Opening design using bayesian optimization

Accurate light-transport simulation has been aided by significant improvements in the convergence speed inrecent years. This is largely attributed to the more advanced statistical light-transport solvers, denoisingalgorithms, and inclusion of hardware-accelerated ray tracing in commodity GPUs. This has unlocked asubstantial potential for its utilization in demanding applications of light simulation (such as architecturallighting) in terms of accuracy and consistency. In this work, we leverage photorealistic simulation via interactive path tracing to propose an automatic solution to the opening design problem that maps effectively tothe luminance distribution characteristics of natural light scattering.

Opening design is an important subspace in architectural building design. An opening is a geometricformation that connects the interior of a building to its exterior. In our context, the problem of designing suchopenings is approached as an inverse geometry problem, i.e., a goal-driven definition of the design parameters(commonly referred to as the inverse opening problem).

This workfocus on the design of openings as natural light sources and their potential tomaximize natural light usage and minimize the requirements for artificial light sources during the daytime. We describe and develop a method as well as the respective working system that combines robustphotorealistic rendering, a comprehensive set of tools for describing the designer′s illumination requirementsand aesthetic preferences, and an optimization framework adapted to the opening design problem.

The contributions of this work are as follows: a) a problem formulation based on Bayesian optimization,which acknowledges the designer′s workflow and allows for a convenient expression of each part of theirdesign process; b) accurate physically-based lighting evaluation during optimization by utilizing maximumluminance environment maps for glare prevention; and c) support for arbitrary opening designs using daylighting systems of any shape, size, and orientation.