# ACCESS: Stampede3 Simulations Advance AI Models for Extreme Hypersonic Ceramics — retained record

> Retained news record from news:hpcwire.

Canonical: https://highsignal.app/data/records/133bd9eb2bce310a

[← sources](https://highsignal.app/data)


source: [news](https://highsignal.app/data/news) (news:hpcwire)  published 2026-06-30 18:24 ingested 2026-07-01 09:42

[original source ↗](https://www.hpcwire.com/off-the-wire/access-stampede3-simulations-advance-ai-models-for-extreme-hypersonic-ceramics/)

## retained text



```
Off the Wire Press Releases
June 30, 2026 — Hypersonic flight exposes materials to brutal conditions. Once a vehicle is screaming through the atmosphere above Mach 5, it’s wrapped in a sheath of superheated gas that can reach temperatures hot enough to melt many metals and chemically eat away at exposed surfaces in seconds.
To make this kind of flight practical, engineers are turning to advanced ceramics for the outer “skin” and control surfaces. Increasingly, researchers like Peter Kroll at The University of Texas at Arlington are designing ceramics with the help of artificial intelligence running on powerful high-performance computing (HPC) systems like those available by the U.S. National Science Foundation (NSF) ACCESS program.
“At the speeds and temperatures involved in hypersonics, several problems hit at once,” said Kroll, professor in the Department of Chemistry and Biochemistry at UT Arlington. “The air itself becomes a chemically aggressive plasma, heating the vehicle, attacking materials at the surface and sometimes blacking out radio communications.”
Kroll explained that ultra-high-temperature ceramics and ceramic composites can tolerate far higher temperatures than traditional alloys, but they are notoriously difficult to design, process and test because their behavior depends sensitively on chemistry, microstructure and how they are actually made. Answering what might sound like simple questions, such as “Which ceramic recipe should we choose?” and “How will it behave during flight?”, quickly becomes a challenge that can only be tackled in detail with large-scale simulation and huge amounts of compute units.
Traditionally, researchers rely on quantum-mechanical calculations, such as density functional theory, to predict how atoms bond and how materials respond under extreme conditions. These methods are accurate but slow; they might handle a few hundred atoms over a few nanoseconds of simulated time, which is far too small and too short to represent a real component.
On the other side of the spectrum, classical models can simulate millions of atoms for much longer, but they often miss important chemistry, especially in complex ceramic systems containing multiple elements like silicon, carbon and nitrogen. That’s where machine-learning interatomic potentials, essentially AI models that learn how atoms interact from high-quality quantum data, are beginning to change the story.
Kroll’s recent study, which was published in the Journal of the American Ceramic Society, showcases this approach for a family of materials known as polymer-derived ceramics, which are made by heating special polymers until they convert into ceramic.
The team used NSF ACCESS allocations to build a new machine-learning model for a four-element system of silicon, carbon, nitrogen and hydrogen, tailored specifically to follow these materials as they transform from soft polymers into hard ceramics. Specifically, they utilized the Stampede3 supercomputer at the Texas Advanced Computing Center (TACC) to train the model on more than 2,000 carefully chosen atomic configurations, including crystalline structures, disordered glasses, real preceramic polymers and high-temperature snapshots from quantum-mechanical simulations, capturing over a million force values at quantum accuracy.
“With the Stampede3-trained model in hand, we could run simulations that would be impossible with traditional quantum methods alone,” Kroll said. “For example, we studied how atoms move in amorphous silicon nitride using a model containing more than 7,000 atoms and still reproduced diffusion behavior that agrees with earlier, much smaller quantum calculations.”
He said that they also examined an unusual ceramic that shrinks when heated, a carbodiimide phase called β-Si(NCN)₂, and found that their model predicts thermal expansion values close to experiment. In another test, they built a virtual composite containing tens of thousands of atoms of silicon carbide and silicon nitride grains and watched how it responds to stress from room temperature up to 2,300 Kelvin, revealing a transition from brittle fracture at low temperature to more plastic deformation at high temperature.
