← All research

Device modeling · Neuromorphic computing

Bringing measured device behavior into computing

Crossbar simulations that connect measured memristor nonidealities to synaptic updates and neural-network accuracy.

Wang’s Thin Film Group · Purdue UniversityRelated manuscript under major revision

The research question

How do the measured nonidealities of a memristor affect a neural network that uses conductance states as synaptic weights?

An idealized weight update does not capture the asymmetry, variability, and history dependence of a physical device. I built a device-aware crossbar simulation for a dual-mode memristor to connect measured pulse behavior with network-level performance. The work uses differential-pair synapses and write verification to evaluate computation under a model informed by the device.

This is the computing component of the dual-mode memristor research in Wang’s Thin Film Group. It is related to, but distinct from, the Au–CeO₂ compact-model and programming project.

Measured
pulse data
Device
nonidealities
Differential-pair
synapses
Crossbar
simulation

My contribution

I built the device-aware simulation and extracted parameters from raw pulse data. My focus was the connection between experimental behavior and the assumptions used to represent synapses in computation.

  • Pulse-data analysis. I extracted update nonlinearity and the number of steps associated with polarity reversal, alongside cycle-to-cycle and device-to-device variation.
  • Nonideality modeling. I incorporated measured characteristics relevant to write/read noise and retention into the device-aware simulation.
  • Synapse representation. I used write-verified differential pairs to represent synaptic weights and evaluated their network-level performance against a software baseline.

These choices make the device model part of the computational experiment. The resulting accuracy reflects the modeled device behavior and synapse implementation, rather than only the capacity of the neural network.

Generic differential-pair representation of a signed synaptic weight
Conceptual illustration of the differential-pair representation. Two nonnegative conductances encode a signed weight through their difference; device updates affect the resulting effective weight.

Network-level results

The device-aware simulation reached 92.7% MNIST accuracy, compared with a 96.2% software baseline: a difference of 3.5 percentage points.

Bar comparison of 96.2 percent software and 92.7 percent device-aware accuracy
Visual summary of the two accuracy values reported above, on a common 0–100% scale. The difference is an aggregate comparison; it does not isolate the contribution of any individual device nonideality.

These are simulation results informed by measured device data. They describe the modeled crossbar and synapse configuration; they are not measurements from a fabricated inference chip.

The comparison connects materials and device research to a system-level question: how much computational performance remains when the model accounts for the device’s nonideal behavior? It also provides a way to evaluate the implications of programming and synapse choices before committing to a physical array implementation.

Related research

Related manuscript: Z. Lin*, Henan Wang*, et al., “Realizing Dual-Mode Memristor Based on Ni₀.₈Fe₀.₂-BTO/BTO/Ni₀.₈Fe₀.₂-BTO Tri-layer Vertically Aligned Nanocomposites,” Small, under major revision. *Co-first authors.

Research background and results in my CV (PDF)

Explore the related RRAM programming project