Molecular Property Prediction with Photonic Chip‐Based Machine Learning

Abstract

Machine learning methods have revolutionized the discovery process of new molecules and materials. However, the intensive training process of neural networks for molecules with ever‐increasing complexity has resulted in exponential growth in computation cost, leading to long simulation time and high energy consumption. Photonic chip technology offers an alternative platform for implementing neural networks with faster data processing and lower energy usage compared to digital computers. Photonics technology is naturally capable of implementing complex‐valued neural networks at no additional hardware cost. Here, the capability of photonic neural networks for predicting the quantum mechanical properties of molecules is demonstrated. To the best of knowledge, this work is the first to harness photonic technology for machine learning applications in computational chemistry and molecular sciences, such as drug discovery and materials design. It is further shown that multiple properties can be learned simultaneously in a photonic chip via a multi‐task regression learning algorithm, which is also the first of its kind as well, as most previous works focus on implementing a network in the classification task.

Three-Dimensional Modelling of Near-field Beam Profiles from Grating Couplers using Deep Neural Network

“…Meanwhile, Dwivedi et al reported the use of machine learning models on different datasets vis-a-vis fiber-optic plasmonic sensor devices. The modeled data is found to be highly consistent with the data in terms of trend matching, and the values of other parameters such as R^2 and mean absolute error [29]. On the other hand, in our previous studies, we used deep neural network (DNN) techniques to predict the beam properties of light coupled from grating structures to free space.…”

Inverse design of electromagnetic metamaterials: from iterative to deep learning-based methods

“…Although ANNs inherently introduce systematic errors into optimization processes, these errors, typically within a few percent, do not negate the advantages of DL-based methods characterizing by strong fitting capabilities and high evaluation speeds. Driven by currently emerging technologies such as photonic chips [159][160][161][162] and quantum computing [163][164][165][166], these advantages can be even further expanded. Overall, as DL theories advance and computer performance improves, it is reasonable to anticipate that DLbased inverse design methods will play a vital role in realizing disruptive EM metamaterial designs in the future.…”

Physics‐Aware Analytic‐Gradient Training of Photonic Neural Networks

“…On-chip (online) PNN training can improve prediction accuracy, and it is classified into two main categories: gradient-free and numerical-gradientbased. Gradient-free methods employ non-gradient search algorithms, such as genetic algorithms, [13,14] biologically inspired method [15] and particle swarm optimization [16] to obtain optimal solutions. While these approaches yield high accuracy in simple classification tasks, their computation complexity increases significantly with the number of trainable parameters, restricting their scalability for complicated machine-learning tasks.…”