Research by S M Hozaifa Hossain: devices that compute like brains.

Research

Ferroelectric FETs

TCAD & device physics

Spiking neural networks

Dhaka, Bangladesh

S M Hozaifa Hossain researches semiconductor devices and neuromorphic hardware: how ferroelectric transistors can act as low-energy synapses, and how TCAD, high-level synthesis and machine learning connect devices to algorithms. See also awards and publications.

Research · Thesis · CREST 2027

Undergraduate thesis · defended September 2026

What did my undergraduate thesis show?

My thesis models a gate-all-around ferroelectric FET with a 7 nm HZO gate stack in Synopsys Sentaurus TCAD, then uses it as an artificial synapse in a spiking network. The device shows a 5410x memory window and a 63.5 mV/dec subthreshold slope, and the network classifies ECG heartbeats with 83.04% accuracy.

Title: Ferroelectric Gate-All-Around FETs for Next-Generation Neuromorphic Systems: Device Modeling and Spiking Network Implementation. Advisor: Dr. Mainul Hossain, University of Dhaka. The sub-100 nm device was calibrated with Preisach polarization dynamics against experimental data. Conductance-level placement mattered more to SNN accuracy than the number of levels. The calibrated conductance states were deployed in a PyTorch spiking recurrent network (SpikingJelly, leaky integrate-and-fire neurons) for 4-class arrhythmia classification on the MIT-BIH database.

Transfer curves of the erased and programmed GAA FeFET showing a 5,410x ratio at zero gate voltage and a retained threshold shift of 0.405 V
Erased and programmed transfer curves: 5,410x at VGS = 0 V, retained ΔVT 0.405 V.
Four panels showing the HZO dipole state and channel conductance at the erased state and after 3, 9 and 15 programming pulses
Channel conductance rises from the erased state to pulse 15 as HZO dipoles flip.

Key numbers

Memory window 5410x at VGS = 0 VRetained ΔVT 0.405 VSS 63.5 mV/dec<1% retention loss over 9.9e5 read cycles15 analog conductance states over 3.79 decades~9 aJ per switching pulse83.04% accuracy, macro-F1 0.75535.8 pJ per heartbeat

CREST 2027 · Silicon River Bangladesh

What is the CREST 2027 research project?

As a CREST 2027 Research Fellow at Silicon River Bangladesh, I will study how the curvature of the channel limits the number of reliable analog states in gate-all-around FeFET synapses. The study combines Sentaurus TCAD simulation with machine learning. A synapse that holds more distinguishable conductance levels stores more weights per device, so the question is where the geometry sets a ceiling.

The fellowship runs January to December 2027 at the Center for Research Excellence on Semiconductor Technology, supported by Neural Semiconductor Ltd. It builds on the thesis, where 15 analog conductance states were extracted for a GAA device and a planar reference supported 10. My advisor is Dr. Mainul Hossain.

Fellowship

Selected 30 October 2026Fellowship: Jan to Dec 2027Advisor: Dr. Mainul HossainHost: Silicon River Bangladesh

IEEE EDS TCAD Hackathon 2026

How does stress improve a sub-10 nm gate-all-around nanowire transistor?

Applying 1 GPa of uniaxial tensile stress along a 6 nm silicon nanowire repopulates its conduction-band subbands. In my Sentaurus TCAD design this gave 35% more on-current and 32% more transconductance, while band-to-band tunneling and quantum corrections kept off-state leakage at 0.99 pA.

The nMOS has a 15 nm gate length, 6 nm diameter and 0.9 nm EOT (0.4 nm SiO2 interface layer plus 2.82 nm HfO2, κ = 22), built as a six-material 3D cylinder in Sentaurus Structure Editor with 0.2 nm radial mesh refinement. Simulation used deformation-potential and eSubBand models, density-gradient quantum correction and nonlocal-path band-to-band tunneling. This design made me one of three international winners of the IEEE EDS TCAD Hackathon 2026, and the first winner from Bangladesh.

Cross-section of the gate-all-around nanowire showing a 6 nm silicon core, 15 nm gate length and the oxide, HfO2, TiN and nitride layers
Device cross-section: 15 nm gate length, 6 nm silicon diameter.
Drain current versus gate voltage on log and linear scales comparing 1 GPa tensile stress with the unstressed device, showing a 35 percent gain in on-current
Drain current with and without 1 GPa tensile stress: +35% ION.

Key numbers

+35% ION (5.42 µA, 287 µA/µm)+32% gm (2.10 mS/µm)IOFF 0.99 pA, ION/IOFF 5.5e6SS 61 mV/dec at VDD 0.7 V

Tesseract Hackathon 2026 · Best-in-Track

What is Differentiable Silicon?

Differentiable Silicon is a pipeline that passes one gradient from an ECG classification loss through a PyTorch spiking network and a JAX device-to-neuron model into Synopsys Sentaurus TCAD. That tells you how four FeFET fabrication parameters should change to improve accuracy, even though the TCAD solver is a closed binary.

The closed solver has no adjoint, so the missing gradient was manufactured with trust-region finite-difference probes and a Broyden update. Descent cut balanced cross-entropy from 1.400 to 1.018 and raised accuracy from 0.250 to 0.688 in 64 solver calls. It won Best-in-Track in the Hybrid ML x Mechanistic Models track, run by Pasteur Labs and ISI.

Code: github.com/hozaifa1/differentiable-silicon

Source: Pasteur Labs winners blog

Gradient path from ECG loss through a PyTorch spiking network and a JAX circuit model into Sentaurus, ending at the four fabrication parameters
Gradient path from the ECG loss back to four fabrication parameters. The solver has no adjoint, so the last step is estimated.
Plots of memory window against subthreshold slope over the optimization and where each of the four fabrication parameters moved
How the memory window and slope traded off, and where the four parameters moved.

Key numbers

4 FeFET fabrication parametersCross-entropy 1.400 to 1.018Accuracy 0.250 to 0.68864 solver calls

Siemens EDA HLS Hackathon 2026 · 1st place

How was a CNN accelerator made 915 times more energy efficient?

I designed an ultra-low-power CNN accelerator for roadside-object recognition by exploring many high-level synthesis configurations, then measuring real gate-level power. Energy per inference fell from 6,877 nJ to 7.51 nJ, a 915x reduction, at 95.3% accuracy. The result took first place in the Siemens EDA hackathon.

The flow used QKeras and hls4ml to move the quantized network into Siemens Catapult HLS, targeted a 45 nm ASIC library, and measured power with PowerPro at gate level. The final design came from 365 synthesis runs.

Source: HLS Academy results page

Layer-by-layer comparison of the reference CNN and the final design, which drops one pooling and dense stage and uses 6-bit activations
Reference network (top) against the final design (bottom).

Key numbers

6,877 nJ to 7.51 nJ per inference915x lower energy95.3% accuracy365 synthesis runs45 nm ASIC library

Let's build something worth building.