Physics-informed AI Accelerated Retention Analysis of Ferroelectric Vertical NAND: From Day-Scale TCAD to Second-Scale Surrogate Model

(2) Semiconductor Research; (3) NVIDIA; Asif Khan (1); CA; Computer Engineering; Daewon Ha (2); Development; GA; Georgia Institute of Technology; Gyujun Jeong (1)

arxiv: 2603.06881 · v3 · submitted 2026-03-06 · 💻 cs.LG · cs.AI· physics.comp-ph

Physics-informed AI Accelerated Retention Analysis of Ferroelectric Vertical NAND: From Day-Scale TCAD to Second-Scale Surrogate Model

Gyujun Jeong (1) , Sungwon Cho (1) , Minji Shon (1) , Namhoon Kim (1) , Woohyun Hwang (2) , Kwangyou Seo (2) , Suhwan Lim (2) , Wanki Kim (2)

show 21 more authors

Daewon Ha (2) Prasanna Venkatesan (3) Kihang Youn (3) Ram Cherukuri (3) Yiyi Wang (3) Suman Datta (1) Asif Khan (1) Shimeng Yu (1) ((1) School of Electrical Computer Engineering Georgia Institute of Technology GA USA (2) Semiconductor Research Development Samsung Electronics Co. Ltd South Korea (3) NVIDIA Santa Clara CA USA)

This is my paper

Pith reviewed 2026-05-15 14:34 UTC · model grok-4.3

classification 💻 cs.LG cs.AIphysics.comp-ph

keywords ferroelectric vertical NANDretention analysisphysics-informed neural operatorsurrogate modelTCAD accelerationFeFETthreshold voltage shiftdata retention

0 comments

The pith

Physics-informed neural operator predicts ferroelectric vertical NAND retention 10000 times faster than TCAD.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to establish that a Physics-Informed Neural Operator framework can accurately surrogate the computationally expensive TCAD simulations for threshold voltage shifts and retention in Fe-VNAND devices. By integrating physical principles of charge detrapping and ferroelectric depolarization into the neural architecture, the model delivers over 10000x speedup. This acceleration makes it feasible to explore broad parameter spaces for device optimization. Readers should care because it provides a practical data engine for SPICE simulations and compact model development in emerging memory technologies.

Core claim

Embedding fundamental physical principles into a neural operator trained on limited TCAD data for a single FeFET configuration produces a surrogate model that predicts Vth shifts and retention behavior with maintained physical accuracy at a speedup exceeding 10000x compared to conventional TCAD tools.

What carries the argument

The Physics-Informed Neural Operator (PINO) that embeds standard physical principles to capture interactions between charge detrapping and ferroelectric depolarization.

Load-bearing premise

The interaction between charge detrapping and ferroelectric depolarization is accurately represented by embedding standard physical principles into a neural operator trained on a limited set of TCAD simulations for one FeFET configuration.

What would settle it

Running TCAD and the PINO model on a new set of device parameters or longer retention times and observing large discrepancies in predicted Vth shifts.

read the original abstract

Ferroelectric field-effect transistors (FeFET)-based vertical NAND (Fe-VNAND) has emerged as a promising candidate to overcome z-scaling limitations with lower programming voltages. However, the data retention of 3D Fe-VNAND is hindered by the complex interaction between charge detrapping and ferroelectric depolarization. Developing optimized device designs requires exploring an extensive parameter space, but the high computational cost of conventional Technology Computer-Aided Design (TCAD) tools makes such wide-scale optimization impractical. To overcome these simulation barriers, we present a Physics-Informed Neural Operator (PINO)-based AI surrogate model designed for high-efficiency prediction of threshold voltage (Vth) shifts and retention behavior. By embedding fundamental physical principles into the learning architecture, our PINO framework achieves a speedup exceeding 10000x compared to TCAD while maintaining physical accuracy. The resulting surrogate provides a physics-consistent data engine for compact model parameter extraction and look-up-table (LUT) generation, directly supporting reliability-aware SPICE simulation of Fe-VNAND. This study demonstrates the model's effectiveness on a single FeFET configuration, serving as a pathway toward modeling the retention loss mechanisms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PINO surrogate accelerates Fe-VNAND retention sims but needs more validation data to back the accuracy claims.

