🔬 applied physics

Rapid FinFET Modelling Using an Autoencoder

This paper presents a machine learning framework using an autoencoder to efficiently model FinFETs by compressing calibrated I-V data into a low-dimensional latent space, enabling rapid and accurate reconstruction of device characteristics and extraction of key metrics with minimal training data.

Original authors: Amit Sarkar Suman Sau, Swagata Mandal

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Amit Sarkar Suman Sau, Swagata Mandal

Original paper licensed under CC BY 4.0 (http://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 you are trying to teach a computer to understand how a tiny, high-tech switch called a FinFET works. These switches are the building blocks of modern smartphones and computers, but they are so complex that simulating them on a computer is like trying to solve a massive, intricate puzzle that takes a long time and requires a PhD in physics to even start.

This paper presents a clever shortcut using Artificial Intelligence (AI) to make this process fast and easy. Here is how they did it, explained simply:

1. The Problem: The "Heavy" Simulation

Usually, engineers use a heavy, complex tool (called TCAD or BSIM-CMG) to predict how these switches behave. It's accurate, but it's slow and requires a lot of expert knowledge to set up. It's like trying to bake a perfect cake by calculating the exact chemical reaction of every single ingredient from scratch every time.

2. The Solution: The "Smart Summarizer" (Autoencoder)

The authors built a special AI tool called an Autoencoder. Think of this AI as a super-efficient librarian or a summarizer.

The Input: They fed the AI thousands of "stories" about how the switch behaves. These stories are graphs showing how much electricity flows through the switch at different settings (specifically, the Gate Voltage and Drain Voltage).
The Compression (The Magic): Instead of memorizing every single detail of the graph, the AI learns to compress the whole story into a tiny, 3-number "secret code" (called a latent space).
- Analogy: Imagine you have a 100-page novel. The AI reads it and realizes the whole story can be perfectly summarized by just three words: "Hero," "Villain," and "Twist." It doesn't need the whole book to remember the plot; it just needs those three keywords.
The Innovation: The authors made sure to include the Drain Voltage (how hard the electricity is being pushed) as one of the inputs. This is like telling the librarian, "Don't just summarize the story; also tell me how the story changes if the weather is rainy." This helps the AI understand how the switch behaves under different pressures.

3. The Training: Learning from a "Calibrated" Teacher

Before teaching the AI, the researchers used the heavy, slow simulation tool (BSIM-CMG) to generate a "textbook" dataset. They made sure this textbook matched real-world hardware data perfectly. They then fed this textbook to the AI.

The Process: The AI tried to read a compressed "secret code" and then rebuild the original graph from it.
The Goal: If the AI could rebuild the graph perfectly, it meant it truly understood the physics of the switch.

4. The Results: Fast and Accurate

Once trained, the AI proved to be incredibly good at two things:

Rebuilding the Graphs: It could take a tiny "secret code" and draw the full electricity curve almost perfectly.
Extracting Key Stats: It could look at the curve and instantly tell you important numbers, like the Threshold Voltage (the exact moment the switch turns on) and Transconductance (how well it amplifies signals).

The Proof:
When they compared the AI's predictions to the heavy, slow simulation tool, the results matched almost perfectly (over 99% accuracy in some cases). The AI made very few mistakes in predicting how the switch turns on or how fast it reacts, even when the voltage settings changed.

5. Why This Matters (According to the Paper)

The paper claims this method is a powerful new tool for rapid device modeling.

Instead of waiting hours for a slow simulation, engineers can use this AI to get the same results instantly.
It removes the need for deep, manual physics expertise to interpret the data because the AI learned the patterns automatically.
It is specifically useful for designing Analog, Digital, and RF circuits (the brains of our electronic devices).

In a nutshell: The authors taught an AI to read the "language" of a tiny electronic switch, summarize it into a few numbers, and then instantly redraw the behavior and calculate the important stats. It's a fast, accurate shortcut that replaces the need for slow, complex simulations.

