Inverse Problems

Inverse Problems

1. Introduction to Inverse Problems

1.1 Mathematical Definition

In mathematical terms, a forward problem is defined as:

where:

• $x$ = input parameters (process conditions)
• $f$ = forward operator (physical model)
• $y$ = output observations (measurements, wafer state)

The inverse problem seeks to find $x$ given $y$:

1.2 Hadamard Well-Posedness Criteria

A problem is well-posed if it satisfies:

1. Existence : A solution exists for all admissible data
2. Uniqueness : The solution is unique
3. Stability : The solution depends continuously on the data

Most semiconductor inverse problems are ill-posed , violating one or more criteria.

1.3 Why Semiconductor Manufacturing Creates Ill-Posed Problems

• Non-uniqueness : Multiple process conditions $\{x_1, x_2, \ldots\}$ can produce indistinguishable outputs within measurement precision
• Sensitivity : Small perturbations in measurements cause large changes in estimated parameters:

• Incomplete information : Not all relevant physical quantities can be measured

2. Lithography Inverse Problems

2.1 Optical Proximity Correction (OPC)

2.1.1 Forward Model

The aerial image intensity at the wafer plane:

where:

• $H(f_x, f_y)$ = optical transfer function (pupil function)
• $M(f_x, f_y)$ = Fourier transform of the mask pattern
• $(f_x, f_y)$ = spatial frequencies

2.1.2 Inverse Problem Formulation

Find mask pattern $M$ that minimizes:

where:

• $T(M)$ = printed pattern from mask $M$
• $D$ = desired (target) pattern
• $R(M)$ = regularization for mask manufacturability
• $\lambda$ = regularization weight

2.1.3 Regularization Terms

Common regularization terms include:

• Mask complexity penalty :

• Minimum feature size constraint :

• Sidelobe suppression :

2.2 Source-Mask Optimization (SMO)

Joint optimization over source shape $S$ and mask $M$:

This is a higher-dimensional inverse problem with:

• Source degrees of freedom: pupil discretization points
• Mask degrees of freedom: pixel-based mask representation
• Coupled nonlinear interactions

2.3 Inverse Lithography Technology (ILT)

Full pixel-based mask optimization using gradient descent:

Gradient computation via adjoint method :

where $\psi$ is the complex field at the wafer plane.

3. Thin Film Metrology Inverse Problems

3.1 Ellipsometry

3.1.1 Measured Quantities

Ellipsometry measures the complex reflectance ratio:

where:

• $r_p$ = p-polarized reflection coefficient
• $r_s$ = s-polarized reflection coefficient
• $\Psi$ = amplitude ratio angle
• $\Delta$ = phase difference

3.1.2 Forward Model (Fresnel Equations)

For a single film on substrate:

where:

• $r_{01}, r_{12}$ = interface Fresnel coefficients
• $\beta = \frac{2\pi d}{\lambda} \tilde{n}_1 \cos\theta_1$ = phase thickness
• $d$ = film thickness
• $\tilde{n}_1 = n_1 + ik_1$ = complex refractive index

3.1.3 Inverse Problem

Given measured $\Psi(\lambda), \Delta(\lambda)$, find:

• Film thickness(es): $d_1, d_2, \ldots$
• Optical constants: $n(\lambda), k(\lambda)$ for each layer

Objective function :

3.2 Scatterometry (Optical Critical Dimension)

3.2.1 Forward Model

Rigorous Coupled-Wave Analysis (RCWA) solves Maxwell's equations for periodic structures:

The grating is represented as Fourier series:

where $G = \frac{2\pi}{\Lambda}$ is the grating vector.

3.2.2 Profile Parameterization

A trapezoidal line profile is characterized by:

• CD (Critical Dimension) : $w$
• Height : $h$
• Sidewall Angle : $\theta_{\text{SWA}}$
• Corner Rounding : $r$
• Footing/Undercut : $\delta$

Parameter vector: $\mathbf{p} = [w, h, \theta_{\text{SWA}}, r, \delta, \ldots]^T$

3.2.3 Inverse Problem

Challenges :

• Non-convex objective with multiple local minima
• Parameter correlations (e.g., height vs. refractive index)
• Sensitivity varies dramatically across parameters

4. Plasma Etch Inverse Problems

4.1 Etch Rate Modeling

4.1.1 Ion-Enhanced Etching Model

where:

• $\Gamma_{\text{ion}}$ = ion flux
• $\Gamma_{\text{neutral}}$ = neutral radical flux
• $E_a$ = activation energy
• $a, b$ = reaction orders

4.1.2 Aspect Ratio Dependent Etching (ARDE)

Etch rate in high-aspect-ratio features:

where $AR = \frac{\text{depth}}{\text{width}}$ is the aspect ratio.

4.2 Profile Reconstruction from OES

4.2.1 Optical Emission Spectroscopy Model

Emission intensity for species $j$:

where:

• $n_e$ = electron density
• $n_j$ = species density
• $\langle \sigma v \rangle$ = rate coefficient for excitation

4.2.2 Inverse Problem

From observed $I_j(t)$ time traces, determine:

• Etch front position $z(t)$
• Layer interfaces
• Process endpoint

State estimation formulation :

5. Ion Implantation Inverse Problems

5.1 As-Implanted Profile

5.1.1 LSS Theory (Lindhard-Scharff-Schiøtt)

The implanted concentration profile:

where:

• $\Phi$ = implant dose (ions/cm²)
• $R_p$ = projected range
• $\Delta R_p$ = straggle (standard deviation)

5.1.2 Dual-Pearson for Channeling

For crystalline substrates with channeling:

where $P_i$ are Pearson IV distributions and $f$ is the channeled fraction.

