13. Device Characterization, TCAD, and Data Analytics

Introduction and Context

Semiconductor devices are understood not merely by their design equations, but by how closely fabricated structures reproduce those equations in practice. Electrical measurements reveal current, voltage, capacitance, resistance, noise, switching, and breakdown behaviour. Optical, structural, and thermal measurements explain why that behaviour occurs. Technology computer-aided design (TCAD) connects material properties and process conditions to device operation, while data analytics converts large collections of measurements into evidence for process control, reliability assessment, and yield improvement.

At postgraduate level, characterization is therefore an inverse problem. A measured result is an observable consequence of hidden parameters such as doping concentration, mobility, interface-trap density, contact resistance, geometry, defect density, or thermal resistance. The engineer must determine which physical model explains the observations, estimate parameters with uncertainty, and distinguish genuine device physics from measurement artefacts. This chapter develops that integrated workflow.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Select suitable electrical, optical, structural, and thermal characterization methods for semiconductor devices.

2. Extract physical and compact-model parameters from measured data.

3. Quantify uncertainty, identify systematic error, and judge parameter identifiability.

4. Formulate and validate TCAD simulations using appropriate physical models and boundary conditions.

5. Design experiments efficiently and apply statistical process control to device manufacturing.

6. Use data analytics and AI-assisted methods responsibly for defect classification and yield improvement.

7. Interpret characterization results in technical, manufacturing, and professional contexts.

From Observable Behaviour to Physical Parameters

A characterization experiment consists of a device under test, an excitation, a measurement chain, and an inference procedure. For a diode, the measured current may be represented approximately by

\[

I=I_0[\exp(qV/nkT)-1],

\]

but the measured curve may also contain series resistance, shunt leakage, self-heating, instrument offsets, and contact effects. A useful model must therefore be judged over a specified bias and temperature range rather than fitted indiscriminately to the entire curve.

Parameter extraction is the estimation of quantities such as threshold voltage, subthreshold swing, transconductance, channel-length modulation, ideality factor, series resistance, breakdown voltage, carrier lifetime, interface-trap density, or thermal resistance. Extraction methods include graphical constructions, local derivatives, nonlinear least squares, Bayesian inference, and physics-informed regression. A parameter is identifiable only when the available data distinguish it from competing parameters. For example, mobility and effective channel length can both influence transconductance; an independent length measurement or additional bias data may be necessary.

A sound workflow separates three stages. First, data are acquired under controlled conditions and checked for instrument compliance, grounding, temperature, and repeatability. Second, a physical or empirical model is selected before fitting. Third, the fitted model is validated using data not used for estimation. A small residual error does not prove physical correctness: an over-parameterized model can fit noise exceptionally well.

Electrical Characterization

Current–Voltage Measurements

DC I–V characterization is the starting point for diodes, transistors, photodetectors, LEDs, and power devices. Important observables include turn-on behaviour, leakage, rectification ratio, contact resistance, breakdown, hysteresis, and current crowding. Measurements should specify sweep direction, step size, compliance limits, settling time, temperature, and whether a two-wire or Kelvin connection was used.

For a diode, the ideality factor can be estimated from

\[

n=\frac{q}{kT}\left(\frac{dV}{d\ln I}\right),

\]

in a region where series resistance and high-level injection are negligible. At higher current, the voltage drop across series resistance distorts the slope. At low current, instrumentation noise and shunt leakage dominate. Consequently, a single straight-line fit to a semilogarithmic plot is often misleading.

For a MOSFET, transfer characteristics yield threshold voltage and transconductance, while output characteristics reveal channel resistance, saturation, velocity saturation, and output conductance. Threshold voltage may be defined by a constant-current method, a linear extrapolation, or a charge-based method. These definitions are not interchangeable; comparisons must state the extraction convention. Subthreshold swing is commonly obtained from the slope of log drain current against gate voltage:

\[

S=\left(\frac{d\log_{10}I_D}{dV_G}\right)^{-1}.

\]

A change in subthreshold swing can indicate altered interface traps, oxide capacitance, temperature, or measurement leakage.

Capacitance, Impedance, and Pulsed Tests

C–V measurements probe depletion width, doping, oxide thickness, flat-band voltage, interface states, and mobile charge. Frequency, AC amplitude, DC sweep rate, and series resistance must be controlled. In a Schottky or p–n junction, a one-sided abrupt-junction approximation gives a relationship between inverse capacitance squared and applied voltage, but nonuniform doping, deep levels, and parasitic capacitance can invalidate a simple linear interpretation.

