⚡ electrical engineering

Physics-Based Compact Modeling and Design–Technology Co-Optimization of MoS2-Based Optoelectronic Synapses for Neuromorphic Vision

This paper presents a physics-based compact model for MoS2-based optoelectronic synapses that accurately captures trap dynamics to reveal emergent behaviors and enable design–technology co-optimization, achieving significant energy reductions and retention-aware performance improvements in neuromorphic vision systems.

Original authors: Muhammet Arucu

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Muhammet Arucu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain as a bustling city where thoughts are like delivery trucks zipping between neighborhoods. In a traditional computer, the "sensors" (like a camera) are the delivery trucks that pick up packages (images), but they have to drop them off at a central warehouse (the processor) to be sorted and recognized. This round trip takes time and burns a lot of fuel (energy). Scientists are now trying to build "smart sensors" that can do the sorting right where the image is taken, mimicking how our own brains work. This field is called neuromorphic computing. To make this happen, researchers are looking for tiny electronic switches called "synapses" that can remember light patterns, just like our brain cells remember experiences. One promising material for these switches is a super-thin, two-dimensional sheet called Molybdenum Disulfide (MoS2). Think of MoS2 as a microscopic trampoline; when light hits it, it traps tiny electrical charges in its surface, changing how easily electricity flows through it. This change acts like a memory. However, figuring out exactly how these tiny traps work and how to design them efficiently has been like trying to guess the rules of a game just by watching the players, without ever seeing the rulebook.

This paper introduces a new "rulebook" for these MoS2 light-sensing synapses. Instead of guessing or using messy trial-and-error methods, the author, Muhammet Arucu, built a precise physics-based model that acts like a digital twin of the real device. Imagine the synapse as a multi-story parking garage for electrical charges. When light hits the device, it's like cars (charges) driving in. The model treats the garage as having several levels (trap states), where some cars park in the top, easy-to-reach spots (short-term memory) and others park in the deep, hard-to-reach basement (long-term memory). The paper shows that this "parking garage" model can perfectly predict how the device behaves when hit with different patterns of light pulses, matching real-world data from two different types of devices with incredible accuracy.

The study makes a few exciting discoveries while debunking some old assumptions. First, it proves that the "memory" of these devices isn't just a simple blur; it's a complex dance of charges filling up different levels of the parking garage. The model revealed that the speed at which these devices seem to "forget" isn't just about the traps themselves, but is actually limited by how long the light pulse lasts. If you shine a light for too long, you can't see the fast-moving charges; they get hidden, like trying to spot a hummingbird with a slow camera shutter. Second, the paper found a hard limit on how much "facilitation" (getting more excited by a second light pulse) these devices can show. If the device is behaving linearly, the response can never more than double (200%). If researchers see numbers higher than that, it means the device's internal amplifier is working extra hard, a detail that was previously missed.

Most importantly, the paper uses this model to find a "magic knob" for engineers: the brightness of the light used to program the memory. By turning down the light intensity, the team showed they could train a digital brain (a 640-synapse array) to recognize handwritten digits just as well as before, but using 6.7 times less energy. It's like realizing you can drive a car just as fast by shifting gears earlier rather than slamming the gas pedal. The model also allowed them to simulate how long the memory lasts before fading, showing that with a tiny bit of "refreshing" (a quick flash of light), the system could keep working for hours while using almost no power—less than a microwatt. This work doesn't just describe the device; it provides a clear, physics-based map for designing the next generation of ultra-efficient, brain-like vision sensors, moving us from guessing to precise engineering.

Technical Summary: Physics-Based Compact Modeling and DTCO of MoS2-Based Optoelectronic Synapses

Problem Statement
Current neuromorphic vision systems based on in-sensor computing rely on charge-trapping optoelectronic synapses, particularly those utilizing two-dimensional semiconductors like monolayer MoS2. While experimental devices have demonstrated synaptic behaviors such as excitatory post-synaptic currents (EPSC), paired-pulse facilitation (PPF), and long-term retention, the field lacks a predictive, physics-based modeling framework. Existing methodologies rely heavily on empirical double-exponential fits and lookup-table (LUT) simulations. These approaches fail to capture the underlying trap kinetics, cannot predict responses to stimulation protocols different from those measured, and are structurally blind to the volatility inherent in charge-trapping devices. Consequently, there is a gap in connecting device-level trap parameters to system-level performance metrics, hindering Design–Technology Co-Optimization (DTCO).

Methodology
The authors propose a physics-based compact model that couples a multi-level trap-kinetics ladder to a smooth field-effect transistor (FET) transducer.

