Neural network designs five hard metallic glasses, all confirmed by experiment

The work

Technical · from the record

A neural network called VIBANN, combining a self-attention encoder with a variational information bottleneck, was trained on 673 literature records of bulk metallic glass composition, indentation load and Vickers hardness (538 training, 135 held-out test; test R about 0.943 and mean absolute error about 55.6 HV). Its latent space was sampled with Gaussian-mixture-guided MCMC and refined by gradient-based optimisation of an uncertainty-penalised hardness objective, and decoded compositions were filtered on latent support, novelty and uncertainty. Five B-Nb-Fe-based compositions were arc melted and suction cast into 2 mm rods that X-ray diffraction and TEM showed to be amorphous, with B68Nb24Fe4W4 measuring 2447 ± 44 HV at 0.5 N; measured hardness across alloys and loads lay within the model's prediction intervals. Machine-learning-potential molecular dynamics was then used to relate the hardness ordering across the five alloys to local packing, coordination and bond-angle statistics.

How AI was used

A curated literature dataset of bulk metallic glass compositions, indentation loads and Vickers hardness values was encoded as 56-dimensional simplex-normalised composition vectors with a standardised scalar load and split by composition-based clustering. An attention network with a variational information bottleneck was trained from scratch to predict standardised hardness while a decoder head reconstructed the composition from the 16-dimensional latent vector, with the bottleneck weight set by a feedback controller targeting a fixed KL per dimension, hyperparameters chosen by Optuna, and epistemic uncertainty obtained from Monte Carlo dropout passes with batch normalisation frozen. For inverse design at a fixed 0.5 N load, a three-component Gaussian mixture model was fitted to the training latents to define distributional support; mixture-aware multi-chain MCMC proposed latent points under a risk-aware utility combining predictive mean and uncertainty, selected seeds were refined by backpropagating an uncertainty-aware objective into the latent coordinate and re-decoding, and surviving candidates had to pass latent-likelihood, novelty-distance, uncertainty-cap and lower-confidence-bound gates before synthesis. The trained model was also run to predict hardness for the synthesised alloys across loads, to produce attention scores, integrated-gradient attributions and latent traversals, and separately a pre-trained SevenNet machine learning interatomic potential drove LAMMPS molecular dynamics quenches of the five compositions for structural analysis.