Small-signal impedance measurements provide frequency-dependent information about traps, recombination, contacts, and transport. Equivalent circuits are useful representations, not unique physical proofs. A constant-phase element, for example, may describe distributed relaxation but does not by itself identify a particular microscopic defect.

Pulsed I–V and transient measurements reduce self-heating and expose trapping or detrapping. They are particularly valuable for wide-bandgap transistors, where dynamic on-resistance and current collapse may not be visible in slow DC sweeps. The pulse width and duty cycle must be short enough to control temperature but long enough for the intended electrical response to develop.

Noise and Reliability Measurements

Noise spectroscopy examines random fluctuations such as shot noise, thermal noise, generation–recombination noise, and low-frequency 1/f noise. Noise is often sensitive to traps before a large DC degradation becomes visible. Reliability tests may include constant-voltage stress, constant-current stress, bias-temperature stress, hot-carrier stress, time-dependent dielectric breakdown, electromigration, and accelerated thermal cycling. Acceleration models must be physically justified; extrapolating far beyond the tested range can create large errors.

Optical, Structural, and Thermal Characterization

Optical methods are essential for materials and optoelectronic devices. Photoluminescence reveals bandgap-related transitions, strain, alloy composition, and nonradiative recombination. Raman spectroscopy can indicate crystal stress, temperature, phonon behaviour, and damage. Spectral response and external quantum efficiency connect photon wavelength to carrier generation and collection. Time-resolved photoluminescence provides recombination lifetimes, although the measured decay may contain multiple channels and instrument-response effects.

Structural techniques establish whether the intended physical structure was fabricated. X-ray diffraction measures lattice spacing, crystallinity, orientation, and strain. Scanning electron microscopy and transmission electron microscopy reveal morphology, interfaces, dislocations, and layer thickness. Atomic force microscopy quantifies surface roughness. Secondary-ion mass spectrometry provides depth-dependent impurity profiles, while four-point-probe mapping measures sheet resistance without requiring a fully fabricated device.

Thermal characterization is increasingly important as dimensions shrink and power density rises. Infrared thermography offers spatial temperature maps but depends on emissivity and optical calibration. Raman thermometry can provide local temperature estimates when the phonon response is properly calibrated. Electrical thermal transient testing extracts junction-to-case or junction-to-ambient thermal impedance. A common compact relation is

\[

T_J=T_A+P R_{\theta JA},

\]

but transient thermal behaviour requires a thermal impedance function rather than a single resistance. Self-heating can alter mobility, threshold voltage, leakage, and reliability during the very measurement intended to characterize the device.

Measurement Uncertainty and Data Quality

Uncertainty is not the same as error. Error is the difference from an unknown true value; uncertainty expresses limited knowledge of that value. Sources include calibration uncertainty, repeatability, reproducibility, resolution, drift, environmental variation, model assumptions, and sample-to-sample variation.

A measurand should be defined before analysis. If \(y=f(x_1,x_2,\ldots,x_n)\), the first-order combined standard uncertainty is approximately

\[

u_y^2=\sum_i\left(\frac{\partial f}{\partial x_i}\right)^2u_{x_i}^2

+2\sum_{i<j}\frac{\partial f}{\partial x_i}\frac{\partial f}{\partial x_j}\operatorname{cov}(x_i,x_j).

\]

The covariance terms matter when quantities share calibration or are derived from the same measurement. Repeated measurements estimate random variation, whereas calibration and reference standards address systematic components. Reporting a fitted parameter without its confidence interval, posterior distribution, or sensitivity information gives an incomplete picture.

Good practice includes instrument verification, open/short/load corrections, shielding, guarding for low-current work, four-terminal connections for low resistance, thermal equilibration, randomized measurement order, replicate samples, and preservation of raw data. Outliers should not be deleted merely because they disagree with a preferred model. Their physical origin should be investigated.

TCAD as a Physics-Based Laboratory

TCAD usually combines process simulation, device simulation, and visualization or extraction tools. Process simulation predicts implantation, diffusion, oxidation, deposition, etching, and annealing. Device simulation solves semiconductor equations on the resulting structure. The basic equations include Poisson’s equation, carrier continuity equations, and transport relations. Drift–diffusion is often sufficient for conventional devices, while hydrodynamic, energy-balance, quantum-correction, tunnelling, or Monte Carlo models may be required in advanced regimes.

A simulation begins with a question, not with a mesh. The engineer defines geometry, materials, contacts, doping, physical models, boundary conditions, and output quantities. Mesh refinement is needed near junctions, interfaces, high-field regions, and thin oxides, but excessive refinement increases computation without necessarily improving the answer. Mesh-convergence testing should demonstrate that the reported quantity is stable.

