TCAD-Assisted Deep-Learning Surrogate Modeling of a Dielectric-Modulated FinFET Biosensor for HER2-Equivalent Dielectric-State Sensing

Technical Summary: TCAD-Assisted Deep-Learning Surrogate Modeling of a Dielectric-Modulated FinFET Biosensor

Problem Statement
Breast cancer prognosis relies heavily on HER2 status, yet conventional biochemical assessment requires complex laboratory infrastructure. While label-free semiconductor biosensors offer a compact alternative, the computational cost of repeatedly simulating dielectric-modulated Field-Effect Transistor (FET) devices using Technology Computer-Aided Design (TCAD) is prohibitive. TCAD simulations for various dielectric states, gate biases, and cavity geometries are time-intensive, creating a bottleneck for rapid design exploration. Furthermore, existing literature often lacks deep-learning surrogates specifically tailored to FinFET architectures that can predict electrical responses under effective dielectric states without claiming clinical biomarker quantification.

Methodology
The study proposes a two-part framework combining physics-based simulation with machine learning:

1. Device Modeling and TCAD Simulation:

   • A three-dimensional silicon FinFET with an analyte-accessible nanocavity near the gate stack is modeled.
   • The sensing mechanism relies on dielectric modulation: the nanocavity is treated as air-filled ($k=1$) for the reference state and filled with an effective dielectric medium ($k=8$) to represent a HER2-equivalent loading condition.
   • Simulations utilize drift-diffusion transport with standard physical models (Shockley-Read-Hall, Auger recombination, bandgap narrowing, etc.) to ensure physical consistency across subthreshold and inversion regimes.
   • The device is calibrated against reported FinFET characteristics to verify electrical behavior before biosensor analysis.

2. Deep-Learning Surrogate Framework:

   • Data Generation: TCAD generates $I_D-V_{GS}$ transfer characteristics for dielectric anchor states ($k = 1, 2.1, 5, 10$).
   • Model Architecture: A Multilayer Perceptron (MLP) is trained using gate voltage ($V_{GS}$) and the logarithm of the dielectric constant ($\log_{10}(k)$) as inputs. The target output is $\log_{10}(I_D)$ to preserve the exponential nature of subthreshold current.
   • Validation Strategy: The $k=8$ state (HER2-equivalent) is held out during training to serve as a blind interpolation validation case, ensuring the model operates within the bounded simulation window rather than extrapolating to arbitrary conditions.
   • Metric Extraction: Electrical metrics (Threshold Voltage $V_{TH}$, Sensitivity $S_{VTH}$, Sensitivity, $I_{ON}/I_{OFF}$, and Transconductance $g_m$) are extracted from both TCAD and MLP-predicted curves using standard definitions (e.g., constant-current method for $V_{TH}$).

Key Contributions

• FinFET-Specific Surrogate: Development of a deep-learning surrogate specifically for dielectric-modulated FinFETs, extending previous work on NSFETs and GAA structures.
• Bounded Interpolation Validation: Successful blind prediction of the HER2-equivalent ($k=8$) dielectric state, demonstrating the model's ability to interpolate between anchor states without overclaiming clinical diagnosis capabilities.
• Geometry-Aware Screening: Extension of the framework to screen cavity length and cavity thickness variations, linking geometric parameters to electrical performance metrics.
• Physics-Consistent Extraction: Demonstration that the surrogate preserves not only curve shapes but also derived physical quantities like transconductance and sensitivity, which are critical for biosensor design.

Results

• Reconstruction Accuracy: The MLP successfully reconstructed TCAD anchor states ($k=1, 2.1, 5, 10$) with $R^2$ values ranging from 0.9923 to 0.9950.
• Blind Validation ($k=8$): For the held-out HER2-equivalent state, the model achieved an $R^2$ of 0.9878 on linear current and 0.9986 on logarithmic current. The Root Mean Square Error (RMSE) was $2.79 \times 10^{-6}$ A.
• Sensing Metrics: For the predicted $k=8$ state, the extracted parameters were:
  • $I_{ON}/I_{OFF}$ ratio: $1.72 \times 10^9$
  • Threshold-voltage sensitivity ($S_{VTH}$): 0.26912
  • Sensitivity: $\approx 6300$
  • Relative errors compared to TCAD were minimal (e.g., 0.58% for $I_{ON}/I_{OFF}$).
• Geometry Screening:
  • Increasing cavity length ($6 \to 18$ nm) increased the $I_{ON}/I_{OFF}$ ratio by nearly an order of magnitude.
  • Decreasing cavity thickness ($5 \to 3$ nm) significantly improved electrostatic coupling and readout margin.
  • Validation errors for geometry screening remained low (0.23%–1.25%).
• Computational Efficiency: While a single TCAD sweep takes ~45 minutes, the trained MLP predicts a full 101-point transfer curve in ~0.02 seconds, reducing the time for geometry screening from hours to less than a second.

Significance and Claims
The paper claims that the proposed framework provides a computationally efficient route for device-level dielectric-state sensing and geometry-dependent FinFET biosensor design exploration. It explicitly states that the work is restricted to predicting electrical responses under effective dielectric states and does not model:

• HER2-antibody binding kinetics.
• Concentration-dependent biomolecule loading.
• Biological sample variability.
• Fabricated-device validation or clinical diagnosis.

The significance lies in bridging the gap between expensive TCAD simulations and rapid design iteration. By using a surrogate model trained on TCAD "anchor" data, researchers can rapidly screen dielectric-geometry trends and predict electrical metrics (such as $I_{ON}/I_{OFF}$ and sensitivity) for specific effective dielectric states (like the HER2-equivalent $k=8$) without running new, time-consuming simulations. The study emphasizes that while the results are promising for device physics and design optimization, they do not constitute a validated clinical screening tool for HER2 concentration.