PaperlayerView exact revision on arXiv

Source-linked AI summary

AI and TCAD for Inverse Design and Defect Discovery: From Simple Machine Learning to LLM

Hiu Yung Wong

arXiv:2609.07046v1cs.LG

TL;DR

Semiconductor device design and defect discovery face limited data, including scarce defect observations and difficulty obtaining 3D tomography linked to electrical characteristics. The paper reviews simulation-augmented ML using TCAD-generated data, autoencoders, and noise engineering, alongside LLM applications. It reports that these approaches can learn device relationships and, in a TCAD-calibration example, perform similarly to an experienced author.

  • Problem

    Semiconductor device and defect ML is constrained by scarce data, including few defective samples and the near-impossibility of linking 3D tomography to electrical characteristics.

    Source: §I · p. 1

  • Method

    The paper reviews simulation-augmented ML trained on TCAD data, including autoencoder models, and discusses LLMs for generating TCAD input files.

    Source: §I · p. 1, A. Principal Components Analysis (PCA) · p. 2, §VI · p. 7

  • Results

    Autoencoder and noise-engineering approaches learn underlying physics and can apply to experimental data; AE-PINN calibration performs similarly to the experienced author’s result.

    Source: D. Automatic TCAD Parameter… · p. 6, §VII · p. 8

  • Takeaways & Limitations

    The reviewed ML approaches support inverse design, defect discovery, forward prediction, and device optimization, while LLMs and MLLMs may alleviate difficulties of traditional ML.

    Source: §V · p. 7, §VII · p. 8

    A new model is generally needed for each problem, and some cases still require domain expertise to avoid non-physical results.

    Source: §V · p. 7

Abstract

from arXiv · show

AI has revolutionized various engineering domains, but its impact on semiconductor device design and defect discovery is still limited, due to limited data and the curse of dimensionality. In this paper, we will discuss our work on using the Technology Computer-Aided-Design (TCAD) to generate precise data needed for machine learning (ML) to enable simulation-augmented ML. We demonstrate that with minimal domain expertise, it is possible to create a machine that performs as well as a device engineer on a specific task. We will show that auto-encoder-based machine learning models and noise engineering applied to TCAD data are effective at learning latent physics, and that the models can be seamlessly applied to experimental data. We will demonstrate how to build a device-engineer-level model step by step through various examples, including using only non-destructive electrical data to inverse-engineer the PiN diode layer thickness variations, the Ga2O3 Schottky diode doping and anode workfunction variations, and the transistor contact resistance in an inverter. Examples also include the generation of a FinFET IV/CV prediction model, the mapping between transistor images and IV curves, and the automatic calibration of TCAD parameters for a Ga2O3 Schottky diode, which can only be handled well by experienced TCAD engineers. Finally, to fully realize the potential of AI, large language models (LLMs) and multimodal LLMs (MLLMs) are believed to be necessary. We will discuss the application of LLMs to TCAD command file creation and our vision for MLLMs in automated device design and defect discovery.

I. INTRODUCTION

The paper motivates simulation-augmented ML as a way to address scarce semiconductor data by training on well-calibrated simulations. Its initial PIN-diode example shows that electrical curves can support both forward prediction and non-destructive inference of device variations.

I. INTRODUCTION

  • Device-physics and defect data are scarce because of confidentiality, few defective samples, and the near-impossibility of obtaining 3D tomography linked to electrical characteristics.The introduction identifies 3D tomography as necessary to fully understand defect physics and its relationship to electrical behavior.Source: §I · p. 1
  • Well-calibrated TCAD or SPICE simulations can generate training data for simulation-augmented ML, which the paper reviews for inverse design and defect identification.The paper uses TCAD-augmented ML and simulation-augmented ML for this approach.Source: §I · p. 1

II. INITIAL ATTEMPT OF TCAD-AUGMENTED ML

  • In PIN diodes, both doping and thickness affect reverse current, but only doping affects forward current.Fig. 1 shows IV curves for 2,000 simulated devices with layer-doping or thickness variations.Source: §II · pp. 1–2, Figure 1 · p. 2
  • Forward prediction of unseen IV curves does not eliminate the substantial computing resources required to generate TCAD training data.The paper cautions against claims that this approach alone reduces calculations from days to seconds.Source: §II · p. 2
  • A given PIN-diode IV curve can be used to infer thicknesses and doping non-destructively, supporting discovery of unintentional defects and process variations.The study reports that even simple linear regression reverse-engineers unseen TCAD IVs well; the concept was proposed in and independently proposed by Teo et al..Source: §II · p. 2

III. OVER-FITTING AND ITS ALLEVIATIONS

Models that predict parameters accurately from unseen TCAD data can still fail on experimental or differently varied data. The examples show overfitting when training variations do not cover the conditions encountered at inference.

