Simulation-Driven Graphics and Fabrication

点击翻译按钮获取中文摘要

Algorithms for computational physics serve as a computer's imagination, allowing the machine to ``visualize'' the outcome of complex, real-world interactions. These algorithms are used to drive the games we play, the movies we watch, to design cars, ships and planes and in countless other areas from architecture to medicine. But, while a person can imagine an object interacting with the world almost instantaneously, accurate physical simulations operate on time scales of hours, days or weeks. Computational Fabrication and Computer Graphics amplify the need for as-fast-as-a-person physics simulation. Additive manufacturing (identified by the World Economic Forum as a key emerging technology) allows for the ultra high-resolution control of geometry and material properties of fabricated objects. Predicting the performance of such designs in the real world often requires the solution of computational problems thousands of times larger than what was previously required in engineering disciplines. The over-arching goal of Professor Levin's research is to manifest this computational imagination by developing new machine learning techniques that understand physics and mechanics. His work tackles major difficulties in adapting standard machine learning techniques to large-scale problems in physics simulation. Levin focuses on creating algorithms that are robust to real-world input data, from LIDAR data generated by self-driving cars, to computer-aided designs created during the engineering design process. By leveraging these increasing sources of geometric information, Levin utilizes learning and data-driven approaches to create fast algorithms for physics simulation that can be orders-of-magnitude faster than standard approaches, but retain predictive accuracy. These methods enable scalable generative design across a myriad of engineering and computer graphics domains -- from aeronautical design to visual effects. Finally, Levin's work focuses on methods that are generalizable across large families of geometries and phenomenologies which can alleviate the need to train neural models for new examples -- conserving computational power and energy. Professor Levin's research enables a game changing shift in simulation performance and accuracy along with more environmentally friendly machine learning approaches than in other disciplines.