Accelerated discovery with differentiable programming: From nanomaterials to chiplets design
Tuesday, May 14, 2024
11 a.m. — 11:45 a.m. ET
Giuseppe Romano
Research Scientist, Institute for Soldier Nanotechnologies
Automatic Differentiation (AD) software initially developed for machine learning may benefit applications beyond neural networks. A prominent example is given by physics solvers, where language-wide differentiability enables large-scale inverse design. This talk will report on recent materials and systems optimization efforts, along with the implemented open-source software. The first application is thermal energy harvesting, for which the research team identifies an optimal nanostructured Si membrane. This system, which has a minimum feature of about 100 nm, has been fabricated with the Focused Ion Beam (FIB) at MIT.nano. The second application focuses on the inverse design of spatio-termporal-modulated materials, an emerging class of materials that has the potential to unlock unprecedented performances, such as time interfaces. The latest application pertains to chiplets floorplan design, where, in collaboration with the MIT-IBM Watson AI Lab, the researchers design a framework for minimizing the maximum temperature during operation. This thermally-aware approach, combined with the maximum wire length constraint, enables a faster design compared to common gradient-based approaches. The last part of the talk will briefly describe MatInverse, the in-house JAX-based software underlying these research directions.
Attendees can join and participate in the series via Zoom.