Approaches using graph neural networks (GNNs) are studied. GNN has the potential to significantly reduce calculation costs while maintaining calculation accuracy by representing the atomic and molecular structures ...
Materials informatics (MI) is attracting attention as a new approach in materials science, and progress has been made in building predictive models using machine learning while utilizing first-principles calculations and molecular dynamics (MD) calculations. However, there are some major challenges to putting MI into practical use. First, although first-principles calculations (DFT) have very high accuracy, the calculation time increases exponentially as the system grows, making it unsuitable for large-scale materials exploration. Second, in molecular dynamics (MD) calculations, there is a problem in which the accuracy of calculations decreases significantly if an appropriate force field does not exist. Furthermore, a method that utilizes neural network potential (NNP) has appeared, but in order to make highly accurate predictions, it is necessary to obtain a large amount of first-principles calculation results as learning data and data collection have become a major bottleneck.
Against this background, new approaches using graph neural networks (GNNs) are expected. GNN has the potential to significantly reduce calculation costs while maintaining calculation accuracy by representing the atomic and molecular structures of materials as graphs and building machine learning models. Among these, M3GNET (Materials Graph Neural Network) is attracting attention as a method that can improve the efficiency of large-scale material searches while having accuracy equivalent to first-principles calculations.
In this article, we predict and analyze the physical properties of high-entropy alloys (HEA) using M3GNET and verify their effectiveness. We will also consider the possibility that M3GNET can be applied to fields such as MI. HEA is a new material group containing multiple major elements, and while it has excellent properties, it has a high compositional diversity, making it difficult to find the optimal composition. We will explain how this search process is accelerated by using M3GNET, using specific examples.
In this study, we created a slab model (288 atoms) based on platinum (Pt) and evaluated the combination of elements that affect catalyst properties and structural stability. Specifically, we uniformly mixed elements such as Co, Cr, Mo, Mn, Cu, Fe, Zn, Sn, Pd, Ga, and Au at a ratio of 5 to 10 wt% and investigated their effects.
When constructing the slab model, the experimental conditions were reproduced by fixing the bottom three layers of Pt, which served as the substrate, and giving the surface a degree of freedom. In addition, in order to evaluate OH adsorption, we considered three types of adsorption sites: Top, Hollow, and Bridge, and compared their adsorption energies.
Furthermore, we aimed to statistically analyze the effects of different elemental compositions from the viewpoints of adsorption energy (E1) and structural stability (E2), and to clarify the optimal composition conditions. Adsorption energy (E1) was determined from formula (1), and structural stability (E2) was determined using formula (2) below. Since E(OH) is a common term, it is omitted here.
E1 = E(slab + OH) – E(slab) – E(OH) （１）
E2 = E(slab) – (E(each FCC structure) * Number of atoms of each element in slab model/4) (2)
In this study, we conducted molecular dynamics (MD) simulations using LAMMPS-linked M3GNET provided by NanoLabo. M3GNet-MP-2021.2.8-PES potential was applied to perform structural optimization. For calculations, a single NVIDIA A100 GPU was used to achieve high-speed energy calculations and optimization processes. In addition, parallel calculations were performed for different compositions to efficiently predict physical properties under a wide range of conditions.
In this research, we performed structural optimization of a five-element slab model and optimized M3GNET using A100 GPU. Calculations for the slab model shown in Fig. 1 were completed in just 4 minutes and 50 seconds.
Taking advantage of this high calculation speed, we prepared 300 types of basic slab models for each five-element slab model, and calculated 300 models for each of the three types of adsorption sites: Top, Hollow, and Bridge. As a result, we were able to evaluate a total of 1,200 models and conduct detailed analysis on seven types.
Furthermore, by operating multiple A100 GPUs in parallel in the Mat3ra environment, it became possible to collect data on 8,400 models in just 3 days while changing the type and wt% of each element. This is a process that would take several weeks using traditional methods, demonstrating that M3GNET's high-speed computational capabilities are extremely effective in large-scale materials exploration. For the collected data, the energy was calculated for each of the three types of adsorption sites: Top, Hollow, and Bridge, and the average value was evaluated as the adsorption energy (E1). This allowed us to quantitatively compare the catalytic properties of each composition and serve as an index for identifying the optimal element combination. By changing the wt% of each element in the 288-atom model, we were able to analyze in detail the influence of each element on the catalytic properties and adsorption energy. The figure below shows the change in adsorption energy (E1) under conditions where the wt% of each element is different. When the proportion of Pt is high, the adsorption energy is somewhat better (Fig. 2), but as the proportion of elements 1 and 4 increases, it can be observed that the adsorption energy tends to decrease. (Fig.3)
Next, we comprehensively evaluated the adsorption energy (E1) and structural stability (E2) for the seven models. E1 in each model serves as an indicator of catalytic activity, and E2 is an important factor in stable structure formation. Figures 4 and 5 visually show the changes in E1 and E2, respectively, providing a detailed view of how different wt% elemental ratios affect the catalytic properties and structural stability. In particular, the adsorption energy (E1) was highly dependent on the proportion of Pt and the composition of other elements, and significant variations were observed with increasing specific elements. Furthermore, in the evaluation of structural stability (E2), it was confirmed that the stability was significantly improved under certain compositional conditions, suggesting that it may contribute to improving the long-term durability of the catalyst. In particular, when specific elements shown by the purple line were included, the adsorption energy (E1) and structural stability (E2) tended to improve significantly.
Although we cannot disclose the details of the specific elements, the experimental results using these combinations showed very high agreement with the simulations. This suggests that the predictions made by M3GNET are consistent with experimental data and are promising as a practical material exploration method.
In this study, we predicted and analyzed the physical properties of high-entropy alloys (HEA) using M3GNET, and demonstrated that calculations are significantly faster than first-principles calculations. In particular, it has been revealed that by operating M3GNET using multiple GPUs and parallel calculations, large-scale composition screening of thousands of models can be performed within just a few days.
Using a 288-atom model, we quantitatively analyzed the influence of each element on catalytic properties and adsorption energy while changing the wt% of each element. After considering seven different models, we found that even if the proportion of Pt was lowered, the increase in certain elements affected the stability. Therefore, by analyzing the correlation between adsorption energy (E1) and structural stability (E2), it was suggested that a specific elemental composition may promote the formation of a stable structure and improve catalytic performance.
By utilizing M3GNET, the composition search process, which previously took several months using first-principles calculations, has been dramatically shortened, making it possible to develop efficient catalyst materials. In particular, by making full use of the GPU installed in Mat3ra, it will be easier to explore complex composition spaces, and material design is expected to be accelerated. In the future, it is thought that by applying this method to even more diverse compositions and conditions, it will become possible to apply it to more optimal material design.
keywords:
#M3GNET #High-entropy alloy #Materials informatics #MI #First-principles calculation #Molecular dynamics #Neural network potential #Graph neural network #Adsorption energy #Fast screening #Material design #Catalytic properties #Data-driven materials exploration
The content is prepared by kazuki.mori@mat3ra.com