5.2 Diffusion Inversion

5.2.1 Fick's Second Law with Concentration Dependence

For dopants like boron:

5.2.2 Inverse Problem

Given final SIMS profile $C_{\text{final}}(x)$, find:

• Initial implant conditions: $\Phi, E$ (energy)
• Anneal conditions: $T(t)$, time $t_a$
• Diffusion parameters: $D_i^*, \beta_1, \beta_2$

Regularized formulation :

6. Deposition Inverse Problems

6.1 CVD Step Coverage

6.1.1 Thiele Modulus

Conformality characterized by:

where:

• $L$ = feature depth
• $k_s$ = surface reaction rate
• $D_{\text{Kn}}$ = Knudsen diffusion coefficient

Step coverage:

6.1.2 Inverse Problem

Given target step coverage $SC_{\text{target}}$, find:

• Pressure $P$
• Temperature $T$
• Precursor partial pressures
• Carrier gas flow

6.2 ALD Thickness Control

6.2.1 Growth Per Cycle (GPC)

where:

• $\Theta_{\text{sat}}$ = saturation coverage (0 to 1)
• $d_{\text{ML}}$ = monolayer thickness

6.2.2 Inverse Problem

For target thickness $d$:

Optimize $(T, t_{\text{pulse}}, t_{\text{purge}})$ for throughput and uniformity.

7. CMP Inverse Problems

7.1 Preston Equation

Material removal rate:

where:

• $K_p$ = Preston coefficient
• $P$ = applied pressure
• $V$ = relative velocity

7.2 Pattern Density Effects

7.2.1 Effective Density Model

Local removal rate depends on pattern density $\rho$:

where $\eta$ is the selectivity ratio.

7.2.2 Dishing and Erosion

• Dishing (over-polish of metal in trench):

• Erosion (over-polish of dielectric):

7.3 Inverse Problem

Given target post-CMP topography, find:

• Polish time
• Pressure profile (zone control)
• Slurry chemistry
• Potentially: design rule modifications for pattern density

8. TCAD Parameter Extraction

8.1 Device Model

MOSFET drain current:

8.2 Inverse Problem Formulation

Given measured $I_D(V_{GS}, V_{DS})$ characteristics, extract:

• $V_{th}$ = threshold voltage
• $\mu_{\text{eff}}$ = effective mobility
• $L_{\text{eff}}$ = effective channel length
• $\lambda$ = channel length modulation

Optimization :

8.3 Interface Trap Density from C-V

From measured capacitance $C(V_G)$:

where $C_s$ is the semiconductor capacitance.

9. Mathematical Solution Approaches

9.1 Regularization Methods

9.1.1 Tikhonov Regularization

Closed-form solution:

9.1.2 Total Variation Regularization

Preserves edges while smoothing noise.

9.1.3 L1 Regularization (LASSO)

Promotes sparse solutions.

9.2 Bayesian Inference

9.2.1 Posterior Distribution

By Bayes' theorem:

where:

• $p(y|x)$ = likelihood
• $p(x)$ = prior
• $p(x|y)$ = posterior

9.2.2 Maximum A Posteriori (MAP) Estimate

For Gaussian likelihood and prior:

This recovers Tikhonov regularization with $\lambda = \frac{\sigma_n^2}{\sigma_x^2}$.

9.3 Adjoint Methods for Gradient Computation

For objective $\mathcal{L}(x) = \|F(x) - y\|^2$ with expensive forward model $F$:

Forward solve :

Adjoint solve :

Gradient :

Computational cost: $O(1)$ forward + adjoint solves regardless of parameter dimension.

9.4 Machine Learning Approaches

9.4.1 Neural Network Surrogate Models

Train $\hat{F}_\theta(x) \approx F(x)$:

Then use $\hat{F}_\theta$ for fast inverse optimization.

9.4.2 Physics-Informed Neural Networks (PINNs)

Loss function includes physics residual:

where:

for PDE operator $\mathcal{N}$.

10. Key Challenges and Considerations

10.1 Non-Uniqueness

• Definition : Multiple solutions $\{x_1, x_2, \ldots\}$ satisfy $\|F(x_i) - y\| < \epsilon$
• Mitigation : Additional measurements, physical constraints, regularization
• Quantification : Null space analysis, condition number $\kappa(A) = \frac{\sigma_{\max}}{\sigma_{\min}}$

10.2 High Dimensionality

• Parameter space : $\dim(x) \sim 10^2$ to $10^6$ (e.g., ILT masks)
• Curse of dimensionality : Sampling density scales as $N^d$
• Approaches : Dimensionality reduction, sparse representations, hierarchical models

10.3 Computational Cost

• Forward model cost : RCWA: $O(N^3)$ per wavelength; TCAD: hours for full 3D
• Inverse iterations : Typically $10^2$ to $10^4$ forward evaluations
• Mitigation : Surrogate models, multi-fidelity methods, parallel computing

10.4 Model Uncertainty

• Sources : Unmodeled physics, parameter drift, measurement bias
• Impact : Inverse solution may fit model but not reality
• Approaches : Model calibration, uncertainty propagation, robust optimization

11. Emerging Directions

11.1 Digital Twins

• Real-time state estimation combining physics models with sensor data
• Kalman filtering for dynamic process tracking:

11.2 Multi-Fidelity Methods

• Hierarchy of models: analytical → reduced-order → full numerical
• Efficient exploration with cheap models, refinement with expensive ones
• Multi-fidelity Gaussian processes for Bayesian optimization

11.3 Uncertainty Quantification

• Full posterior distributions, not just point estimates
• Sensitivity analysis: which measurements reduce uncertainty most?
• Propagation to downstream process steps and device performance

11.4 End-to-End Differentiable Simulation

• Automatic differentiation through entire process flow
• Enables gradient-based optimization across traditionally separate steps
• Requires differentiable forward models

12. Summary