FET/SHF: Small: Reinforcement learning and transformer inspired smart photonics inverse design

光学和光子学领域通过光通信、太阳能、太空望远镜和高分辨率光刻等现代技术改变了我们的生活。长期以来，人类专家只能依靠自己积累的经验和物理直觉手动进行光学设计。对于光子逆设计尤其如此，因为必须找到具有所需光学功能的适当光子结构。近年来，科学界采用深度学习方法来自动设计满足设计目标的复杂结构，这一方向取得了实质性进展。该研究项目将使光子学非专业人士能够使用开发的人工智能（AI）模型来获得他们个人光学设计问题的解决方案。此外，了解潜在的操作原理将推进更广泛的光子学知识，并使研究人员能够更快地开发新结构。为了实现他们的目标，研究团队将探索和应用两种强大的人工智能技术。要研究的第一个工具是深度强化学习（deep reinforcement learning），这是一种顺序生成过程，通过试与奖励来学习设计结构，以模仿人类和动物学习与世界互动的方式。追求机器与人的合作学习，人教机器学习，机器启发人理解。这种组合输入将有利于实现各种类型的光学结构。第二种方法是利用变压器方法，这是非常成功的强大大型语言模型背后的动力，用于智能光学设计。研究团队将利用基础模型，大型机器学习模型，处理各种下游任务，一旦训练不同的，大规模的数据，解决大规模和复杂纳米结构的光学逆设计。通过研究获得的知识将进一步应用于几个试验台进行实验演示。随着越来越多的人使用这项技术，越来越多的数据可用于训练神经网络，人们可以预测它的学习和生成能力将会提高，用户也会更有效。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The field of optics and photonics has transformed our lives with such modern technologies as optical communications, solar energy sources, space telescopes and high-resolution photolithography. For a long time, human experts had to perform optical design manually relying on their accumulated experience and physical intuitions. This is especially true for photonic inverse design, where one must find the appropriate photonic structures with the desired optical functions. Recently, deep learning approaches have been pursued by scientific community to automatically design sophisticated structures that can satisfy the design objective, leading to substantial progress in this direction. This research project will enable photonics non-experts to use the developed Artificial Intelligence (AI) model to obtain solutions to their individual optical design problems. Furthermore, understanding the underlying operational principles will advance more generalizable photonics knowledge and enable researchers to develop new structures faster.To accomplish their goals, the research team will explore and apply two powerful AI technologies. The first tool to be studied is deep reinforcement learning, a sequential generation process that learns to design structures with trial-and-reward, in a way to mimic how human and animals learn to interact with the world. Cooperative learning between machine and human will be pursued, where human teaches machine to learn, and machine inspires human to understand. This combined input will benefit the realization of various types of optical structures. The second method utilizes the transformer method, the powerhouse behind the highly successful powerful large language models, for smart optical design. The research team will leverage the Foundation model, the large machine learning models that tackle various downstream tasks once trained on diverse, and large-scale data, to address the optical inverse design of large-scale and complicated nanostructures. The knowledge gained through the study will be applied further to a few testbeds for experimental demonstrations. With more people using the technology and increased data available for training the neural network, one can anticipate its learning and generative capabilities will advance and will the users even more effectively.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.