Abstract
Deep neural networks (DNNs) have shown marked achievements across numerous research and commercial settings. Part of their success is due to their ability to “learn” internal representations of the input (x) that are ideal to attain an accurate approximation ((Formula presented.)) of some unknown function (f) that is, y = f(x). Despite their universal approximation capability, a drawback of DNNs is that they are black boxes, and it is unknown how or why they work. Thus, the physics discovered by the DNN remains hidden. Here, the condition of causality is enforced through a Lorentz layer incorporated within a deep neural network. This Lorentz NN (LNN) takes in the geometry of an alldielectric metasurface, and outputs the causal frequencydependent permittivity (Formula presented.) and permeability (Formula presented.). Additionally, this LNN gives the spatial dispersion (k) inherent in the effective material parameters, as well as the Lorentz terms, which constitute both (Formula presented.) and (Formula presented.). The ability of the LNN to learn metasurface physics is demonstrated through several examples, and the results are compared to theory and simulations.
Original language  English 

Article number  2200097 
Journal  Advanced Optical Materials 
Volume  10 
Issue number  13 
DOIs  
State  Published  Jul 4 2022 
Keywords
 Lorentzian oscillators
 deep learning
 metamaterials
 metasurface physics
 neural networks
Fingerprint
Dive into the research topics of 'Learning the Physics of AllDielectric Metamaterials with Deep Lorentz Neural Networks'. Together they form a unique fingerprint.Press/Media

Teaching Physics to AI Can Allow It To Make New Discoveries All on Its Own
06/16/22
2 items of Media coverage
Press/Media

Duke incorporates physics into ML algorithms for new insights into materials
05/17/22 → 05/18/22
5 items of Media coverage
Press/Media