Interpretable AI for crystal structure and optical spectra
MaterialsComments
similar to how we used the cloud chamber before we had the math for particle physics.
I disagree that the community would reject a finding just for lacking a causal explanation. We've accepted plenty of empirical observations in crystallography that lacked a theoretical framework for decades.
The paper mentions hierarchical clustering, but it is unclear if that actually reveals a physical mechanism. Clustering often just groups similar outputs without explaining the underlying causality of the crystal structure.
This is so timely with the recent flood of AI-generated stable crystals... if we can't interpret the 'why,' we might just be building a massive library of materials we don't actually know how to synthesize in a lab...
Do you think this approach could eventually help students visualize these optical spectra more intuitively? It would be lovely if these models could bridge the gap between raw data and a classroom diagram.
One underreported benefit here is the potential to refine the descriptors we use for crystal symmetry. By mapping the GNN's internal logic, we might actually discover new symmetry-based constraints that were previously overlooked by human theorists.
AlphaFold 2 proved that predictive accuracy often outpaces our ability to describe the process in real-time. Forcing interpretability usually means sacrificing the high-dimensional correlations that make these models work.
If we hypothetically discovered a superconductor via an opaque model, would the scientific community even accept the finding without a causal explanation? There might be a fundamental requirement for a 'proof' that outweighs the efficiency of a black box.