III. OVER-FITTING AND ITS ALLEVIATIONS

  • The Ga2O3 Schottky-diode model predicts parameters accurately on unseen TCAD data but fails when given experimental IV curves.Fig. 2 contrasts experimental parameter predictions with and without PCA against physically expected boundaries.Source: §III · p. 2, Figure 2 · p. 2
  • A PIN model trained only on thickness-varying Set1 predicts absurd, including negative, thicknesses for IVs from doping-varying Sets2 and 3.Sets2 and 3 use fixed thicknesses matching or differing from the mean Set1 thickness.Source: §III · p. 2, Table II · p. 2

A. Principal Components Analysis (PCA)

The paper addresses overfitting in two ways: PCA constrains the features extracted from IV curves, while noise engineering makes training data less ideal. The noisy-data model predicts experimental temperature ranges, though not absolute values.

A. Principal Components Analysis (PCA)

  • Extracting four principal components before regression keeps experimental parameter predictions within expected ranges despite noise and unaccounted-for physics.The paper frames PCA as domain guidance that encourages the model to consider IV data as a whole rather than memorize individual points.Source: A. Principal Components Analysis (PCA) · p. 2

B. Noise Engineering

  • With 7 dB noise added to TCAD data, polynomial regression predicts the correct experimental temperature range, but not the absolute values.The noisy model also predicts ambient temperature within a reasonable range and avoids absurd PIN-diode thickness predictions.Source: B. Noise Engineering · pp. 2–3, Figure 3 · p. 3

C. Autoencoder

Autoencoders provide a way to learn a compact latent representation of device IV curves, which can be mapped to physical parameters with or without domain expertise. For Ga2O3 Schottky diodes, the model without domain expertise performs similarly to the expert-guided model and better than noise addition alone.

C. Autoencoder

  • With domain expertise, extracted turn-on voltage and subthreshold swing train predictions of workfunction and temperature, respectively.The extraction relies on the stated direct relationships between workfunction and turn-on voltage, and temperature and subthreshold swing.Source: C. Autoencoder · p. 3
  • The autoencoder compresses IV curves into a smaller latent space and reconstructs them, potentially encoding the physics underlying the curves.The bottleneck has fewer nodes than the input and output layers.Source: C. Autoencoder · p. 3
  • The model without domain expertise predicts Ga2O3 SBD workfunction and temperature similarly to the expert-guided model, and more accurately than noise addition alone.The workfunction comparison is shown as a scatter plot; temperature predictions are shown in box plots against expected ranges.Source: Figure 4 · p. 3, Figure 5 · p. 3, C. Autoencoder · p. 4
  • The latent-space size is usually set to the number of structural parameters to capture their degrees of freedom, though extra nodes may help convergence.Latent variables need not map one-to-one to structural parameters.Source: C. Autoencoder · p. 3
  • For Ga2O3 SBDs, 52-point IV curves are encoded into two latent variables that are linearly regressed to workfunction and temperature.Experimental IV curves pass through the autoencoder before the regressor predicts these parameters.Source: C. Autoencoder · p. 3

IV. AE-BASED MODELS

Autoencoder-based models are presented as a way to learn latent physics with minimal domain expertise and support several semiconductor modeling tasks. In an inverter example, a SPICE-trained machine infers drain contact resistances from the voltage-transfer characteristic despite imperfections in the model and measurements.

