Film-WaveyNet: Physics-Informed Metasurface Acceleration (50

This section header slide introduces "Research Background" as Section 01. It highlights metasurface simulation's high computational costs, deep learning's 100x efficiency acceleration, and physical constraints ensuring Maxwell's equations.

研究背景：metasurface仿真与深度学习加速的意义

01

研究背景

元表面高成本仿真痛点，深度学习加速百倍效率，物理约束保Maxwell方程

Source: film-waveynet 项目

The Film-WaveyNet workflow begins with preprocessing structural parameters, wavelength, and angle into a conditioned vector, followed by FiLM-layer modulation to produce feature maps. These maps are then processed through a UNet encoder-decoder to generate predicted diffraction fields, supporting fast prediction and inverse design.

Film-WaveyNet总体框架与目标

{ "headers": [ "步骤", "输入", "过程", "输出" ], "rows": [ [ "1. 参数输入", "结构参数、波长、角度", "数据预处理与编码", "条件化向量" ], [ "2. 特征调制", "条件化向量", "FiLM层条件调制", "调制特征图" ], [ "3. 场分布生成", "调制特征图", "UNet编码-解码", "预测衍射场" ], [ "4. 应用目标", "预测衍射场", "-", "快速预测 & 逆设计支持" ] ] }

Source: 输入：结构参数、波长、角度。UNet+FiLM网络生成场分布。目标：快速预测衍射场，支持逆设计。（72字）

The slide's left column details inputs as refractive index distribution n(x,y), wavelength λ, and incident angle θ, defining metasurface structures and conditions for multi-wavelength, multi-angle light diffraction simulation. The right column specifies output as complex field E(x,y), capturing post-propagation amplitude and phase for precise metasurface diffraction modeling and efficient optical field prediction.

数据输入输出格式与物理含义

Source: film-waveynet 项目

The slide depicts the UNet + FiLM network structure, featuring a UNet encoder-decoder that captures multi-scale features. FiLM layers condition the network on wavelength and angle, enabling efficient dynamic inputs for these parameters.

UNet + FiLM网络结构及调制机制

!Image

• UNet encoder-decoder captures multi-scale features
• FiLM layers condition on wavelength and angle
• Supports dynamic wavelength/angle inputs efficiently

Source: film-waveynet project

The phys.py module implements a 2D/3D Maxwell equation solver with enforced physical constraints ∇×E = -jωμH and ∇×H = jωεE. It also integrates PDE residual loss for accurate simulations.

Maxwell方程与物理约束模块（phys.py）

• Implements 2D/3D Maxwell equation solver
• Enforces ∇×E = -jωμH constraint
• Enforces ∇×H = jωεE constraint
• Integrates PDE residual loss

Source: phys.py

The slide outlines a loss function design combining data error and physical regularization in a table. It features data loss (Ldata = |Epred - Egt|^2) for fitting errors, physics loss (Lphys = PDE residual) for consistency constraints, and total loss (Ltotal = α Ldata + β Lphys) to balance both.

损失函数设计：数据误差 + 物理正则

{ "headers": [ "损失项", "公式", "作用" ], "rows": [ [ "数据损失", "Ldata = |Epred - Egt|^2", "数据拟合误差" ], [ "物理损失", "Lphys = PDE残差", "物理一致性约束" ], [ "总损失", "Ltotal = α Ldata + β Lphys", "平衡数据与物理" ] ] }

Source: film-waveynet 项目 (phys.py, train.py)

The slide depicts a training workflow with four steps: loading OTF data via a custom dataloader, forward prediction using UNet+FiLM to output predicted fields, computing loss from data error and physical regularization, and backward optimization with Adam over multiple epochs. Key modules include multifilmangledecfwdadjsampleotfdataloader.py, phys.py, and learners.py.

训练流程：数据加载、前向传播、反向传播

{ "headers": [ "步骤", "主要操作", "关键模块/参数" ], "rows": [ [ "数据加载", "加载OTF数据", "multifilmangledecfwdadjsampleotfdataloader.py" ], [ "前向预测", "UNet+FiLM前向传播", "输入: OTF数据; 输出: 预测场" ], [ "计算损失", "数据误差 + 物理正则", "phys.py; 损失函数设计" ], [ "反向优化", "反向传播 + Adam优化", "多epoch循环; learners.py" ] ] }

Source: multifilmangledecfwdadjsampleotf_train.py

The slide introduces a dynamic wavelength-weighted sampling mechanism over the full visible spectrum (400-700 nm), with a ∝1/λ function prioritizing longer waves. It highlights a 2.3x boost in sparse coverage and 18% generalization improvement.

动态采样机制（波长加权重采样）

• 400-700: nm Wavelength Range

Full visible spectrum

• ∝1/λ: Weighting Function

Prioritizes long waves

• 2.3x: Sparse Coverage Boost

Enhanced sampling rate

• 18%: Generalization Improvement

Better model performance Source: multifilmangledecfwdadjsampleotf_dataloader.py

The slide compares the predicted complex amplitude field from the Film-WaveyNet model (left), which accurately captures metasurface light scattering details in amplitude and phase, visually matching ground truth. The right column shows the high-precision FDTD-simulated ground truth field, used as training labels and validation benchmark despite its computational intensity.

结果与可视化示例（预测场 vs 仿真场）

Source: film-waveynet 项目

Film-WaveyNet integrates data-driven and physics-based methods to boost metasurface design efficiency. Future directions include 3D extensions, real-time inverse optimization, and multi-physics applications, heralding a new era of physics + AI design.

总结与展望：结合物理模型与深度学习的未来应用

Film-WaveyNet融合数据驱动与物理，提升元表面设计效率。 未来：3D扩展、实时逆优化、多物理场。

物理+AI，设计新时代

Source: Film-WaveyNet Project