The Era of AI Driving Materials Research Has Arrived?

About two years ago, DeepMind, a subsidiary of Google, announced the discovery of 2.2 million new crystal materials using deep learning technology. Earlier this year, Microsoft claimed that its AI model MatterGen can generate inorganic materials from scratch, potentially revolutionizing the paradigm of inorganic material design.

The new era of materials research driven by artificial intelligence (AI) seems to have begun, but criticism has also emerged. Critics argue that some compounds proposed by AI lack originality and practicality. Will AI completely transform the field of materials discovery, or will it become a case of overhyped expectations? A recent report on the British website Nature pointed out that most researchers acknowledge the enormous potential of AI in materials science, but it needs to work closely with experimental chemists while also recognizing current limitations of AI and continuously improving, in order to fully unleash its potential.

AI-driven materials design boom

Before the intervention of AI, researchers primarily relied on "Density Functional Theory" (DFT), a traditional computational method, to predict new materials and their properties. DFT had predicted high-quality new materials such as super magnets and superconductors.

However, DFT calculations are extremely computationally intensive, and screening millions of compounds at once would be unimaginably costly, highlighting the value of AI. DeepMind has developed the "Graph Network of Materials Exploration" (GNoME) AI system, which has discovered 2.2 million novel crystalline materials, encompassing various elements from the periodic table. Among these are 52,000 graphene-like layered compounds and 528 lithium-ion conductors that hold promise for improving rechargeable battery performance.

The Lawrence Berkeley National Laboratory has developed the A-Lab robotic system. This system masters the ability to design formulas by studying thousands of synthesis papers on inorganic compounds and can synthesize compounds that have been predicted by DFT but have never been prepared before. At the same time, A-Lab can control robots to perform experiments, analyze whether the products meet standards, and adjust formulas for closed-loop optimization when necessary.

Shortly after the publication of the GNoME and A-Lab papers, Microsoft launched an AI tool called MatterGen. Compared to GNoME, MatterGen is more targeted, capable of directly generating materials that meet design specifications. Scientists can not only specify the type of material but also set requirements for mechanical, electrical, magnetic, and other properties, providing a powerful tool for precise research and development. Additionally, the foundational AI team of a metaverse platform company collaborated with Georgia Institute of Technology, focusing on "Metal-Organic Frameworks" (MOFs), to predict over 100 MOF structures with strong carbon dioxide adsorption. This supports AI in accelerating the development of direct air capture carbon technology.

The Debate between Originality and Practicality

Despite the strong momentum of exploration by industry giants, controversy has never ceased. Many scientists bluntly state that some AI systems' envisioned compounds lack originality and practical value.

Materials scientist Anthony Cheetham from the University of California, Santa Barbara, and others discovered that the AI predictions from DeepMind's hypothetical crystal list included more than 18,000 compounds containing rare radioactive elements such as promethium and actinium, questioning their practical value. Robert Palgrave, a solid-state chemist at University College London, also pointed out errors in the descriptions of some materials when verifying the A-Lab's research results, noting that some of the 41 synthesized inorganic compounds were already known materials that had been synthesized previously.

In response, A-Lab personnel stated that further analysis proved that A-Lab's description of the material properties is reliable and that the claimed compound was indeed synthesized. A spokesperson for DeepMind said that over 700 compounds predicted by GNoME have been independently synthesized by other researchers, and the model has also guided the discovery of several unknown cesium-based compounds, which hold promise for use in the optoelectronic and energy storage fields.

Microsoft's MatterGen has also become embroiled in controversy. During testing, the team had it recommend new materials with specific hardness, and it synthesized a "tantalum chromium oxide" disordered compound. However, a preprint paper in June this year pointed out that this material was first synthesized as early as 1972 and was even included in MatterGen's training data.

The collaboration project between the metaverse platform company and Georgia Tech has also faced scrutiny. Berend Smit, a computational chemist at the Swiss Federal Institute of Technology in Lausanne, confirmed through calculations that the new materials proposed in the collaboration project cannot achieve direct air capture. The model overestimates the material's ability to bind with carbon dioxide, partly due to errors in the foundational database used for training.

Practical application needs to break through multiple barriers.

Despite the controversy, most researchers still believe that with continuous optimization, AI models will strongly drive the advancement of materials science.

To ensure the reliability of AI results, the Microsoft team developed a supporting AI system called MatterSim, specifically to verify whether the structures proposed by MatterGen are stable under real temperature and pressure conditions. However, even if the AI-assisted material discoveries are proven effective, humans still face significant challenges: such as how to optimize the process according to market demands, and how to achieve large-scale manufacturing of new materials and integrate them into commercial products.

Citrine Informatics, an American company, is using its AI system to help clients optimize existing materials and manufacturing processes. The company's CEO, Greg Mulholland, stated that each client has a customized Citrine model, which is trained on the client's proprietary experimental data and incorporates researchers' "chemical intuition" to enhance AI judgment.

Undeniably, the urgent demand for new materials in society will continue to drive AI exploration in this field. Many major social challenges currently faced by humanity are constrained by material bottlenecks. Scientists hope to leverage AI to design advanced materials that can be produced at scale and truly impact daily life, thereby realizing the true value of AI in the field of materials science.