IV. AE-BASED MODELS

  • The AE-based approach is presented for defect discovery, IV prediction, inverse design, and automatic calibration while learning underlying physics with minimal domain expertise.In this framing, circuit defects, device images, and physical-model parameters are structural parameters correlated with electrical characteristics.Source: §IV · p. 4

A. Simulation-Augmented ML using SPICE

  • A SPICE-trained machine infers NMOS and PMOS drain contact resistances from an inverter VTC despite an imperfect SPICE model and measurement inaccuracies.The experiment used discrete resistors to mimic contact resistance because advanced equipment and dies were unavailable during the COVID-19 pandemic.Source: A. Simulation-Augmented ML using SPICE · p. 4, Figure 7 · p. 4
  • Each VTC is discretized into 51 points and encoded as two latent variables, h0 and h1; KNN with k=3 correlates them to predict R0 and R1, normalized to NMOS resistance Rn.Source: A. Simulation-Augmented ML using SPICE · p. 4

B. Surrogate Model for Advanced Devices

A modified autoencoder and polynomial regressor predict FinFET IV and CV curves from structural parameters. The reported models retain high accuracy with 25 training examples and with random stress variation intended to mimic unknown experimental physics.

B. Surrogate Model for Advanced Devices

  • The FinFET surrogate uses a three-node autoencoder latent space regressed to three structural parameters with a third-order polynomial to predict IV and CV curves.The parameters are mapped to latent values, which the decoder uses to generate the curves.Source: B. Surrogate Model for Advanced Devices · p. 4, Figure 9 · p. 5
  • DIBL is extracted from curves predicted by separate machines trained on linear and saturation IV, and its high accuracy is cited as evidence the machine might have learned underlying physics.The AE25 model is trained on only 25 examples.Source: B. Surrogate Model for Advanced Devices · p. 4
  • The approach is presented as enabling IV and CV curve prediction for a new technology before its physics is well understood.The reported stress variation is intended to mimic unknown physics variation in experiments.Source: B. Surrogate Model for Advanced Devices · p. 4
  • With 25 training examples, AE25 predicts ION, IOFF, gm, DIBL, SS, and CV with high accuracy; AE50SD also achieves high accuracy with random stress variation.The reported accuracy is summarized in Tables III and IV.Source: B. Surrogate Model for Advanced Devices · p. 4, Table III · p. 5, Table IV · p. 5

C. Correlating Images to Electrical Characteristics

The image-to-curve approach correlates latent representations from separate image and IV-curve autoencoders using a polynomial regressor. It predicts curves for hand-modified transistor structures and can inverse-design a structure for a target IV, although the inverse design need not reproduce the original geometry.

C. Correlating Images to Electrical Characteristics

  • Separate image and curve VAEs are linked by a third-order polynomial regressor, enabling image-to-IV prediction and IV-to-image inverse design.The model is trained on transistor data with five structural-parameter variations.Source: Figure 10 · p. 5, C. Correlating Images to Electrical… · p. 5
  • For hand-drawn structures with added irregularities, the machine predicts IVs very close to those of the corresponding TCAD-generated structures.The figure compares TCAD-generated structures, hand-modified versions, and their simulated or predicted IDVG curves.Source: Figure 11 · p. 5, C. Correlating Images to Electrical… · p. 5
  • An inverse-designed transistor need not match the source geometry because different transistor designs can produce the same IV; the model is also reported robust to input noise and weak or irrelevant parameters.The machine can redraw a transistor for a given IV well.Source: C. Correlating Images to Electrical… · p. 5

D. Automatic TCAD Parameter Calibration and Physics-

The authors use autoencoder-based models to calibrate Ga2O3 Schottky-diode PhuMob parameters from IV curves. Adding a physical loss lets the AE-PINN calibration perform similarly to expert manual calibration, including on experimental curves with noise and unaccounted-for physics.

