Researchers have used artificial intelligence to predict dozens of previously unknown carbon crystal structures, including at least one with a calculated hardness exceeding that of diamond. The findings, drawn from separate but converging lines of computational research, suggest that machine learning paired with evolutionary algorithms can rapidly identify stable carbon arrangements that traditional lab methods might never find. If any of these predicted materials can be synthesized, the implications for cutting tools, electronics, and industrial drilling could be significant.
Evolutionary Algorithms Meet Machine Learning
The foundational work behind these predictions traces back to a study in npj Computational Materials, in which researchers integrated the XTALOPT evolutionary crystal-structure prediction tool with a machine-learning model trained within the AFLOW high-throughput materials framework. The approach worked in two stages: XTALOPT generated candidate crystal geometries through an evolutionary process that mimics natural selection, while the machine-learning component evaluated elastic properties and estimated Vickers hardness for each structure. According to reporting from the University at Buffalo, the team reported 43 new carbon allotropes, with several predicted to be superhard. That count, however, carries a caveat: the structures exist only as computational predictions, and no experimental confirmation of synthesis has been published for any of them.
What makes this method distinct from earlier crystal-prediction efforts is the feedback loop between generation and evaluation. Rather than producing thousands of random structures and then running expensive quantum-mechanical calculations on each one, the machine-learning filter screens candidates in real time. Structures unlikely to be hard are discarded early, which lets the evolutionary search focus computational resources on the most promising geometries. The result is a far more efficient scan of the vast space of possible carbon arrangements, where even a tiny fraction of viable structures can represent entirely new classes of materials.
Within this computational landscape, the predicted allotropes span a range of densities and bonding patterns. Some resemble distorted versions of diamond’s familiar tetrahedral network, while others feature more complex ring systems and channels. The machine-learning models infer hardness from how these networks distribute stress, highlighting configurations where strong, directional bonds create resistance to both compression and shear. Even if only a subset of these structures prove experimentally accessible, they expand the conceptual toolkit for designing carbon-based materials.
CrystaLLM and the Push Beyond Diamond
A newer AI-driven framework known as CrystaLLM has pushed the search further into unexplored territory. In work described on Phys.org, researchers used this system to discover several allotropes with combinations of exotic properties, including at least one whose calculated hardness exceeds even that of diamond. That claim is the headline result, but it demands careful framing. “Calculated hardness” here refers to values derived from models of bonding and elastic response, rather than from direct indentation tests on a physical crystal. The difference matters: simulations can overestimate performance if they miss defects, grain boundaries, or subtle instabilities that arise during real-world synthesis.
CrystaLLM builds on language-model concepts to represent crystal structures as sequences that can be learned and modified. By treating atomic arrangements somewhat like sentences, the framework can generate new “phrases” in the language of crystallography, then score them for stability and mechanical properties. This generative capability, coupled with physics-informed filters, allows the system to explore structural motifs that might not emerge from more conventional evolutionary algorithms alone.
Crucially, the CrystaLLM results do more than list hypothetical lattices. The same coverage of the project notes that the framework points to potential synthesis pathways, such as pressure–temperature regimes or precursor phases that could transform into the desired allotropes. By identifying not just what structures might exist but how they might be made, the approach narrows the path from digital design to laboratory reality. Still, as of the latest reports, no peer-reviewed experimental confirmation of these specific superhard structures has appeared, leaving them in the realm of informed prediction.
Why Carbon Keeps Surprising Scientists
Carbon’s ability to form wildly different materials from the same element is not new. Graphite and diamond are both pure carbon, yet one is soft enough to mark paper while the other is the hardest known natural material. Over recent decades, theorists have added a growing catalog of proposed carbon crystal families, such as M-carbon, W-carbon, Z-carbon, and bct-carbon, many of which are summarized in a review in National Science Review. That work traces how advances in computational methods have steadily expanded the list of plausible allotropes, while experimentalists use high-pressure synthesis and X-ray diffraction to hunt for matching phases in the lab.
The reason so many predicted structures remain unconfirmed is partly practical. Creating new carbon phases typically requires extreme pressures and temperatures, conditions that are expensive and difficult to sustain long enough to produce measurable crystals. Even when researchers do create unusual forms of carbon in diamond-anvil cells or shock-compression setups, the products may be nanocrystalline, mixed with other phases, or too small for straightforward characterization. Matching an experimental X-ray diffraction pattern to a theoretical structure can become a complex puzzle with many nearly overlapping solutions.
Those challenges are compounded by the very properties that make superhard materials attractive. As highlighted in a ScienceDaily report, superhard solids can slice, drill, and polish other objects, making them valuable for machining and wear-resistant coatings. Yet the same resistance to deformation that underpins these applications also makes such materials difficult to grow, cut, or polish into the high-quality samples needed for detailed study. This tension, between industrial utility and experimental tractability, helps explain why theory often runs ahead of confirmed discoveries in carbon science.
Exascale Computing Expands the Search
Another thread in this story comes from large-scale simulations of how carbon behaves under extreme conditions. At Argonne National Laboratory, researchers have been using exascale-class supercomputers and AI tools to simulate the evolution pathways of post-detonation nanodiamonds. Their work maps how tiny diamond particles produced in explosive environments can transform into carbon nano-onions, carbon dots, and other forms under intense thermodynamic stress. Rather than proposing entirely new crystal topologies, this line of research focuses on phase transitions among known or closely related structures.
The Argonne simulations matter for the broader prediction effort because they test AI models against processes that can be probed experimentally. If machine-learning methods can reproduce the observed evolution of nanodiamonds as they anneal, coalesce, or graphitize, that success builds confidence in the same classes of models when they are applied to more speculative structures. Conversely, discrepancies between simulation and experiment can reveal missing physics or biases in the training data, prompting refinements that ultimately improve predictive power.
Exascale computing also changes the scale of what is computationally tractable. Instead of simulating a handful of candidate structures in detail, researchers can now explore vast swaths of parameter space, running many shorter simulations that collectively map out how carbon responds across a range of pressures, temperatures, and defect landscapes. This statistical view complements the more targeted crystal-structure predictions from tools like XTALOPT and CrystaLLM, linking idealized lattices to the messy realities of growth, shock, and degradation.
From Prediction to Practice
For now, the superhard carbon allotropes identified by AI remain virtual. Turning them into practical materials will require coordinated advances in high-pressure synthesis, in situ diagnostics, and post-synthesis characterization. Experimentalists will need to design pressure–temperature pathways that follow the theoretical suggestions, then verify whether the resulting products match the predicted structures within experimental uncertainty.
Even partial success could be impactful. A carbon phase that is easier to synthesize than diamond but nearly as hard, or one that combines high hardness with unusual electronic or thermal properties, could find roles in cutting tools, heat spreaders, or protective coatings. AI-guided discovery does not guarantee such outcomes, but it does stack the odds by focusing effort on the most promising regions of an enormous design space.
The convergence of evolutionary algorithms, generative models like CrystaLLM, and exascale simulations signals a shift in how materials are discovered. Instead of relying primarily on serendipity or incremental tweaks to known compounds, researchers can now start from desired properties and work backward to candidate structures and synthesis routes. In carbon, where the combinatorial possibilities are immense, this strategy is already revealing architectures that stretch beyond diamond’s long-held benchmark. The coming years will show how many of these digital crystals can be coaxed into physical form, and whether any of them can truly out-hard the hardest natural material known.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
