materials-chemistry/ai in a supporting role/Advanced Science 2025 · v2
Thin molybdenum disulfide memory cells tested, with reinforcement learning fitting their circuit model
Researchers built flash memory cells with channels of molybdenum disulfide just a few nanometres thick and measured how they stored charge. A reinforcement-learning agent then tuned a compact circuit model until its curves matched the measurements.
spectrum · one line per step, placed by what the step does · bright lines used AI
MoS 2 Channel‐Enhanced High‐Density Charge Trap Flash Memory and Machine Learning‐Assisted Sensing Methodologies for Memory‐Centric Computing Systems
Advanced Science, 2025
doi:10.1002/advs.202501926 · record aix-00165 v2 · checked 2026-10-09
- AI was for
- Simulation surrogate
- Model family
- Multilayer perceptron
- Checked by
- None stated
- Code
- not reported
AI processed or interpreted data, but the main finding does not rest on it.
What this research was about
Flash memory stores a bit by parking electric charge in an insulating layer above a tiny transistor. The trapped charge shifts the voltage at which the transistor switches on, and reading that shift reads the bit. Squeezing more memory into the same area means making each cell smaller, and the thin silicon channel that carries current through the cell starts to behave badly at those sizes. One alternative is molybdenum disulfide, a material that can be peeled into sheets only atoms thick while still conducting. How such a cell behaves depends on how thick that sheet is, and thickness affects writing, erasing and how long the charge stays put.
The team made charge-trap memory cells with molybdenum disulfide channels of 1.3, 3.2 and 6.5 nanometres, using gold nanoparticles as the charge store and a thin low-k layer for charge to tunnel through. They measured the memory window, the speed of programming and erasing, how many cycles the cells survived and how well they held charge. They also wanted to know whether such cells could be read reliably by real circuitry, so they fabricated a page-buffer chip in a 180-nanometre process and characterised its noise.
Where AI came in
Reading a memory cell in a circuit simulator needs a compact model: a set of equations with adjustable numbers, here the industry-standard BSIM form, whose parameters are normally hand-tuned until the simulated behaviour matches the real device. The researchers handed that tuning to a reinforcement-learning agent, an algorithm that learns by trying actions and being scored on the result. The agent proposed values for parameters such as threshold voltage, gate width, gate length and mobility; the simulator was run with them; the mismatch against the measured curves became the score. The loop repeated until the mean squared error fell below 0.1 per cent.
So the AI stood in for expert judgement in curve fitting, and nothing more. The devices were made and measured without it, the physics simulations ran without it, and the sensing-margin checks used the fitted model rather than any learned prediction. The record notes the only accuracy figure reported is the error against the same curves the agent was fitting.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The study fabricated charge-trap flash memory devices using mechanically exfoliated MoS2 channels of 1.3, 3.2 and 6.5 nm thickness, a 13 nm low-k pV3D3 tunnelling layer and Au nanoparticles as the charge-storage layer, then measured memory window, program and erase speed, endurance and retention against channel thickness. A deep reinforcement learning agent (DDPG) tuned BSIM compact-model parameters until simulated transfer curves matched the measured curves to a mean squared error below 0.1%, producing a SPICE model of the MoS2 cell. That model was combined with a page-buffer circuit fabricated in a 180 nm CMOS process to simulate sensing-node voltage differentials for word-line deviations of 10, 25 and 50 mV. The memory window saturated at about 6 V and was independent of channel thickness, thicker channels programmed more slowly but erased faster, the 1.3 nm device showed the poorest retention after endurance cycling, and the 1.3 nm device gave the largest simulated sensing margin.
How AI was used
A Deep Deterministic Policy Gradient reinforcement-learning agent was used to calibrate BSIM compact-model parameters so that a SPICE netlist reproduced device characteristics obtained from TCAD simulation or physical measurement. Measured transfer characteristics of MoS2 devices with 1.3, 3.2 and 6.5 nm channels were loaded as target model data. The agent's actor network predicted parameters including threshold voltage, gate width, gate length and mobility; the SPICE environment was then run with those parameters, and the mean squared error between the simulated and target curves served as the reward signal. The actor-critic formulation was chosen for its continuous action space, and the loop iterated until the mean squared error fell below a set threshold, reported here as less than 0.1%. The resulting calibrated BSIM model was then placed in a testbench with a fabricated page-buffer circuit for circuit-level sensing-margin simulation; no learned model was used in device fabrication, electrical measurement, TCAD simulation or the sensing-margin evaluation itself.
The shape of the work
Structural · the record, drawn
no AI
Fabricate MoS2 charge-trap memory devices
Physical execution, by hand or by robot.
MoS2 channels were formed through mechanical exfoliation onto a 90‐nm‐thick SiO2 layerwhere the paper describes this · verbatim
no AI
Measure electrical, memory and reliability characteristics
Obtaining raw data, whether by measurement, download or retrieval.
The fabricated MoS2 memory devices were evaluated using a Keithley 4200 semiconductor parameter analyzerwhere the paper describes this · verbatim
no AI
Calibrate TCAD device simulation to measured I-V curves
Numerical or physics simulation, including where a learned surrogate replaces it.
We successfully calibrated the experimental I‐V curves for 1.3 and 6.5 nm MoS2 channelswhere the paper describes this · verbatim
no AI
Simulate candidate BSIM netlist in SPICE
Numerical or physics simulation, including where a learned surrogate replaces it.
the agent environment modifies BSIM model parameters in the SPICE netlist to facilitate simulation executionwhere the paper describes this · verbatim
AI
Optimise BSIM parameters with DDPG agent
Iterative search over a space. The AI stood in for expert judgement. Its result feeds back into an earlier step.
The learning process involves running SPICE simulations with the predicted parameters and comparing the resulting curves against the target model curveswhere the paper describes this · verbatim
no AI
Fabricate page-buffer chip and characterise noise
Physical execution, by hand or by robot.
the prototype page buffer chip was fabricated using TSMC 180 nm technologywhere the paper describes this · verbatim
no AI
Verify sensing margin at circuit level
Testing outputs against ground truth.
the sensing margin characteristics of the NAND device during sensing operations were verifiedwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The device-level findings (memory window, program/erase speed, endurance, retention) come from fabrication and electrical measurement. The reinforcement-learning agent only fitted a compact SPICE model to already-measured transfer curves so that circuit-level sensing margins could be simulated.
leverages Deep Deterministic Policy Gradient (DDPG) reinforcement learning to optimize BSIM model parameterswhere the paper describes this · verbatim
The data that support the findings of this study are available from the corresponding author upon reasonable request.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- ValidationNo validation of the AI is described.
- Version of DDPG agent (actor-critic) for BSIM parameter optimisationWhich version of the model was used is not stated.
About this article
Record aix-00165, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error