AI material discovery: hype check

When Google DeepMind, a pioneering AI firm, announced nearly two years ago that a deep-learning AI technique had led to the discovery of 2.2 million novel crystalline materials, it seemed like a revolutionary moment for materials research. This collection included 52,000 layered compound simulations akin to graphene, a remarkable substance, along with 528 potential lithium-ion conductors that could enhance rechargeable batteries, and much more.

However, this initiative, along with similar endeavors from Microsoft and Meta, soon faced criticism. Researchers argued that some AI-generated compounds lacked originality, practicality, or suitability. According to Anthony Cheetham, a materials scientist at UCSB, many of the hypothetical crystals on DeepMind’s list were “ridiculous.” Cheetham and his UCSB colleague Ram Seshadri pointed out that over 18,000 predicted compounds contained extremely scarce radioactive elements like promethium and protactinium, making them unlikely to be useful materials.

Cheetham stated that discovering a compound differs significantly from discovering a new functional material. Meta's work proposed over 100 materials with the potential to capture carbon dioxide from the atmosphere, aiding in global warming reduction. Yet, these suggestions also faced criticism. Computational chemist Berend Smit from EPFL noted that the candidates weren't viable for this purpose, implying that the AI tool's novelty might have overshadowed practical considerations.

So, will AI transform materials discovery, or is it overhyped? Since the initial critiques, scientists have more thoroughly assessed AI’s potential. The teams involved have responded by moderating initial claims or suggesting alternative approaches. Many scientists believe that AI holds significant promise for materials science. However, they emphasize that closer collaboration with experimental chemists, along with acknowledging the current limitations of these systems, is crucial to fully realize AI's potential.

AI's Role in Materials Science

Throughout history, from the Bronze Age's copper and tin blend to stainless steel, material discovery has fueled innovation. The use of AI in materials science has grown significantly in the last decade. Recent AI-driven efforts to accelerate material discovery primarily focus on crystalline inorganic solids, essential in various technologies like semiconductors and lasers.

The properties of these solids are determined by their atomic composition and arrangement in repeating patterns. Scientists planning new inorganic crystals aim to predict the structure atoms will adopt, rather than just creating new combinations of atoms. Before AI, researchers relied on conventional computational methods, such as density functional theory (DFT), to approximate electron behavior in materials.

DFT can reveal a hypothetical compound's most stable structure and properties. It has been used to predict materials with exceptional properties, including strong magnets and superconductors that don’t require extremely cold temperatures. The Materials Project at LBNL has an open-access database containing DFT-calculated structures for about 200,000 crystals.

However, DFT requires significant computational resources. While academic labs can perform DFT calculations on a few compounds, analyzing millions would be prohibitively expensive.

AI Approaches to Material Discovery

This is where AI becomes valuable. DeepMind’s approach, for instance, used a machine-learning algorithm (GNoME) that learned from existing DFT calculations to predict the stability of randomly generated crystal structures much faster than conventional DFT. Promising predictions were then checked using DFT, with the results feeding back into GNoME to improve performance, enabling the system to envision a vast array of potentially stable compounds.

According to Kristin Persson at LBNL and UC Berkeley, using such methods will soon be essential. DeepMind researchers have also employed AI to aid in material synthesis. Persson co-authored a paper describing A-Lab, a robotic system that learned to create recipes for synthesizing target compounds, using tens of thousands of published papers. Robots then made and analyzed these compounds, adjusting recipes as needed.

Shortly after, Microsoft introduced MatterGen, an AI tool for materials discovery. Similar to GNoME, MatterGen is a machine-learning model designed to generate stable crystal structures but is more targeted, suggesting hypothetical materials with specific properties. Tian Xie from Microsoft Research AI for Science explained that this approach is more efficient than creating millions of candidates through brute force.

Meta's project focused on identifying metal–organic frameworks (MOFs) for CO2 capture. Researchers used DFT to assess the CO2-binding ability of over 8,000 MOFs and trained an AI model, demonstrating similar accuracy with greater speed than DFT. A May 2024 paper predicted that over 100 of these MOFs would strongly bind to CO2, proving AI tools could expedite MOF development for direct air capture.

Controversies and Challenges

These AI-driven explorations have sparked controversy. Robert Palgrave, a solid-state chemist at University College London, found that the A-Lab project mischaracterized some of the 41 inorganic compounds it claimed to have produced, and in some instances, synthesized materials that were already known. Palgrave, along with Leslie Schoop from Princeton University, critiqued A-Lab's work, citing characterization shortfalls and concluding that the project didn't discover new materials.

They highlighted a fundamental issue with DFT, which typically predicts highly ordered crystal structures that are stable only at extremely low temperatures (-273 °C). Palgrave noted that while the ordered DFT structures A-Lab aimed to create seemed new, they existed as known, disordered structures, which A-Lab eventually made.

Gerbrand Ceder, co-leader of the A-Lab work, disagreed, stating that reanalysis confirmed the reliability of A-Lab's characterizations. He asserted that A-Lab successfully made the compounds it claimed, even without synthesis information, and that creating disordered versions of predicted ordered compounds is a recognized success.

Johannes Margraf, a computational chemist, noted that disorder also affects AI-based DFT surrogates like GNoME. His team trained a machine-learning system on experimentally determined crystal structures and found that 80–84% of about 380,000 stable compounds highlighted by DeepMind as synthesis targets would likely be disordered in reality. This suggests that many GNoME suggestions might not be realized in ordered forms and could have different properties than predicted.

Margraf added that AI models trained on DFT data can overlook useful properties arising from structural disorder, which can lead to both false negatives and false positives.

Ekin Dogus Cubuk, a lead author of the GNoME paper, acknowledged that many predicted ordered structures would likely be disordered. He stated that the tool's main purpose is to guide further investigation of promising compounds.

Some criticized DeepMind's claim of achieving “an order-of-magnitude expansion in stable materials known to humanity” as overly optimistic. Jonathan Godwin, a former DeepMind employee, found it implausible that unsynthesized materials could be considered new. A DeepMind spokesperson noted that over 700 GNoME-predicted compounds were independently created by other researchers, and that GNoME structures guided the synthesis of novel caesium-based compounds with potential applications in optoelectronics and energy storage.