Calibration is the disciplined adjustment of uncertain model or process parameters against selected experimental observables. Validation uses independent observables or conditions. If threshold voltage and on-current are both used for calibration, breakdown behaviour should be reserved for validation where possible. TCAD results are conditional predictions: they inherit assumptions about material constants, defect distributions, interface charge, mobility, recombination, and geometry.

A useful TCAD workflow is:

1. State the physical hypothesis and measurable outputs.

2. Build the simplest defensible structure.

3. Select models based on the operating regime.

4. Check mesh, solver convergence, and conservation laws.

5. Perform sensitivity analysis on uncertain parameters.

6. Calibrate against reliable measurements.

7. Validate with independent data.

8. Record versions, input files, numerical settings, and extraction scripts.

Convergence is necessary but not sufficient. A numerically converged solution may still be physically inappropriate. Conversely, solver failure may result from poor scaling, abrupt bias stepping, or an unsuitable initial solution rather than from impossible device physics.

Design of Experiments and Statistical Process Control

One-factor-at-a-time experimentation is inefficient when factors interact. Design of experiments (DoE) varies process factors systematically to estimate main effects and interactions. Screening designs identify influential variables; factorial designs quantify interactions; response-surface methods optimize a response within a process window. Factors may include implant dose, anneal temperature, film thickness, gas flow, lithographic bias, or contact-metal condition. Responses may include leakage, threshold voltage, breakdown voltage, sheet resistance, or yield.

Randomization protects against time-related confounding, blocking groups similar wafers or lots, and replication estimates experimental variation. The distinction between statistical significance and engineering significance is essential: a small effect may be statistically detectable yet irrelevant to specification performance.

Statistical process control monitors variation over time. Control limits are estimated from a stable process and are not the same as specification limits imposed by design or customers. An \(X\)-bar chart monitors subgroup means; an R or S chart monitors within-subgroup dispersion. Capability indices such as

\[

C_p=\frac{USL-LSL}{6\sigma},\qquad

C_{pk}=\min\left(\frac{USL-\mu}{3\sigma},\frac{\mu-LSL}{3\sigma}\right)

\]

summarize potential and actual capability under appropriate assumptions. Capability indices are meaningless when the process is unstable or the data are strongly non-normal without suitable treatment. A control-chart signal should trigger technical investigation, not automatic adjustment that may increase variation.

Data Analytics and Responsible AI

Device manufacturing generates wafer maps, parametric test results, spectra, images, reliability traces, and equipment logs. Analytics can support defect classification, virtual metrology, anomaly detection, root-cause analysis, and yield prediction. Feature engineering should be grounded in device physics: spatial gradients, hysteresis area, derivative features, spectral peak shifts, transient time constants, and temperature-normalized quantities may be more informative than raw values alone.

Supervised learning requires labelled examples and careful separation of training, validation, and test data. Randomly splitting measurements from the same wafer into all three sets can cause data leakage and overstate performance. A wafer-, lot-, or time-based split is usually more realistic. Class imbalance is common because severe defects are rare; precision, recall, F1 score, receiver operating characteristic analysis, and the cost of false negatives should be considered rather than accuracy alone.

Unsupervised methods such as principal-component analysis, clustering, and autoencoders can reveal structure or anomalies, but an anomaly is not automatically a defect. Explainability tools, physical constraints, uncertainty estimates, and human review are important in high-consequence decisions. A model should be monitored for sensor drift, process changes, and population shift. Training data may encode historical operator bias or underrepresent new failure mechanisms.

Responsible deployment requires traceability, access control, data privacy, documented performance limits, audit logs, and a defined human override. AI may prioritize wafers for inspection or suggest a likely root cause; it should not silently alter a safety-critical process without authorization and verification.

Illustrative Case Study: Diagnosing Excess Leakage in a Wide-Bandgap Diode

Background

A fictional power-device manufacturer observes that a new batch of vertical wide-bandgap diodes meets the forward-voltage target but has leakage current approximately three times higher than the historical median at elevated reverse bias.

Situation and Analysis

Electrical screening shows that leakage is spatially clustered on several wafers rather than uniformly distributed. Reverse I–V curves display a temperature-sensitive component, while capacitance measurements indicate that the intended depletion width is approximately correct. Optical inspection finds no obvious surface contamination. A wafer-level map of leakage is aligned with a map of local sheet resistance.

The engineering team forms three hypotheses: bulk generation centres, edge termination damage, and a contact-related shunt. Temperature-dependent I–V data favour generation–recombination in the high-leakage population. Deep-level optical and electrical measurements are then performed on representative dies. Cross-sectional microscopy identifies a process-dependent defect concentration near an implanted termination region.

