ReflectorML

Project page: https://alexin.cclaude.rocks/projects/reflectorml/

Introduction

Free-form reflector design is essential in optics for precisely shaping light distributions, with applications in automotive lighting, energy-efficient LED optics, laser-based manufacturing, aerospace systems, and medical imaging.

This problem is mathematically formulated as a non-linear Monge-Ampère equation (Wikipedia), which defines the mapping between a given light source and a prescribed target intensity.

Far-field (parallel-beam) Monge–Ampère equation

$\det\big(D^2\varphi(x)\big)=\dfrac{f(x)}{g(\nabla\varphi(x))},;x\in\Omega$

• $\varphi$ : convex potential $\Omega \rightarrow \mathbb{R}$
• $D^2\varphi$ : Hessian of $\varphi$
• $f$ : source density on $\Omega$
• $g$ : target density at $y=\nabla\varphi(x)$

However, traditional numerical solvers for this equation are computationally expensive and often struggle with convergence, particularly in complex boundary conditions. Developing efficient and robust methods to solve this problem is crucial for advancing high-performance optical designs in both scientific and industrial applications.

By embedding the governing equations - such as the Monge-Ampère equation - directly into the learning process, our approach ensures physically consistent solutions while significantly reducing computational costs. Unlike purely data-driven models, PINNs do not rely solely on labeled data but instead enforce optical constraints during training, improving solution accuracy for specific problem instances. This framework accelerates the inverse design process and provides a computationally efficient alternative to traditional numerical solvers.

Description

This project implements a hybrid method that aims to use both:

• PyTorch raytracer (with automatic differentiation) with a transport loss
• Physical loss based on the Monge-Ampere equation.

Implementations

This repository contains six parallel implementations of the same reflector problem, sharing one pipeline in reflectorml/core/ and differing only in network parameterization:

Directory	Approach	Network	Network output	Monge-Ampère loss
reflectorml/impls/siren/	SIREN scalar potential	MirrorSurface (SIREN)	Scalar potential phi(x)	Hessian of phi
reflectorml/impls/icnn/	Input Convex Neural Network potential	MirrorSurface (ICNN)	Scalar potential phi(x)	Hessian of phi
reflectorml/impls/spd_field/	Gradient-field	MirrorSurface (SIREN)	SPD Jacobian J(x) = D^2 phi	Jacobian of the map + curl-free penalty
reflectorml/impls/hycnn/	Hyper Input Convex NN potential	MirrorSurface (HyCNN)	Scalar potential phi(x)	Hessian of phi
reflectorml/impls/hinge/	Smoothed-hinge potential (frozen ridges + linear readout)	MirrorSurface (hinges)	Scalar potential phi(x)	Hessian of phi
reflectorml/impls/quad_zernike/	Zernike spectral potential	MirrorSurface (Zernike)	Scalar potential phi(x)	Hessian of phi

Run each implementation from the repository root (library shapes and tests are shared):

make
python -m reflectorml.core.main --impl siren        # potential-based
python -m reflectorml.core.main --impl icnn         # ICNN potential
python -m reflectorml.core.main --impl spd_field    # gradient-field
python -m reflectorml.core.main --impl hycnn        # HyCNN potential
python -m reflectorml.core.main --impl hinge        # hinge potential
python -m reflectorml.core.main --impl quad_zernike  # Zernike potential

Getting Started

Prerequisites

Ensure you have the following installed:

• Python 3.7 or higher
• Required libraries (listed in requirements.txt)

Installation

Clone the repository:

git clone  https://github.com/Alexin-CH/ReflectorML.git
cd ReflectorML

Install the required dependencies:

make

---

Results

Each implementation documents its results in its own README:

• siren/ — SIREN scalar potential
• icnn/ — ICNN scalar potential
• spd_field/ — gradient-field (SPD Jacobian)

---

Acknowledgments

This project is inspired by several papers:

• "A Neural Network Approach for Solving the Monge-Ampère Equation with Transport Boundary Condition"
  You can read the paper here.
• "Input Convex Neural Networks"
  You can read the paper here
• "Convex Physics Informed Neural Networks for the Monge-Ampère Optimal Transport Problem"
  You can read the paper here