Technical Summary: Rapid FinFET Modelling Using an Autoencoder

Problem Statement
The scaling of semiconductor devices, particularly FinFETs, has introduced significant complexity in understanding device physics and modeling. Traditional Technology Computer-Aided Design (TCAD) simulations, while indispensable, are often hindered by high computational costs, convergence issues, and the necessity for extensive domain expertise. While machine learning (ML) has emerged as a tool to accelerate simulations and extract physical insights, a central challenge remains: developing models that are accurate within their training domain, capable of generalization, and able to minimize reliance on specialized manual feature extraction. Existing approaches often struggle to balance data-driven efficiency with the accurate capture of bias-dependent variations without explicit physical constraints.

Methodology
The authors propose a machine learning framework utilizing an autoencoder (AE) to model FinFET current-voltage ( $I_D$ - $V_G$ ) characteristics. The methodology proceeds as follows:

Data Generation: A dataset was generated using a calibrated BSIM-CMG compact model, tuned to correlate with experimental data from 12-nm p-channel bulk FinFETs. The devices featured a 14 nm gate length, 30.5 nm fin height, and 14 nm fin width.
Preprocessing: Drain current versus gate voltage curves were generated at specific drain-to-source voltages ( $V_{DS}$ ), with $V_G$ swept from 0 V to -0.8 V. To ensure consistency, curves were interpolated onto a common $V_G$ vector. Given the wide dynamic range of drain current, a base-10 logarithmic transformation was applied. The data underwent z-score normalization using mean and standard deviation calculated from the training pool.
Network Architecture: A symmetric encoder-decoder neural network was employed.
- Input: A flattened vector comprising normalized $I_D$ points concatenated with the normalized $V_{DS}$ value.
- Encoder: Two fully connected hidden layers (50 neurons each, ReLU activation) compressing the input into a 3-neuron latent space bottleneck.
- Decoder: Two fully connected layers mirroring the encoder to reconstruct the original input vector.
- Training: The model was trained using the Adam optimizer for up to 500 epochs.
Evaluation: Beyond reconstruction error, the model's efficacy was assessed by extracting critical device metrics from the reconstructed curves: threshold voltage ( $V_{TH}$ ), subthreshold slope ($SS$), and peak transconductance ( $g_m$ ). These were compared against the original BSIM-CMG data.

Key Contributions

Latent Space Compression: The work demonstrates the ability to compress full $I_D$ - $V_G$ curves into a low-dimensional latent space that intrinsically encodes key device physics.
Explicit Bias Incorporation: A notable innovation is the explicit inclusion of drain-to-source voltage ( $V_{DS}$ ) as an input feature. This enhances the model's ability to capture bias-dependent variations, a factor often treated separately in simpler regression models.
Direct Metric Extraction: The framework successfully reconstructs full curves and directly extracts critical parameters ( $V_{TH}$ , $SS$, $g_m$ ) without requiring manual feature engineering or complex post-processing.
Data Efficiency: The approach leverages manifold learning principles, suggesting that high accuracy can be achieved with minimal training data (referencing prior work where 25–50 samples sufficed for similar tasks).

Results
The proposed AI-based model was validated against the industry-standard BSIM-CMG model:

Curve Correlation: The model demonstrated strong correlation with BSIM-CMG for both drain current and transconductance across various gate voltages.
Statistical Accuracy: The coefficient of determination ( $R^2$ ) consistently exceeded 0.9 across different $V_{DS}$ values.
Parameter Extraction Error:
- Threshold Voltage ( $V_{TH}$ ): The percentage error between the AI-predicted and BSIM-extracted $V_{TH}$ was consistently low, confirming accurate capture of threshold dependence on drain bias.
- Subthreshold Slope ($SS$): Similarly, low percentage errors were observed in the predicted subthreshold swing.
- Transconductance ( $g_m$ ): The error in peak transconductance prediction remained within a 0% to 5% range. The authors note that $g_m$ exhibits slightly more variability than $V_{TH}$ or $SS$, which is expected as it is a differential parameter highly sensitive to minor curve shape changes.

Significance and Claims
The paper concludes that the proposed autoencoder framework provides a powerful tool for rapid device characterization and modeling. By accurately reproducing key device characteristics with low prediction error, the model demonstrates robustness suitable for application in Analog, digital, and RF circuit design. The authors position this data-driven approach as a viable alternative to traditional compact models, offering a pathway to faster modeling and high accuracy for both existing and upcoming device technologies, all while reducing dependency on manual domain expertise for feature extraction.