6. The basin-hopping method based on a machine learning recommendation system according to any one of claims 1-5 in the perovskite catalyst Ba 0.5 Sr 0.5 Co 0.8 Fe 0.2 Application in O3 occupancy configuration optimization.

Images