Inverse Problems

Keywords: inverse problems,inverse problem,ill-posed problems,regularization,parameter estimation,OPC,scatterometry,virtual metrology

Inverse Problems

1. Introduction to Inverse Problems

1.1 Mathematical Definition

In mathematical terms, a forward problem is defined as:

$$ y = f(x) $$

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 $y$:

$$ x = f^{-1}(y) $$

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:

$$ \|x_1 - x_2\| \gg \|y_1 - y_2\| $$

• 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:

$$ I(x, y) = \left| \int \int H(f_x, f_y) \cdot M(f_x, f_y) \cdot e^{i2\pi(f_x x + f_y y)} \, df_x \, df_y \right|^2 $$

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:

$$ \mathcal{L}(M) = \|T(M) - D\|^2 + \lambda R(M) $$

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 :

$$ R_{\text{complexity}}(M) = \int | abla M|^2 \, dA $$

• Minimum feature size constraint :

$$ R_{\text{MFS}}(M) = \sum_i \max(0, w_{\min} - w_i)^2 $$

• Sidelobe suppression :

$$ R_{\text{SRAF}}(M) = \int_{\Omega_{\text{dark}}} I(x,y)^2 \, dA $$

2.2 Source-Mask Optimization (SMO)

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

$$ \min_{S, M} \|T(S, M) - D\|^2 + \lambda_1 R_S(S) + \lambda_2 R_M(M) $$

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:

$$ M^{(k+1)} = M^{(k)} - \alpha abla_M \mathcal{L}(M^{(k)}) $$

Gradient computation via adjoint method :

$$

abla_M \mathcal{L} = \text{Re}\left\{ \mathcal{F}^{-1}\left[ H^ \cdot \mathcal{F}\left[ \frac{\partial \mathcal{L}}{\partial I} \cdot \psi^ \right] \right] \right\} $$

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:

$$ \rho = \frac{r_p}{r_s} = \tan(\Psi) \cdot e^{i\Delta} $$

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:

$$ r_{012} = \frac{r_{01} + r_{12} e^{-i2\beta}}{1 + r_{01} r_{12} e^{-i2\beta}} $$

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 :

$$ \chi^2 = \sum_{\lambda} \left[ \left(\frac{\Psi_{\text{meas}} - \Psi_{\text{calc}}}{\sigma_\Psi}\right)^2 + \left(\frac{\Delta_{\text{meas}} - \Delta_{\text{calc}}}{\sigma_\Delta}\right)^2 \right] $$

3.2 Scatterometry (Optical Critical Dimension)

3.2.1 Forward Model

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

$$

abla \times \mathbf{E} = -\frac{\partial \mathbf{B}}{\partial t}, \quad abla \times \mathbf{H} = \frac{\partial \mathbf{D}}{\partial t} $$

The grating is represented as Fourier series:

$$ \varepsilon(x, z) = \sum_m \varepsilon_m(z) e^{imGx} $$

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

$$ \hat{\mathbf{p}} = \arg\min_{\mathbf{p}} \sum_{\lambda, \theta} \left( R_{\text{meas}}(\lambda, \theta) - R_{\text{RCWA}}(\lambda, \theta; \mathbf{p}) \right)^2 $$

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

$$ \text{ER} = k_0 \cdot \Gamma_{\text{ion}}^a \cdot \Gamma_{\text{neutral}}^b \cdot \exp\left(-\frac{E_a}{k_B T}\right) $$

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:

$$ \text{ER}(AR) = \text{ER}_0 \cdot \frac{1}{1 + \alpha \cdot AR^\beta} $$

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 $j$:

$$ I_j(\lambda) = A_j \cdot n_e \cdot n_j \cdot \langle \sigma v \rangle_{j}^{\text{exc}} $$

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 :

$$ \hat{z}(t) = \arg\min_{z} \|I_{\text{obs}}(t) - I_{\text{model}}(z, t)\|^2 + \lambda \left\|\frac{dz}{dt}\right\|^2 $$

5. Ion Implantation Inverse Problems

5.1 As-Implanted Profile

5.1.1 LSS Theory (Lindhard-Scharff-Schiøtt)

The implanted concentration profile:

$$ C(x) = \frac{\Phi}{\sqrt{2\pi} \Delta R_p} \exp\left[-\frac{(x - R_p)^2}{2(\Delta R_p)^2}\right] $$

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:

$$ C(x) = (1-f) \cdot P_1(x; R_{p1}, \Delta R_{p1}, \gamma_1, \beta_1) + f \cdot P_2(x; R_{p2}, \Delta R_{p2}, \gamma_2, \beta_2) $$

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

$$ \frac{\partial C}{\partial t} = \frac{\partial}{\partial x}\left[D(C) \frac{\partial C}{\partial x}\right] $$

For dopants like boron:

$$ D(C) = D_i^* \left[1 + \beta_1 \left(\frac{C}{n_i}\right) + \beta_2 \left(\frac{C}{n_i}\right)^2\right] $$