Trap Kinetics: The model treats the device as an n-type MoS2 phototransistor where photogenerated holes are captured by $K$ independent trap populations at the channel/dielectric interface. The occupancy of each trap population is governed by linear rate equations involving capture cross-sections, trap densities, and release time constants. Crucially, for arbitrary optical pulse trains with constant-irradiance segments, the model derives an exact closed-form solution for trap occupancy propagation, avoiding numerical integration and enabling fast array-level simulation.
Transducer: The trapped charge shifts the threshold voltage, which is mapped to the drain current using a smooth, single-piece FET expression valid from subthreshold to strong inversion. This expression captures exponential, square-law, and linear regimes continuously.
Calibration: The model was calibrated against digitized measurements from two independently published devices with different material compositions and stimulation protocols: a monolayer MoS2/SiO2 transistor (Device A) and a CdSe-sensitized MoS2 heterostructure (Device B). The number of trap levels ( $K$ ) was selected per device using Akaike and Bayesian information criteria.
System-Level Simulation: The calibrated model was deployed in a single-layer optoelectronic classifier for handwritten digit recognition. Training followed a Manhattan rule using optical potentiation only, with differential device pairs representing synaptic weights.

Key Contributions

Exact Closed-Form Model: A compact model coupling a multi-level trap ladder to a smooth FET transducer, providing an exact analytical solution for arbitrary optical pulse trains without numerical integration.
Unified Framework: Demonstration that a single trap-ladder framework, with the trap count selected by information criteria, can reproduce both PPF and multi-decade retention across two distinct device types, independently recovering empirically reported relaxation constants.
Analytical Insights:
- Emergent Time Constants: The model reveals that apparent PPF time constants extracted via double-exponential fitting are emergent quantities limited by the optical pulse width, not fundamental trap parameters. Dynamics faster than the pulse width are aliased into fitted amplitudes.
- PPF Bound: It is analytically derived that for a linearly accumulating charge-trapping synapse, the PPF index cannot exceed 200%. Measured values exceeding this bound (e.g., 203.5%) necessarily encode transducer superlinearity (e.g., square-law saturation or exponential regimes).
DTCO Knob: Identification of programming irradiance as a co-optimization variable. Reducing irradiance reshapes the conductance ladder, lowering training energy without sacrificing accuracy.
Retention-Aware Analysis: Unlike LUT approaches, the framework explicitly tracks trap occupancies, enabling the analysis of inference-time weight decay, forgetting during slow training, and optical-refresh scheduling.

Results

Model Accuracy: For Device A, the model reproduced the PPF curve within 1.0% error and retention over four decades with a mean absolute error of 0.003. It successfully captured a PPF "floor" (~135%) that empirical double-exponential fits cannot represent. For Device B, the model captured sharp sub-second facilitation and long-term retention (>1500 s) using only two trap levels.
Predictive Power: An out-of-sample test predicted the response to a 50-pulse optical potentiation protocol using parameters fitted solely to PPF and retention. The predicted temporal dynamics were robust, and the measured curvature fell within the band spanned by the model's transducer forms.
Energy-Accuracy Trade-off: In a photo-programmed synaptic array, reducing programming irradiance from 10 mW/cm² to 1.5 mW/cm² reshaped the conductance ladder steps. This reduced training energy by 6.7-fold (from 7.7 mJ to 1.15 mJ) while maintaining or slightly improving recognition accuracy (92.6% to 93.4% vs. a 97.4% floating-point reference).
Retention Dynamics: The framework quantified inference-time weight decay, showing accuracy remaining above 94% for over 30 minutes and dropping to chance levels after ~8 hours. It also demonstrated that an optical-refresh strategy could sustain a 94% accuracy floor at sub-microwatt standby power (e.g., 0.94 µW for 300s refresh intervals).

Significance
The paper claims that replacing empirical fits with a calibrated, analytically solvable trap-kinetics model fundamentally changes the questions that can be asked of optoelectronic synapses. By explicitly linking trap physics to system metrics, the framework enables:

Predictive DTCO: Identifying irradiance as a design knob to optimize energy efficiency without accuracy penalties.
Retention-Aware Design: Quantifying the impact of volatility on training (forgetting) and inference (weight decay), which is inaccessible to static LUT approaches.
Reservoir Computing Bridge: Providing a calibrated fading-memory kernel for reservoir computing applications where volatility is a resource rather than a liability.
Reproducibility: Establishing a path from trap physics to system performance that is reproducible and capable of guiding future trap engineering and circuit design.

The authors note limitations, including the behavioral modeling of electrical depression (rather than a physical de-trapping model) and the lack of temperature dependence in the current release kinetics, suggesting these as areas for future extension.

Technical Summary: Physics-Based Compact Modeling and DTCO of MoS2-Based Optoelectronic Synapses

More like this