Multilayer optical thin film design with deep Q learning

Abstract

Multilayer optical film plays a significant role in broad fields of optical application. Due to the nonlinear relationship between the dispersion characteristics of optical materials and the actual performance parameters of optical thin films, it is challenging to optimize optical thin film structure with the traditional models. In this paper, we present an implementation of Deep Q-learning, which suited for the most part for optical thin film. As a set of concrete demonstrations, we optimize solar absorber. The optimal program could optimal this solar absorber in 500 epoch (about 200 steps per-epoch) without any human intervention. Search results perform better than researchers’ manual searches.

Parameterized reinforcement learning for optical system optimization

“…if we assume discrete layer thicknesses from 0 to 150 nm in steps of 0.1 nm resulting in a total number of |T| = 1500 thickness values [29]. We compare our experimental results to multi-layer thin films designed by human experts and another Q-learning algorithm [29], henceforth referred to as DQN algorithm.…”

“…if we assume discrete layer thicknesses from 0 to 150 nm in steps of 0.1 nm resulting in a total number of |T| = 1500 thickness values [29]. We compare our experimental results to multi-layer thin films designed by human experts and another Q-learning algorithm [29], henceforth referred to as DQN algorithm. However, to enhance comparability between our approach and the DQN algorithm, we enabled the latter to not only optimize over layer thicknesses but also over layer materials.…”

“…Although the existence of a global optimum is mathematically and computationally evinced [14,59,60], algorithms that guarantee finding the global optimum in an exhaustive search tend to be computationally intractable, even if only layer thicknesses of less than four layers in total are considered [16]. Thus, in accordance with some theoretical and analytical investigations [1,15,28,68,71], including genetic and evolutionary approaches, multi-layer thin films are also optimized based on heuristic approaches [10,21,29,41,44,67]. Alike many of the mentioned methods, deep learning-assisted techniques are reported to optimize layer thicknesses only: Roberts et al [47] proposed a variational autoencoder, Liu et al [36] combined forward modeling and inverse design in a tandem of DNNs, and Hegde [23] blended deep learning with evolutionary elements.…”

Investigation of Non-Linear MHD Jeffery–Hamel Blood Flow Model Using a Hybrid Metaheuristic Approach

“…PSO is applied to birds' flocking and fish training in cooperative swarm routines [25]- [27]. Particle swarm optimization has proven to be a highly effective optimization process and has many applications, such as the optical properties of multilayer thin films [28]- [31], auto-regressive moving average model [32]- [34], parameter estimation of chaotic maps [35], [36], clustering high-dimensional data [37], [38], nonlinear benchmark system optimization [39], in electromagnetic plane waves the parameter estimation problem [40], multi-objective problem of core loading pattern optimization in nuclear reactors [41] and optimal reactive power dispatch [42].…”

A Reinforcement Learning Method for Optical Thin-Film Design

“…In contrast, deep learning (DL) based algorithms, which are considered as being able to "learn" Maxwell's equations, were proposed to solve the nonlinear relationships between the structural parameters and film's performance by a large dataset [15][16][17] [18]. Learning the design process by human experts, and reinforcement learning are other solutions to solve this problem, which train an agent to learn about the parameter space of a series by exploration [19][20] [21]. While the previously described algorithms have worked well for structural parameters optimization (nano-structure size and layer thickness) for a in thin-film inverse design well, there is little research on the design of component materials for these processes.…”