Computer Graphics and Puzzles
Computer Graphics techniques have an expanding range of applications in animations, movies, games, virtual reality, etc. today. Whenever you see photo-realistic scenes in cinemas or beautiful animations on your computer screen, there might be several cutting-edge computer graphics algorithms solving hard optimisation problems behind the scene. Somewhat surprisingly, within computer graphics, there is a subfield focusing on puzzles, such maze generation, disentangling two entwined metal pieces (alpha puzzle), or predicting outcomes from Rube Goldberg machines. Although it is seemingly situated on trivial games, the optimisation algorithms developed to automatically figure out the solutions have much wider implications, such as path planning in robotics, automatic terrain generation in games and virtual reality, design of public space, etc.
The PhD project is unique in that it is not designed for a specific application domain but a class of puzzle problems. The research aims to develop new algorithms and theories about new puzzles or new solutions to existing puzzles. The project is designed to let the PhD student choose the type of puzzles to solve. Candidate topics include automatic maze generation, game level completion (e.g. Super mario, or Quake), construction in mine-craft or propose your own!
The supervision team has expertise in computer graphics, computer vision, machine learning, visualization, virtual reality with regularly published research papers at the very top venues in related fields. The research group (vcg.leeds.ac.uk) is also well supported in terms of research equipment including studio-level motion capture systems, eye-tracking devices, deep learning servers and virtual reality kits, to which the candidate will have access.
Supervisors: He and Tom
Searching for Net-Zero Buildings
Net-zero architecture is essential if we wish to live sustainably while growing the UK housing stock. Over the PhD, we will create a tool to interactively design beautiful new developments which do not contribute to our carbon emissions.
Manually designing such developments typically involves alternating phases of simulation and manual (human) design steps in order to minimize energy demands. The simulation stage tells us how efficient our buildings are at using, and generating, power (e.g. insulating against heat loss and the expected production of solar panels); in response, architects and engineers modify their designs to improve the buildings. However, these designs are rarely optimal and it has become increasingly difficult to rely on manually driven design optimization as carbon emission targets become tighter. To design the 300,000 new homes that the UK plans to develop each year, smarter tools are needed. In this project we will explore the potential of procedural architectural modeling and machine-learning accelerated simulation in the interactive design of low-carbon developments.
Simulation lets us compute operational carbon emissions and thermal comfort performance of a building from its geometric representation. This allows our designers to explore and optimise the design space. Techniques such as hourly annual simulation are able to assess the complex thermal heat balances and so energy performance of buildings in detail, but are time consuming and expensive to compute. One aspect of this project will be the generation of fast simulation proxies using state of the art machine-learning techniques.
Procedural models express a multitude of architectural forms by converting a set of parameters to a geometric representation. However, as the number of parameters grows, it becomes increasingly difficult for designers to understand and explore this parameter space. This makes discovering good designs time consuming…and finding optimal designs impossible. The second aspect of this project will be using fast simulations to guide the optimisation of these parameters.
Finally, the development of an interactive system will allow the above two systems to guide architects in the rapid development of designs for large-scale housing developments. This will allow users to quickly estimate and guide procedural optimisation to ensure beautiful designs with minimal environmental impact.
This project is a collaboration between the Schools of Computing (Tom studies machine-learning and urban procedural modeling) and Civil Engineering (Simon Rees studies building energy modelling and net-zero design). Strong applicants will have training in software engineering, machine-learning, or simulation techniques.
Growth Paradigms for Procedural Modeling
There is a history of growth-based procedural modeling for a wide range of geometry domain – from trees to urban environments. Growth based modeling has the advantage of emergent behaviour which is easily understood, and geometrically smooth. Such smoothness is important for domains such as model fitting and rule creation user interfaces, as explored in previous work. Goals for this project may include creating a unified growth language useful for multiple domains and creating user interfaces for growth-based modeling.
Databases for Massive-Scale Procedural Modeling
Procedural modeling systems are typically constrained by a single computer’s memory. To alleviate this constraint, we may build procedural systems which operate on databases. Given different constraints on geometry evaluation speed, bandwidth, and storage, we may wish to evaluate procedural models at different times, at different scales.