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 :

$$ \min_{\theta} \|C_{\text{SIMS}} - C_{\text{simulated}}(\theta)\|^2 + \lambda \|\theta - \theta_{\text{prior}}\|^2 $$

6. Deposition Inverse Problems

6.1 CVD Step Coverage

6.1.1 Thiele Modulus

Conformality characterized by:

$$ \phi = L \sqrt{\frac{k_s}{D_{\text{Kn}}}} $$

where:

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

Step coverage:

$$ SC = \frac{1}{\cosh(\phi)} $$

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)

$$ \text{GPC} = \Theta_{\text{sat}} \cdot d_{\text{ML}} $$

where:

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

6.2.2 Inverse Problem

For target thickness $d$:

$$ N_{\text{cycles}} = \left\lceil \frac{d}{\text{GPC}(T, t_{\text{pulse}}, t_{\text{purge}})} \right\rceil $$

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

7. CMP Inverse Problems

7.1 Preston Equation

Material removal rate:

$$ \text{MRR} = K_p \cdot P \cdot V $$

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 $\rho$:

$$ \text{MRR}_{\text{local}} = \frac{\text{MRR}_{\text{blanket}}}{\rho + (1-\rho) \cdot \eta} $$

where $\eta$ is the selectivity ratio.

7.2.2 Dishing and Erosion

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

$$ D = K_d \cdot w \cdot t_{\text{over}} $$

• Erosion (over-polish of dielectric):

$$ E = K_e \cdot \rho \cdot t_{\text{over}} $$

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:

$$ I_D = \mu_{\text{eff}} C_{\text{ox}} \frac{W}{L} \left[(V_{GS} - V_{th})V_{DS} - \frac{V_{DS}^2}{2}\right] (1 + \lambda V_{DS}) $$

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 :

$$ \min_{\theta} \sum_{i,j} \left( I_{D,\text{meas}}(V_{GS,i}, V_{DS,j}) - I_{D,\text{model}}(V_{GS,i}, V_{DS,j}; \theta) \right)^2 $$

8.3 Interface Trap Density from C-V

From measured capacitance $C(V_G)$:

$$ D_{it}(E) = \frac{1}{qA}\left(\frac{1}{C_{\text{meas}}} - \frac{1}{C_{\text{ox}}}\right)^{-1} - \frac{C_s}{qA} $$

where $C_s$ is the semiconductor capacitance.

9. Mathematical Solution Approaches

9.1 Regularization Methods

9.1.1 Tikhonov Regularization

$$ \hat{x} = \arg\min_x \|Ax - y\|^2 + \lambda\|Lx\|^2 $$

Closed-form solution:

$$ \hat{x} = (A^T A + \lambda L^T L)^{-1} A^T y $$

9.1.2 Total Variation Regularization

$$ \min_x \|Ax - y\|^2 + \lambda \int | abla x| \, dA $$

Preserves edges while smoothing noise.

9.1.3 L1 Regularization (LASSO)

$$ \min_x \|Ax - y\|^2 + \lambda\|x\|_1 $$

Promotes sparse solutions.

9.2 Bayesian Inference

9.2.1 Posterior Distribution

By Bayes' theorem:

$$ p(x|y) = \frac{p(y|x) \cdot p(x)}{p(y)} \propto p(y|x) \cdot p(x) $$

where:

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

9.2.2 Maximum A Posteriori (MAP) Estimate

$$ \hat{x}_{\text{MAP}} = \arg\max_x p(x|y) = \arg\max_x [\log p(y|x) + \log p(x)] $$

For Gaussian likelihood and prior:

$$ \hat{x}_{\text{MAP}} = \arg\min_x \left[\frac{\|y - Ax\|^2}{2\sigma_n^2} + \frac{\|x - x_0\|^2}{2\sigma_x^2}\right] $$

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 :

$$ F(x) = y_{\text{sim}} $$

Adjoint solve :

$$ \left(\frac{\partial F}{\partial u}\right)^T \lambda = \frac{\partial \mathcal{L}}{\partial u} $$

Gradient :

$$

abla_x \mathcal{L} = \left(\frac{\partial F}{\partial x}\right)^T \lambda $$

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

9.4 Machine Learning Approaches

9.4.1 Neural Network Surrogate Models

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

$$ \theta^* = \arg\min_\theta \sum_i \|F(x_i) - \hat{F}_\theta(x_i)\|^2 $$

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

9.4.2 Physics-Informed Neural Networks (PINNs)

Loss function includes physics residual:

$$ \mathcal{L} = \mathcal{L}_{\text{data}} + \lambda_{\text{PDE}} \mathcal{L}_{\text{PDE}} + \lambda_{\text{BC}} \mathcal{L}_{\text{BC}} $$

where:

$$ \mathcal{L}_{\text{PDE}} = \left\|\mathcal{N}[u_\theta(x,t)]\right\|^2 $$

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:

$$ \hat{x}_{k|k} = \hat{x}_{k|k-1} + K_k(y_k - H\hat{x}_{k|k-1}) $$

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

Source: ChipFoundryServices — Search this topic — Ask CFSGPT