Can AI find the conditions that make colloids form quasicrystals?
A neural network that recognises phases from their diffraction patterns can steer simulations to the narrow conditions where particles self-assemble into a chosen structure, even quasicrystals.
Source
Inverse design of soft materials via a deep learning-based evolutionary strategy
Study at a glance
- Design
- Computational / modelling — Monte Carlo simulations of 2D shoulder-potential particles and 3D coronal spherocylinders, with a CNN diffraction classifier as fitness for a CMA-ES optimizer over pressure, temperature, density and shoulder width
- N
- Simulation boxes of 256 particles in 2D and 432 particles in 3D; not a participant count
- Population
- Model colloidal particles: hard-core square-shoulder and softened-core-shoulder disks, and soft-corona spherocylinders
- Outcome
- Whether the optimizer converges to the stability region of a target phase (crystals, quasicrystals, liquid crystals)
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors trained a convolutional neural network to classify simulated phases (fluid, square and hexagonal crystals, and 12-, 10- and 18-fold quasicrystals) from their diffraction patterns. The network's probability for a target phase was used as the fitness in an evolutionary optimizer (CMA-ES) that adjusts pressure, temperature, density or interaction range, running new Monte Carlo simulations each generation. They tested it on the model used for training, on a different interaction potential, and on a 3D system of rod-like particles.
What they found
The classifier sorted all phases with 100% accuracy, and starting from a fluid the search reached the tiny 12-fold quasicrystal region in about 25 generations regardless of starting point. The same network, without retraining, found the 12-fold quasicrystal in a different softened potential within five generations. Letting it also tune the shoulder width revealed a decagonal quasicrystal not previously predicted for that model, and in 3D it located a layered 12-fold quasicrystal phase of rods.
The limits
What it doesn't show
All results are simulations of idealised model potentials with a few hundred particles, not real colloids, so experimental realisability is untested. The method can only target phases the network was trained on and could converge wrongly if an unseen phase is misclassified as the target. Free energies were not computed, so the newly found decagonal phase's thermodynamic stability rests on the classifier and snapshots.
Key terms
- Inverse design
- Working backwards from a desired structure to the particle interactions and conditions that produce it.
- Quasicrystal
- An ordered but non-periodic structure with rotational symmetries (such as 10- or 12-fold) forbidden to ordinary crystals.
- Structure factor / diffraction pattern
- The Fourier-space fingerprint of particle positions that reveals a phase's symmetry.
- CMA-ES
- An evolutionary optimizer that samples parameters from a Gaussian and updates its mean and covariance toward high-fitness regions.
- Order parameter
- A quantity that distinguishes one phase from another.
Flashcards
0 of 9 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What did the CNN take as input?
Common questions
Why use diffraction patterns instead of particle positions?
Diffraction patterns are a translation-invariant fingerprint of global order, and quasicrystal particle positions are not known in advance to compare with.
Does the CNN have to be retrained for every new model?
No; the network trained on one potential guided the search in a different softened potential without retraining.
What was genuinely new?
A decagonal quasicrystal in the softened-core-shoulder model, which had not been predicted before.
More on Phase transitions