Flexcompute, Inc.
Dr. Zongfu Yu, Chairman and Co-Founder, Flexcompute, Inc.
Dr. Zongfu Yu is Chairman and Co-Founder of Flexcompute. Before co-founding the company, he built his career in photonics and applied physics. He earned an M.S. in Management Science and Engineering and a Ph.D. in Applied Physics, both from Stanford University, where he worked under Shanhui Fan. He joined the University of Wisconsin, Madison's Department of Electrical and Computer Engineering in 2013, where he held the Grainger Professorship, an endowed chair reserved for senior faculty, before leaving academia to lead Flexcompute full-time.
His academic research centered on computational photonics and nonreciprocal optics, along with machine learning applied to sensing and imaging, work aimed at extracting more information and performance from physical systems through better computation rather than new hardware alone. That grounding carries into Flexcompute's current work building physics twins, living, physics-accurate models that engineers can query directly, applied to problems like warpage and thermomechanical behavior in semiconductor packaging, where small deviations during assembly can compound into yield and reliability failures. He is a Fellow of Optica, a 2026 IEEE Fellow cited for contributions to computational photonics, and a recipient of an NSF CAREER Award.
Topic:
Tokenized Geometry: Bridging the Gap Between Speed and Precision
Abstract:
Warpage prediction in advanced packaging forces a trade between fidelity and schedule. A full-package model that resolves solder joints discretely exceeds the memory available to conventional implicit FEA, so analysts homogenize: smeared bump arrays, equivalent-property layers, a reduced set of representative thermal excursions. Those approximations sit directly on the quantity of interest, since local joint stiffness and CTE mismatch drive the out-of-plane displacement field they are meant to predict.
This work presents a solver based on a tokenized geometry representation that removes the memory ceiling on package-scale structural analysis. The method enforces conservation explicitly and is not a surrogate regressed on prior simulation output. We apply it to a co-packaged-optics switch package with no geometric simplification, resolving every solder joint, the ASIC, the optical engines, the interposer, and the substrate in a single system of equations at billion-DOF scale. The full assembly process flow, spanning dozens of increments and hundreds of nonlinear solves, completes in under an hour.
We benchmark against commercial implicit FEA on the largest model both codes can run, and report the speedup and the agreement in out-of-plane displacement on identical geometry, material assignment, and thermal load path. We discuss what solve time at this scale changes about where warpage sits in a packaging program, and close with early results on automated process-window exploration.
Key Technologies Covered:
- Tokenized geometry representation for large-scale structural solve
- Neural physics solvers, first-principles rather than trained on simulation data
- Full-package thermomechanical warpage, nonlinear, no geometric homogenization
- Manufacturing traveler simulation across 80 process increments
- Billion-DOF FEA and memory-wall mitigation
- GPU-accelerated multiphysics
- Co-packaged optics and 2.5D/3D heterogeneous integration
- Solder joint reliability at full array resolution
- Agentic design automation with physics in the loop