Disordered atomic structures can now be inverse-designed by property: GaDiT samples two orders of magnitude faster and exposes structure–property clues
Synopsis
The configuration space of disordered materials is vast, and conventional molecular dynamics or Monte Carlo explores it through sequential updates along single trajectories, while rapid quenching traps networks in high-energy defect states and obscures structure–property relations. GaDiT combines a graph transformer on dynamic periodic graphs with masked conditioning and an interval-weighted training-noise distribution, predicting atomic noise vectors directly to generate structures in two-dimensional amorphous carbon and three-dimensional amorphous silicon and silica, sampling two orders of magnitude faster than corrector-based approaches. One model supports single and combined conditioning and generates structure–property combinations beyond the training set in amorphous carbon.
Interpretation
Disordered atomic structures can now be sampled directly by a generative model rather than through trajectory-by-trajectory molecular dynamics or Monte Carlo simulation. The earlier workflow generated structures with molecular dynamics or Monte Carlo and then computed properties, with each run exploring configuration space through sequential updates, so collective rearrangements across energy barriers were rare. For two-dimensional amorphous monolayer carbon, generated networks after fixed-cell relaxation show a mean atomic root-mean-square displacement of 0.0236 Å and a mean energy decrease of 0.0253 eV atom-1, with 95.40% three-coordinated atoms against 98.05% in the training pool.
Generation no longer needs the second derivatives and stochastic corrector steps of energy-based models, which is where the speedup comes from. Energy-based diffusion models obtain denoising forces by differentiating a learned scalar energy, requiring second-order derivatives in training and coordinate gradients at every sampling step, and sometimes add Hamiltonian Monte Carlo correctors that cost thousands of extra evaluations per structure. On an identical GPU, null-mode GaDiT generates an amorphous carbon structure more than two orders of magnitude faster than AMDEN with Hamiltonian Monte Carlo corrections and also faster than a corrector-free AMDEN variant; its Wasserstein distances for relaxed energy and six-membered-ring fraction are less than half those of AMDEN-HMC and less than one-third those of eRSI.
One shared network generates structures under any subset of structural features, physical properties or preparation conditions, and combines several targets for inverse design. Most conditional models assume a fixed set of input labels, so a different combination of targets requires retraining. In amorphous carbon, single-condition generation gives a linear-fit slope of 0.948 for elastic modulus with a post-relaxation targeting mean absolute error of 9.31 N m-1; under joint modulus and six-membered-ring-fraction conditioning, the ring-fraction spread at a fixed modulus is substantially broader than under either single condition, reaching combinations absent from the training set.
Perspective
The demonstrated scope is two-dimensional amorphous monolayer carbon plus three-dimensional amorphous silicon and silica, with training data from melt–quench molecular dynamics and AMDEN source structures and property labels from relaxation and strain calculations using classical potentials such as Tersoff.
Generated energy distributions remain narrower than the deliberately broad training set, and cooling-rate conditioning matches the molecular-dynamics logarithmic trend only within the cooling-rate range covered by training.
