New tool · Johns Hopkins University
ALIGNN 2.0
Dependency-free PyTorch rebuild of ALIGNN unifies property, force-field, spectral and generative prediction in one graph
· 9 claims checked against 1 source
What it is
ALIGNN 2.0 is a dependency-free, pure-PyTorch reimplementation of the Atomistic Line Graph Neural Network (ALIGNN), built by Kamal Choudhary and colleagues at Johns Hopkins University's Department of Materials Science and Engineering and Department of Electrical and Computer Engineering. The rewrite removes the specialized graph libraries that made the original ALIGNN difficult to install or run on newer accelerators, including Blackwell-class GPUs.
How it works
The atom graph and the angular line graph, along with their batching, are constructed directly as tensors. A single shared per-atom representation feeds separate heads for scalar properties, spectra, tensors, per-atom quantities, and conservative forces. The same network also serves as the denoiser inside a conditional crystal-diffusion model used for inverse design.
What is new
- Leads the original ALIGNN on 26 of 30 JARVIS-Leaderboard single-property benchmarks.
- Largest gains appear on piezoelectric maxima, where MAE falls from about 20 to 12.5, along with dielectric response, exfoliation energy, and superconducting Tc.
- The paired force field, ALIGNN-FF, matches universal potentials on Matbench-Discovery and CHIPS-FF with a fraction of their parameters, and scales to hundred-thousand-atom cells in LAMMPS and OpenMM.
- The framework reconstructs infrared, Raman, optical-dielectric, and neutron spectra from a single relaxed structure, in agreement with experiment and DFT.
Limits
coNGN, a connectivity-optimized line-graph competitor, still reaches the lowest MAE on several tasks, including total energy, energy above hull, bulk and shear moduli, magnetic moment, several dielectric components, the Seebeck coefficient, and phonon heat capacity. ALIGNN 2.0 is not uniformly the best model across the JARVIS-Leaderboard.
Availability
Code is available on GitHub and a web app is provided, both named in the preprint, which is available now.
The evidence
Each statement above was written with the sentence from the source that supports it, and a program (not a model) checked that every quote and every figure appears in the pages it read. Here they are.
- 01
ALIGNN 2.0 is a dependency-free, pure-PyTorch reimplementation of ALIGNN with the line graph and its batching built from scratch.
“a dependency-free, pure-PyTorch reimplementation of the Atomistic Line Graph Neural Network, with the line graph and its batching built from scratch” S1
- 02
ALIGNN 2.0 leads on 26 of 30 single-property benchmarks on the JARVIS-Leaderboard against the original ALIGNN.
“ALIGNN 2.0 leads on 26 of 30 single-property benchmarks against the original ALIGNN” S1
- 03
The maximum piezoelectric modulus MAE improves from around 20 in ALIGNN 1.0 to about 12.5 in ALIGNN 2.0.
“the maximum piezoelectric modulus improves from around 20 in the ALIGNN 1.0 to about 12.5 in ALIGNN 2.0” S1
- 04
ALIGNN-FF matches leading universal potentials on Matbench-Discovery and CHIPS-FF with a small fraction of their parameters while scaling to hundred-thousand-atom cells.
“The LAMMPS- and OpenMM-compatible ALIGNN-FF matches leading universal potentials on the Matbench-Discovery and CHIPS-FF benchmarks at a small fraction of their parameters while scaling to hundred-thousand-atom cells” S1
- 05
ALIGNN 2.0 is used as the denoiser in a conditional crystal-diffusion model for inverse design.
“We further use ALIGNN 2.0 as the denoiser in a conditional crystal-diffusion model” S1
- 06
From a single relaxed structure the framework reconstructs infrared, Raman, optical-dielectric, and neutron spectra in agreement with experiment and DFT.
“the same framework reconstructs infrared, Raman, optical-dielectric, and neutron spectra in agreement with experiment and DFT” S1
- 07
coNGN reaches the lowest MAE on total energy, energy above hull, bulk and shear moduli, magnetic moment, several dielectric components, Seebeck coefficient, and phonon heat capacity.
“coNGN is a strong baseline: it reaches the lowest MAE on a number of the property tasks (total energy, energy above hull, bulk and shear moduli, magnetic moment, several dielectric components, the thermoelectric Seebeck coefficient, and phonon heat capacity)” S1
- 08
Code is available on GitHub and a web app is provided at atomgpt.org/alignn.
“Web app: https://atomgpt.org/alignn. Code: https://github.com/atomgptlab/alignn” S1
- 09
Kamal Choudhary is the corresponding author, affiliated with Johns Hopkins University's Department of Materials Science and Engineering and Department of Electrical and Computer Engineering.
“Kamal Choudhary Affiliation: Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, Maryland 21218, USA Affiliation: Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, Maryland 21218, USA Affiliation: Corresponding author: kchoudh2@jhu.edu” S1
