News · October 2026
Anthropic AI for Science award supports machine-learning potentials for ultra-high-temperature ceramics
The Kroll Group has received an award from Anthropic’s AI for Science program to develop machine-learning interatomic potentials (MLIPs) for the chemical space behind polymer-derived ultra-high-temperature ceramics. The award provides Claude API credits that will be used across the first year of a three-year project led by Peter Kroll with postdoctoral associate Dr. Shariq Haseen.
Ultra-high-temperature ceramics based on metal borides, carbides, and nitrides must hold up above 2000 °C in hypersonic and propulsion environments. Making them from pre-ceramic polymers allows molecular-level control over composition and near-net-shape processing, but the chemistry that decides the final microstructure – pyrolysis, oxidation, carbon segregation, nucleation of boride and nitride phases – plays out over length and time scales that quantum-mechanical simulation cannot reach. Density functional theory handles a few hundred atoms for nanoseconds; the processes of interest involve millions of atoms over microseconds. Thermodynamics tells us which phases a ceramic wants to form. Whether it actually forms them is a question of how atoms move through disordered, reacting material, and that is where current methods fall short.
MLIPs close this gap. Trained on DFT energies and forces along reaction trajectories, they reproduce quantum-mechanical accuracy at a tiny fraction of the cost, which makes large-scale reactive molecular dynamics tractable. The group has already built and validated such potentials for Si–C–O–H, Si–C–N–H, and Zr–Si–C–N–H pyrolysis. The new project extends this work to the full multicomponent space and further develops the derivation of moment tensor potentials.
The practical bottleneck in building potentials is not only compute but curation: choosing which of more than 500 million configurations of the group’s DFT library should be included, diagnosing where a potential fails, and feeding those failures back into training. In the project, Claude takes on this supervising role in the active-learning loop – clustering configurations by chemistry and extrapolation grade, computing error statistics against held-out trajectories, ranking candidate calculations by expected information gain, and maintaining the tooling that connects all computations and the code base.
The potentials will be released as LAMMPS-compatible files, and the group intends to publish the active-learning pipeline itself, including an account of where the AI supervision helped and where it did not. Production simulations run under the group’s ACCESS allocation.