Computer programs that learn patterns from examples instead of following rules written by hand. In materials research they are trained on thousands of calculations or measurements. They then predict in a fraction of a second what would take hours or days to compute: whether a crystal might peel into sheets, how wide its is, how its atoms move. They also pick out thin in microscope images and choose the next growth experiment to try. Their answers are only as good as the examples they learnt from, so surprising predictions still have to be checked.
A memristor is a device whose electrical resistance depends on the current that has flowed through it before, so it remembers, and an array of them can perform the multiply-and-add operations at the heart of neural networks directly in memory.
Computational exfoliation screens and 2D databases catalogue thousands of candidate monolayers with stability, band structures, magnetism and topology, and increasingly feed generative models and active-learning discovery loops.
can compute the properties of a known crystal in hours; databases such as C2DB and 2DMatPedia hold such results for thousands of 2D candidates. A model trained on them – often a graph neural network that reads the crystal as atoms joined by bonds – predicts the same quantities for a new structure in a fraction of a second, so millions of hypothetical compounds can be sorted and the promising few passed back to full calculation.
At the largest scale, a model trained on materials databases and its own calculations proposed in 2023 several hundred thousand crystals predicted to be stable that no database contained. Whether most are new, useful or makeable is debated: stability in a calculation is not a route to making the material, and many of the predictions turned out to be close variants of known compounds.
The second large use is : models fitted to DFT energies and forces that then drive molecular dynamics at a small fraction of the cost, making it possible to follow the relaxation of a of a million atoms or the motion of defects over nanoseconds. The entry on molecular dynamics explains how they are built and where they fail.
In the laboratory, image-recognition models scan optical micrographs for flakes of the right thickness, a job that used to take hours by eye, and have been combined with robots that pick up and stack the flakes they find. Bayesian optimisation and active learning choose the next growth or synthesis condition from the results so far, reaching a good recipe in fewer runs than a grid of experiments would.
A model interpolates well inside the range of its examples and can be confidently wrong outside it. A potential trained only on perfect crystals may tear a defect apart; a band-gap model trained on semilocal DFT inherits its systematic underestimate of gaps. Test sets that share near-duplicates with the training set make the accuracy look better than it is, and a structure predicted to be stable says nothing about whether it can be made.
Good practice is to report the uncertainty of each prediction, to test on data the model has not seen in any form – ideally new calculations or experiments made after it was trained – and to treat a surprising prediction as a hypothesis, not a result.
Statistical models – from kernel and tree-based regression to graph neural networks and transformers – fitted to data in order to predict properties, propose structures or choose experiments. For the main uses are screening crystal databases for exfoliable compounds and their properties, with models trained on high-throughput DFT sets such as C2DB and 2DMatPedia; machine-learned interatomic potentials, which reproduce DFT forces at a small fraction of the cost and make million-atom moiré and defect simulations possible; segmentation of optical, and electron-microscope images; and active learning or Bayesian optimisation to choose the next synthesis run.
The usual failure modes are extrapolation beyond the training data, leakage between training and test sets, and errors of the underlying DFT that the model inherits; calibrated uncertainties and prospective tests against new calculations or experiments are the remedies.
The newest items in the site’s news feed that use the term, one from each source.
Atomic reconstruction reshapes moiré materials across multiple scales, from local structure and polarization textures to global electronic , yet direct \textit{ab initio} modeling becomes prohibitive for large superstructures such as marginal- moirés and moiré-of-moirés. We develop MoiréMLIP by…
Atomic reconstruction reshapes moiré materials across multiple scales, from local structure and polarization textures to global electronic topology, yet direct \textit{ab initio} modeling becomes prohibitive for large superstructures such as marginal-twist-angle moirés and moiré-of-moirés. We develop MoiréMLIP by…
We investigate the nanoscale mechanisms determining in-plane lattice (LTC) of pristine and W-doped MX –M X from first principles, using the exact solution of the linearised Boltzmann transport equation in both and relaxon bases. For the W-doped…