Transforming materials development with cutting-edge simulation and causal discovery technologies

Traditional materials development has primarily relied on a trial-and-error approach, heavily dependent on researchers' experience and intuition, requiring significant time and effort. In recent years, simulations have increasingly been deployed to predict material behavior and clarify its mechanisms. This approach aims to gain important insights to help determine which materials offer the most promising prototyping potential. In addition, the combination of simulation and AI is advancing rapidly, allowing for the quicker exploration of promising material candidates from a vast number of options. However, there are still deep-seated challenges in AI simulations for materials development. Firstly, there is a shortage of personnel possessing both the advanced expertise in computational chemistry, which is necessary for simulations, and AI itself. Additionally, improving AI prediction accuracy requires a large amount of high-quality training data, which is both time-consuming and costly to create. Moreover, you also need large-scale and prolonged simulations to unravel the reactions occurring in complex structured materials and their time-dependent behavioral changes. For example, some of these conventional methods would have taken over 100 billion years to compute!

── How are Fujitsu's technologies used in materials development?

Yuta: The materials development process typically begins with generating a large number of possible structures based on a requirement like "we want a material with this kind of property or characteristic." Next, these candidates are screened using simulations and other methods to narrow down the most promising candidates, and finally, they are verified through experiments. To accelerate materials development, Fujitsu has developed two tools: "a tool that easily creates AI for simulations" and "a tool to discover causal relationships between material properties and structure."

The first tool is for creating AI (Machine Learning Interatomic Potential) for molecular dynamics simulations, which predict material properties. Molecular dynamics simulation is a method that calculates forces acting on atoms and reproduces atomic behavior to predict material properties highly accurately. In materials development, there is a high demand for simulating materials composed of numerous atoms, resembling actual materials. This tool can easily create AI that enables long-duration simulations, reproducing material changes over time, for materials consisting of tens of thousands to millions of atoms. Since it automatically generates the dataset required for learning itself, there is no need to prepare large amounts of data beforehand. Furthermore, it allows easy data generation and learning even without specialized AI knowledge. The second tool is an outcomes analysis tool based on causality. It analyzes the complex causal relationships between material structure and properties (for example, how the "cause" of a material's structure influences its chemical reactions and physical properties as an "effect"), using a vast amount of simulation and experimental results. Then, Fujitsu's proprietary causal discovery technology (*1) extracts particularly important causal relationships. Moreover, by visualizing these causal relationships, it becomes immediately clear which structures affect which properties, thereby assisting in determining what materials should be prototyped next.

── What kind of molecular dynamics simulations can be achieved with the first tool?
Yuta: It's a tool that can predict material properties. There are two methods for predicting material properties in materials development. One method, based on quantum mechanics (first-principles calculations), is theoretically accurate but requires an enormous amount of computation time. On the other hand, methods based on classical mechanics, which are conventional physical laws (classical molecular dynamics calculations), are fast but have accuracy issues.

── How are you approaching the challenge of increasing computational speed while also improving accuracy?
Meguru: Our team is focusing on "Machine Learning Interatomic Potential," which balances the accuracy of quantum mechanics with the computational speed of classical mechanics. A Machine Learning Interatomic Potential is an AI model designed to predict interactions between atoms. By training the AI with data generated from first-principles calculations, it enables predictions that are faster than first-principles calculations and more accurate than models based on human empirical rules or experimental data. Based on this approach, we developed GeNNIP4MD (Generator of Neural Network Interatomic Potential for Molecular Dynamics), a tool that creates AI for molecular dynamics simulations to evaluate dynamic properties and chemical reactions of materials. GeNNIP4MD achieves high-accuracy, long-duration simulations by constructing a neural network that predicts interatomic interactions with the same accuracy as quantum chemical calculations but at high speed.