TCAD and Experimental Strategy

A calibrated TCAD model first reproduces the nominal electric-field distribution. Sensitivity analysis shows that a moderate increase in interface-trap density near the termination can reproduce the leakage trend without substantially changing forward voltage. A fractional factorial DoE is used to vary implant dose, anneal temperature, and surface-passivation condition. The results identify an interaction between anneal temperature and passivation condition.

The team introduces a control chart for termination leakage, adds a pre-metal surface inspection step, and repeats the electrical and structural characterization on the next three lots. An analytics model is used to rank likely defect sites from wafer maps, but final disposition remains with device engineers.

Outcome

The revised process reduces the high-leakage tail while preserving forward performance. The important result is not the use of one instrument or one algorithm; it is the convergence of spatial electrical data, temperature dependence, structural evidence, TCAD sensitivity analysis, and controlled experimentation.

Learning Points

1. A single electrical curve rarely identifies a unique physical cause.

2. Spatial and temperature-dependent data improve discrimination between hypotheses.

3. TCAD is most valuable when linked to measurable parameters and sensitivity analysis.

4. DoE can reveal interactions missed by isolated parameter sweeps.

5. AI-assisted ranking should support, not replace, engineering judgment and process governance.

Challenges, Limitations, and Professional Practice

The principal challenges are parasitic effects, sample damage, nonuniformity, model non-uniqueness, insufficient calibration, and confounding between process variables. Contact resistance can be mistaken for channel degradation; probe damage can be mistaken for leakage; thermal drift can be mistaken for instability; and an attractive machine-learning correlation can fail after a process change.

Best practice is to maintain a characterization plan linking each measurement to a physical question; use reference structures and golden devices; record environmental and instrument metadata; combine orthogonal methods; distinguish calibration from validation; quantify uncertainty; and retain raw data and analysis code under version control. Reports should identify measurement conditions, extraction definitions, excluded data with reasons, model assumptions, and limits of inference.

Ethically and professionally, engineers must not suppress inconvenient data, overstate model accuracy, or report capability from an unstable process. Semiconductor data may contain proprietary process information and personal information through operator or equipment records; access and retention should therefore follow organizational and regulatory requirements. Decisions affecting product safety require qualified review and auditable evidence.

Current Developments

Emerging characterization systems combine automated probing, high-speed digitization, spectroscopy, microscopy, and thermal mapping. Digital twins increasingly couple process history, TCAD models, equipment data, and measured device behaviour. Physics-informed machine learning seeks to reduce data requirements and improve extrapolation, while active learning selects the next experiment that is expected to provide the greatest information. Quantum and nanoscale devices demand stronger treatment of nonclassical transport, discrete dopants, variability, and cryogenic measurement. These developments increase capability but also make model validation, data provenance, and uncertainty quantification more important.

Chapter Summary

Device characterization is an inference process connecting measured electrical, optical, structural, and thermal observables to physical parameters. Reliable extraction requires controlled measurement, explicit definitions, appropriate models, validation, and uncertainty analysis. TCAD provides a physics-based means of exploring process and device hypotheses, but its predictions depend on geometry, material models, boundary conditions, calibration, and numerical quality. DoE identifies influential factors and interactions more efficiently than uncontrolled experimentation, while statistical process control distinguishes common-cause variation from special causes and assesses capability only after stability is established. Data analytics and AI can support defect detection, virtual metrology, and yield improvement when data leakage, imbalance, drift, explainability, privacy, and human accountability are addressed. The strongest engineering conclusions arise from converging evidence rather than from any single measurement or algorithm.

Key Terms

Measurand: The quantity intended to be measured.

Parameter extraction: Estimation of physical or compact-model parameters from observations.

Ideality factor: A diode parameter describing departure from ideal exponential conduction.

Subthreshold swing: Gate-voltage change required for a decade change in subthreshold current.

Interface-trap density: Density of electrically active states at a material interface.

Four-terminal measurement: Measurement that separates current-carrying and voltage-sensing leads.

Photoluminescence: Light emitted following optical excitation of a material.

Raman spectroscopy: Analysis of inelastic light scattering to study phonons, stress, or temperature.

Thermal impedance: Temperature response per unit power as a function of time.

Uncertainty: Quantified limited knowledge associated with a measurement result.

Identifiability: Ability of data and a model to distinguish one parameter from alternatives.

TCAD: Computer-aided simulation of semiconductor fabrication and device operation.

Calibration: Adjustment of model parameters using selected experimental data.

Validation: Testing a model against independent observations or conditions.

Design of experiments: Structured planning of factor variations to estimate effects efficiently.

