activity
20242026
most citedPhononic materials with effectively scale-separated hierarchical features using interpretable machine learning

1 citations · 1 across the 4 of their papers we have counts for

collaborators

4 papers