סמינר בחומר מעובה: Learning to predict superconductivity

Predicting the superconducting transition temperature from crystal structure and composition remains unsolved, leaving room for data-driven progress. I will describe two complementary approaches designed to keep interpretability at the center, which converge on the same conclusion: remarkably few descriptors are needed. In the first, experimentally measured crystal structures are encoded as histograms of elemental and geometric properties over atoms and bonds, and the transition temperature is learned with a Gaussian process. The model tells us which of its inputs matter, and 67 descriptors can be pruned to 4 with almost no loss of accuracy. The survivors are chemically transparent, with the leading one being the distribution of electron affinity differences between neighboring atoms. Guided by this model, we predicted, synthesized, and measured a new superconductor, PtPb₃Bi, with a transition at 3 K in a structure type that had never been reported before. In the second approach, we used unsupervised machine learning. A geometry-preserving autoencoder compresses roughly two thousand features of the inorganic crystal structure database onto a three-dimensional manifold whose axes remain physically readable. Superconductors separate from ordinary materials, and families separate from one another, even when no superconductor appears in training. The three coordinates alone suffice to predict the transition temperature accurately. Together these results suggest that the materials landscape has far fewer effective directions than its microscopic complexity implies, and that those directions sit close enough to bonding chemistry to guide synthesis.