Differentiable Silicon
Tesseract Hackathon 2026
Research / Differentiable TCAD · Detailed breakdown
Project overview
One gradient from an ECG classification loss, through a PyTorch spiking network and a JAX device model, into Synopsys Sentaurus TCAD. Best-in-Track winner, Hybrid ML and Mechanistic Models track.
System views
What shipped inside
The architecture, workflow and results that made this project useful.
Sentaurus is a closed binary with no adjoint, so the gradient comes from trust-region finite differences plus a Broyden update
Balanced cross-entropy 1.400 to 1.018 and accuracy 0.250 to 0.688 in 64 solver calls
An open DEVSIM backend can replace Sentaurus without changing the pipeline
Delivery summary