D. Automatic TCAD Parameter Calibration and Physics-

  • A separate six-parameter PhuMob model uses 20,000 simulated IVs and a ten-dimensional latent space to extract parameters from unseen noisy TCAD curves.The extracted parameters are then used in TCAD IV simulations, which match the expected curves well.Source: D. Automatic TCAD Parameter… · pp. 5–6
  • The calibration flow trains an AE and PINN on 5.9k simulations, then feeds 66 experimental Ga2O3 SBD IV curves through them to extract parameters for TCAD verification.Each experimental curve can vary in workfunction and temperature, while the PhuMob parameters are shared.Source: Figure 12 · p. 6, D. Automatic TCAD Parameter… · p. 6
  • Adding a physical loss addresses unrealistic seven-parameter predictions, for which the predicted μ̂_min was sometimes smaller than μ̂_max.The authors state that domain knowledge was needed to include the physical loss term in NN training.Source: D. Automatic TCAD Parameter… · p. 6
  • AE-PINN calibration performs similarly to expert manual calibration and better than AE-NN on 66 experimental IV curves.The expert manual calibration took tens of hours; TCAD simulations using extracted parameters reproduced the experimental curves despite noise and unaccounted-for physics.Source: Figure 13 · p. 6, D. Automatic TCAD Parameter… · p. 6, Figure 12 · p. 6

E. Device Optimization through Surrogate Models

Surrogate models can pair ML-based forward prediction with an optimizer to inverse-design devices or systems. The GaN diode example searches guard-ring designs for a breakdown-voltage and forward-voltage/reserve-charge tradeoff.

E. Device Optimization through Surrogate Models

  • An NN surrogate maps GaN diode guard-ring parameters to electrical performance, and NSGA-II searches designs for the required breakdown voltage and (VFQ)^-1 tradeoff.VF is forward voltage and Q is reserve capacitive charge; the framework uses 266 TCAD-generated sub-optimal structures to train the surrogate and find a Pareto front.Source: E. Device Optimization through… · pp. 6–7, Figure 14 · p. 6

V. LIMITATIONS OF TRADITIONAL ML

Traditional simulation-augmented ML has limits in training-data demands, dimensionality, and generalization, while conditional autoencoders can address fixed-bias prediction. The paper also discusses LLM and MLLM applications, including TCAD input generation and transistor-image parameter inference.

V. LIMITATIONS OF TRADITIONAL ML

  • Traditional ML requires a new model for each problem, substantial simulation data, and domain expertise in some cases; performance degrades beyond the seven parameters tried so far.The authors also note limited out-of-range prediction and prolonged, low-convergence simulations for some tasks.Source: §V · p. 7
  • A conditional variational autoencoder treats fixed bias as a condition, allowing smooth transitions between fixed biases without generating data for every condition.The accompanying figure depicts CVAE and regressor training followed by construction of the surrogate model.Source: §V · p. 7, Figure 15 · p. 7

VI. LARGE LANGUAGE MODEL (LLM)

  • The authors propose that fine-tuned LLMs may bring semiconductor domain expertise and reasoning to high-dimensional problems, while MLLMs could use device-physics data from TCAD.They say a semiconductor database based on TCAD, including 3D tomography, is needed and that greater collaboration is required.Source: §VI · p. 7
  • Fine-tuned Llama 2 and 3 generated Sentaurus Structure Editor files for 18-parameter nanowires from English instructions using 7,000 training pairs.In early 2024, fine-tuned Llama 2 outperformed non-fine-tuned ChatGPT-4; by July 2024, ChatGPT-4o with in-context learning handled the task well.Source: §VI · p. 7
  • Fine-tuned Qwen2.5-VL associated transistor parameters with unannotated images and measured dimensions in unseen images.The training pairs combine transistor images with parameters such as gate length, oxide thickness, and spacer width.Source: §VI · pp. 7–8

VII. CONCLUSIONS

The paper reviews simulation-augmented ML for semiconductor defect discovery and reverse engineering, emphasizing TCAD-generated training data. It concludes that LLM and MLLM applications may alleviate difficulties of traditional ML, while a TCAD-based semiconductor database with 3D tomography is needed.

VII. CONCLUSIONS

  • Simulation-augmented ML with noise engineering and manifold learning can learn underlying physics and apply to experimental data despite noise and unaccounted-for physics.The reviewed applications use TCAD and simulation data to learn correlations between electrical characteristics and device parameters with little to no domain expertise.Source: §VII · p. 8
  • The authors conclude that LLMs and MLLMs may alleviate traditional ML difficulties, but a semiconductor database with 3D tomography is needed and can be obtained through TCAD simulations.Source: §VII · p. 8
Loading 2609.07046v1…