Statistical process control: Monitoring process variation using statistical evidence over time.

Process capability: Degree to which a stable process can meet specification limits.

Data leakage: Unintended use of information during model training that inflates evaluation performance.

Virtual metrology: Prediction of a measured process or device quantity from other data.

Physics-informed machine learning: Learning constrained or guided by physical laws or models.

Review Questions

Short Answer Questions

1. Why can a diode ideality factor extracted from an entire semilogarithmic I–V curve be misleading?

2. Distinguish measurement error from measurement uncertainty.

3. What is the purpose of mesh-convergence testing in TCAD?

4. Why are specification limits different from control limits?

5. State two risks of randomly splitting data from the same wafer into training and test sets.

Long Answer Questions

1. Explain an integrated electrical characterization procedure for extracting MOSFET threshold voltage, transconductance, and subthreshold swing, including relevant sources of error.

2. Discuss how optical, structural, and thermal methods complement electrical characterization in identifying device failure mechanisms.

3. Describe a rigorous TCAD calibration and validation workflow and explain why numerical convergence alone is insufficient.

4. Explain the role of DoE, randomization, replication, and interaction effects in semiconductor process development.

5. Discuss responsible use of AI for defect classification and yield analytics in a manufacturing environment.

Analytical / Application Questions

1. A transistor shows reduced on-current in DC tests but normal pulsed on-current. Develop competing physical hypotheses and propose measurements to distinguish self-heating from trapping.

2. A process has a high capability index but repeated control-chart violations. Determine whether capability should be reported and recommend an investigation strategy.

3. A machine-learning model predicts wafer yield accurately on historical lots but fails after a passivation change. Analyze possible causes and design a technically and ethically sound remediation plan.

22. Multiple Choice Questions with Answers

Question 1

Which measurement condition most directly helps separate self-heating from trapping in a power transistor?

A. A longer DC sweep with higher compliance

B. Comparison of controlled pulsed and DC I–V measurements

C. A single room-temperature optical image

D. Reducing the number of measurement points

Correct Answer: B. Comparison of controlled pulsed and DC I–V measurements

Question 2

Why should threshold-voltage extraction state its extraction convention?

A. Threshold voltage has only one universal experimental definition

B. Different conventions can produce different numerical values

C. The convention determines the wafer material composition

D. It eliminates the need for temperature control

Correct Answer: B. Different conventions can produce different numerical values

Question 3

Which statement best describes TCAD validation?

A. Fitting every available data point with additional parameters

B. Checking solver convergence at one bias only

C. Testing calibrated predictions against independent observations

D. Replacing measurements with simulated data

Correct Answer: C. Testing calibrated predictions against independent observations

Question 4

What is the principal danger of using control limits as specification limits?

A. It confuses process stability with product requirements

B. It always increases measurement resolution

C. It prevents detection of special causes

D. It makes all data normally distributed

Correct Answer: A. It confuses process stability with product requirements

Question 5

Which design feature helps estimate experimental variation in a process study?

A. Replication

B. Removing all repeated runs

C. Changing several factors without recording them

D. Using only the best-performing wafer

Correct Answer: A. Replication

Question 6

A model has high accuracy because measurements from the same wafer appear in both training and test sets. This is primarily an example of:

A. Mesh divergence

B. Data leakage

C. Thermal impedance

D. Blocking

Correct Answer: B. Data leakage

Question 7

Which observation most strongly suggests that a measured leakage increase may be spatially related to a process defect?

A. Every die has exactly the same leakage

B. Leakage clusters in a repeatable wafer-map pattern

C. Forward voltage is measured only once

D. The instrument has a current limit

Correct Answer: B. Leakage clusters in a repeatable wafer-map pattern

Question 8

In uncertainty propagation, covariance terms are important when:

A. All inputs are known exactly

B. Measured quantities are statistically related

C. The model contains no parameters

D. The measurement is performed at zero temperature

Correct Answer: B. Measured quantities are statistically related

Question 9

What is the main value of sensitivity analysis in TCAD?

A. It identifies which uncertain parameters most affect the output

B. It guarantees that the physical model is correct

C. It removes the need for calibration

D. It makes the mesh irrelevant

Correct Answer: A. It identifies which uncertain parameters most affect the output

Question 10

Which practice is most appropriate when deploying AI for semiconductor defect disposition?

A. Allowing an unreviewed model to change safety-critical settings

B. Reporting accuracy only and hiding false negatives

C. Maintaining traceability, uncertainty information, and human oversight

D. Deleting anomalous data before model training without investigation

Correct Answer: C. Maintaining traceability, uncertainty information, and human oversight