Perhaps most striking for real-world applications, the researchers used their model to mimic the actual processing of a commercial preceramic polymer known as Durazane 1800. They constructed a realistic network of polymer chains, heated it virtually up to 2,500 degrees Kelvin, and followed it over tens of nanoseconds of simulated time while allowing small molecules to escape, just as they would in a furnace. By the end of the run, the model produced an amorphous silicon-carbon-nitrogen ceramic with pockets of graphitic carbon and overall composition and energy that lined up with experimental measurements of similar materials. In effect, they created a “digital twin” of the polymer-to-ceramic process, opening the door to optimizing recipes and heating schedules on a supercomputer instead of through trial-and-error in the lab.
This work fits into a broader wave of innovation in hypersonic ceramics highlighted by the American Ceramic Society. Around the world, teams are combining AI, advanced simulation and novel manufacturing methods to push the temperature and durability limits of thermal-protection materials. Some groups are exploring ultra-high-temperature carbides and nitrides in chemical systems like hafnium–tantalum–carbon–nitrogen, where machine-learning-accelerated simulations have predicted compositions with melting points above 4,400 Kelvin by showing how nitrogen can raise melting enthalpy and lower the disorder of the liquid phase. Others are experimenting with laser-based processing routes, such as selectively converting liquid precursors into hafnium carbide coatings using a high-power laser in open air, a fast and flexible approach that nonetheless requires detailed thermal and chemical modeling to control. High-entropy ceramics, which mix five or more different metal species, and hybrid composites that combine ultra-high-temperature carbides with polymer-derived ceramic matrices are also in play, vastly expanding the design space for hypersonic-capable materials.
Across all of these efforts, HPC is a common backbone.
Training a sophisticated interatomic potential demands many thousands of quantum-mechanical calculations, each itself a demanding workload on CPU or GPU clusters. Once trained, these models are deployed inside molecular dynamics codes such as LAMMPS, where they drive long-running, massively parallel simulations that can fill nodes for days as they follow bond breaking, diffusion and microstructural evolution. The software stacks are complex, often combining electronic-structure packages, machine-learning tools, and production-scale simulation engines, and they generate large, carefully curated datasets that become valuable community resources for future model development.
For hypersonic programs, the payoff is a much tighter loop between modeling, processing and testing. Rather than relying purely on expensive furnace campaigns and high-enthalpy tunnel shots to explore the space of possible materials and process conditions, teams can now use AI-enhanced simulations to narrow the field and understand failure modes before hardware is built. Experimental work then shifts toward validating the most promising candidates and exploring edge cases, accelerating the path from lab concepts – like polymer-derived ceramics with embedded carbon phases – to flight-ready leading edges, control surfaces and protective tiles.
As hypersonic systems move closer to operational reality, the limiting factor will not be just aerodynamics or propulsion but the materials that stand between the vehicle and the surrounding plasma. Machine-learning models for complex ceramics, trained and run on modern HPC platforms, are emerging as critical tools in that race, giving scientists a way to design future “plasma shields” from the level of individual atoms all the way up to real components.
“Supercomputing resources like Stampede3 are essential for this kind of research and make it happen,” Kroll said. “Harvesting data at this level of complexity – capturing chemical reactions and phase transitions across millions of atoms in multi-component systems – is what enables us to address future challenges, and it demands computational power far beyond what any single lab can provide.”
Resource Provider Institution(s): Texas Advanced Computing Center
Resources Used: Stampede3
Affiliations: University of Texas at Arlington
Funding Agency: NSF
Grant or Allocation Number(s): DMR190103
The science story featured here was enabled by the U.S. National Science Foundation’s ACCESS program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.
Source: Kimberly Mann Bruch, SDSC; ACCESS
```



document fetched 2026-07-01 09:42 · canonical https://hpcwire.com/off-the-wire/access-stampede3-simulations-advance-ai-models-for-extreme-hypersonic-ceramics