read the letter

The paper's core contribution is a PINO surrogate model that predicts threshold voltage shifts and retention behavior in Fe-VNAND much faster than traditional TCAD simulations. They report over 10,000 times speedup while trying to keep the results physically consistent by embedding physical principles into the neural operator. What stands out is the application to this specific retention problem in ferroelectric vertical NAND. The interaction between detrapping and depolarization is tricky for device design, and using TCAD for wide parameter sweeps is impractical. Turning that into a fast surrogate that can support SPICE modeling is a practical step. The authors show it works on their single FeFET test case, which is a start. The approach is reasonable. Physics-informed operators have worked in other domains for learning operators from data while respecting equations, so applying it here makes sense for better generalization than a plain neural net. Where it falls short is in the supporting evidence. The claims of maintained accuracy and the huge speedup are stated, but without numbers on mean error, maximum deviation, or how they validated against hold-out TCAD runs, it's difficult to judge. Also, everything is shown for one configuration. The dynamics change with different layer thicknesses or operating conditions, so we don't know if the model extrapolates or just interpolates the training set. That matches the stress-test concern about untested generalization. This work is aimed at researchers in semiconductor device simulation, especially those dealing with ferroelectric memories and reliability. Someone building models for 3D NAND variants would see value in the surrogate idea. It is worth sending for peer review because the problem is relevant and the method has potential, but the referees will need to see detailed error analysis and more test cases to be convinced. I would not cite it yet without those details, but it could be useful after revision.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a Physics-Informed Neural Operator (PINO) surrogate model to accelerate prediction of threshold voltage (Vth) shifts and retention behavior in ferroelectric vertical NAND (Fe-VNAND) devices. It claims that embedding fundamental physical principles into the neural operator yields >10000x speedup over TCAD simulations while preserving physical accuracy, demonstrated on a single FeFET configuration, and positioned to enable compact model extraction and SPICE simulations.

Significance. If the speedup and accuracy claims are quantitatively validated across configurations, the work would meaningfully advance device-level modeling for emerging non-volatile memories by making extensive parameter-space exploration feasible. The physics-informed neural operator approach directly targets the computationally expensive interaction of charge detrapping and ferroelectric depolarization. The current single-configuration demonstration, however, limits the assessed significance until broader testing is shown.

major comments (2)

[Abstract] Abstract: the central claim that the PINO 'maintains physical accuracy' while delivering >10000x speedup is unsupported by any quantitative error metrics (e.g., MAE or RMSE on Vth shifts or retention curves) or validation protocol details; without these numbers the speedup claim cannot be evaluated.
[Results] Results (single FeFET configuration): the model is trained and evaluated only on limited TCAD runs for one device geometry; no hold-out tests on varied thickness, voltage, or temperature are reported, leaving open whether the operator has captured the underlying detrapping-depolarization physics or merely interpolated the training distribution.

minor comments (1)

[Abstract] The abstract states the surrogate 'provides a physics-consistent data engine' but does not specify how the embedded physical principles are enforced (e.g., via loss terms or architecture constraints); a brief equation or diagram reference would clarify this.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and have revised the manuscript to provide the requested quantitative support and clarifications while preserving the scope of this initial demonstration study.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that the PINO 'maintains physical accuracy' while delivering >10000x speedup is unsupported by any quantitative error metrics (e.g., MAE or RMSE on Vth shifts or retention curves) or validation protocol details; without these numbers the speedup claim cannot be evaluated.

Authors: We agree that quantitative error metrics are required to substantiate the accuracy claim. In the revised manuscript we have added MAE and RMSE values computed on Vth shifts and full retention curves, together with an explicit description of the validation protocol (including train/test split ratios and the use of physics-informed loss terms during training). The >10000x speedup figure is now accompanied by the precise timing methodology (wall-clock TCAD simulation time versus PINO inference time on identical hardware). These numbers and protocol details have been inserted into both the abstract and the results section. revision: yes
Referee: [Results] Results (single FeFET configuration): the model is trained and evaluated only on limited TCAD runs for one device geometry; no hold-out tests on varied thickness, voltage, or temperature are reported, leaving open whether the operator has captured the underlying detrapping-depolarization physics or merely interpolated the training distribution.

Authors: The manuscript already states that the study demonstrates the approach on a single FeFET configuration as a proof-of-concept. The physics-informed loss functions are constructed directly from the charge-detrapping and ferroelectric-depolarization governing equations, which provides a mechanism for learning the underlying physics rather than pure data interpolation. In the revision we have added a dedicated limitations subsection that discusses generalization bounds, the role of the physics constraints, and the need for future multi-configuration datasets. No new hold-out tests across varied geometries are added in this revision, as they would require substantial additional TCAD data generation outside the present scope. revision: partial

Circularity Check

0 steps flagged

No significant circularity in PINO surrogate derivation for Fe-VNAND retention

full rationale

The paper trains a Physics-Informed Neural Operator on TCAD data for one FeFET configuration to predict Vth shifts and retention curves, claiming >10000x speedup while embedding standard physical principles. No quoted equations or sections reduce any claimed prediction to a fitted parameter by construction, nor invoke self-citations as load-bearing uniqueness theorems. The physics embedding uses external first-principles constraints rather than data-derived ansatzes, and the single-configuration demonstration is consistent with surrogate validation rather than a self-referential loop. The derivation chain remains self-contained against external TCAD benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The abstract relies on standard domain assumptions about charge detrapping and ferroelectric depolarization without introducing new free parameters or invented entities. No explicit fitting constants are named.

axioms (1)

domain assumption Fundamental physical principles of charge detrapping and ferroelectric depolarization govern retention behavior and can be embedded into a neural operator.
Invoked in the abstract as the basis for the PINO framework.

pith-pipeline@v0.9.0 · 5658 in / 1153 out tokens · 37450 ms · 2026-05-15T14:34:35.652231+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Physics-Informed Neural Operator (PINO) ... Poisson residual ... monotonicity constraints ... ReLU(|Pr|) + ReLU(|Qtrapped|)
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Fourier Neural Operator ... resolution-invariant 2D physical fields

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.