nanoRIP and *NanoAI projects take a joint step forward to integrating physics and artificial intelligence

The nanoRIP physicists join forces with the *NanoAI artificial intelligence scientists to find new ways to get details of a sample without going physically close.

Figure: The problem is to reconstruct the digit-patterns shown in black-and-white in the left column. The white color indicates a refractive index contrast of 1 over the background. The contrast of a water drop for example is 0.33. Light is shown on them from different directions and the reflections from them are captured on different measuring elements located far-away from them. Then a chosen computation solver (a traditional physics-based, a conventional AI-based, or a proposed physics-integrated AI approach) is used to estimate the contrast from the measurements. The proposed solutions reconstruct significantly more accurate details of the sample.

The problem of finding the details of a sample without going physically close to it is very challenging. One of the techniques is to shine light onto the sample and collect the light reflected by a sample from afar. However, converting these far-away measurements to the actual details of the sample is a mathematically challenging inverse problem. Both ‘traditional’ physics-based optimization approaches and modern deep learning based approaches are found to be lacking.

The physicists from the nanoRIP project join forces with the artificial intelligence scientists from the *NanoAI project to find new solutions at the junction of these fields. They have recently published an article in IEEE transcations of computational imaging in which they propose powerful mechanisms to integrate techniques from both the fields. The results indeed show a significant improvements over the conventional approaches, see Figure below.

[1]  Z. Liu, M. Roy, D. K. Prasad and K. Agarwal, "Physics-guided Loss Functions Improve Deep Learning Performance in Inverse Scattering," in IEEE Transactions on Computational Imaging, doi: 10.1109/TCI.2022.3158865. Preprint at Early access link:

Published Apr. 1, 2022 11:01 AM - Last modified Apr. 1, 2022 11:01 AM