── What are the features and advantages of GeNNIP4MD compared to conventional molecular dynamics simulations?
Meguru: In conventional AI-driven simulations, building a high-accuracy AI model (Machine Learning Interatomic Potential) requires specialized knowledge, and creating the training data for it incurs significant costs. GeNNIP4MD's key feature is that it automates a series of processes, from the efficient generation of necessary training data to AI model training and accuracy evaluation. Furthermore, it includes a function to efficiently fine-tune (additionally train) the AI model with a small amount of extra data. This makes it possible to build highly accurate AI models easily, even for those without AI expertise. This AI model enables longer molecular dynamics simulations for large-scale material structures. This, in turn, allows for the analysis of complex phenomena (such as polymer entanglement and material fracture) and the reproduction of changes that occur as materials are used over long periods (such as strength degradation and structural collapse). As a result, GeNNIP4MD efficiently explores and evaluates new materials through simulations, reducing experimental trial and error and significantly shortening the overall development period.

── What efforts were made to improve the accuracy of the simulations?
Naoki: In general, large-scale and long-duration simulations based on conventional AI models often become highly unstable, ultimately leading to the collapse of the material structure. This AI simulation instability occurs when the AI encounters data it has not been trained on, resulting in the AI predicting abnormal values. To solve this problem, we adopted an approach of pre-teaching the AI physically important knowledge. Specifically, we devised a method for selecting the training data provided to the AI. Based on our existing knowledge, we improved GeNNIP4MD's algorithm to make it easier for the AI to prioritize learning physically important data. As a result, even if the AI encounters an unknown structure during simulation, it can handle it with its pre-learned knowledge, and the stability of the calculation has dramatically improved.

── Please tell us about a customer case study utilizing GeNNIP4MD.
Naoki: I’d like to illustrate this with a case study from Nippon Steel Corporation on the analysis of hydrogen embrittlement in "nickel and manganese alloys." Hydrogen is attracting attention as a clean energy source that emits no carbon dioxide during combustion, and the entire supply chain, from hydrogen production, transport, storage, to utilization, is rapidly being developed. However, there is a significant challenge, principally hydrogen embrittlement (a phenomenon where metal materials absorb hydrogen atoms, leading to a decrease in strength and ultimately unexpected fracture) in metal materials for hydrogen-related infrastructure such as pipes, valves, and storage tanks used in high-pressure hydrogen environments. To solve this problem and build safer, more reliable hydrogen-related infrastructure, Nippon Steel Corporation was looking to detail the mechanism of hydrogen behavior within materials at the atomic level.

── How was GeNNIP4MD utilized?
Naoki: They created target structures for analysis (structures with varying nickel and manganese ratios, and structures with added hydrogen), and used GeNNIP4MD to create AI models. Simulations utilizing these AI models allowed them to observe the behavior of hydrogen in nickel-manganese alloys, leading to the elucidation of the hydrogen embrittlement mechanism. As a result, new insights have emerged, such as "adding manganese makes hydrogen embrittlement less likely to occur" and "what amount of manganese are effective." In this way, we work alongside our customers, predicting material properties that were previously unpredictable and gaining new insights by observing atomic-level behavior. These achievements have in fact already been published in several academic papers (*2) (*3).

── What are the future plans for the Materials Informatics?
Yuta: We want to enable end-to-end, automatic simulations where, if you tell the AI what kind of material you want, the AI verifies it through simulation, analyzes the causal relationships in the simulation data, and clarify the mechanism. In 10 or 20 years, I hope we can provide solutions that leverage quantum computing.

Meguru: By expanding the scale of simulations and bringing it closer to reality, we aim to enable the development of more efficient solar cell materials than ever before, and simulations of semiconductors targeting entire devices. By analyzing and controlling chemical reactions in more detail, we ultimately want to lead to the creation of innovative devices that we can't even imagine.

Naoki: First, I aim to complete the research we are currently conducting on "reducing AI training costs through advanced data sampling methods." In addition to that, I want to enhance the functions of GeNNIP4MD further, have it utilized in real-world settings, and contribute to solving societal issues.

Sotaro: My goal is to advance the "autonomous exploration cycle" further, where AI considers mechanisms from simulation results and automatically proposes the optimal next experiment to try. In addition, I want to combine this with my hobbies, applying this exploration cycle to familiar processes, such as creating cooking recipes, and challenge new possibilities beyond the scope of materials science.