numerical methods

Every step in a semiconductor process is, underneath, a partial differential equation nobody can solve with pencil and paper. Dopants diffuse, fields settle, plasma etches, light propagates, films grow — each governed by continuous physics on geometries too intricate for closed-form answers. Numerical methods are the bridge that turns those equations into a finite set of arithmetic a computer can grind through, and they are the quiet foundation under every TCAD tool, every optical-proximity-correction run, and every process simulator on this site.\n\n**Three kinds of PDE, three personalities.** Before choosing a method you classify the equation, because its type dictates its behavior and how it must be discretized. Parabolic equations (diffusion, heat) smooth everything out and march forward in time. Elliptic equations (electrostatics, steady-state fields) have no time at all — every point depends on every boundary at once. Hyperbolic equations (wave propagation in lithography) carry information at finite speed along characteristics.\n\n| PDE type | Physics it governs | Canonical equation | Numerical care |\n|---|---|---|---|\n| Parabolic | Dopant diffusion, thermal anneal | $\partial_t C = \nabla\cdot(D\nabla C)$ | Stiff; wants implicit stepping |\n| Elliptic | Poisson, plasma sheaths, device fields | $\nabla^2\phi = -\rho/\varepsilon$ | Global solve; needs good preconditioner |\n| Hyperbolic | Light propagation, acoustics | $\nabla^2 E = \mu\varepsilon\,\partial_{tt}E$ | Respect the CFL speed limit |\n\n**Discretizing space: pick your mesh philosophy.** Finite differences (FDM) replace derivatives with grid-point ratios — trivially simple on a regular grid, awkward on curved geometry. Finite elements (FEM) tile the domain with triangles or tetrahedra and fit local polynomials, which handles complex shapes and adaptive refinement gracefully, at the cost of assembling a global stiffness matrix. Finite volumes (FVM) integrate over little control cells so that mass, charge, and momentum are conserved *exactly* by construction — the natural choice for fluid transport in a CVD reactor. All three convert a continuous PDE into a matrix equation $\mathbf{A}\mathbf{x} = \mathbf{b}$.\n\n```svg\n\n```\n\n**Marching in time is a stability bargain.** Explicit schemes like Forward Euler and RK4 are cheap — each step is pure arithmetic — but a diffusion problem chains them to the CFL condition $\Delta t < \Delta x^2 / 2D$, forcing absurdly small steps on a fine grid. Implicit schemes like Backward Euler and Crank-Nicolson are unconditionally stable and take giant strides, but each step demands solving a linear (often nonlinear) system. For the *stiff* systems typical of dopant diffusion and thermal anneal — where fast and slow physics coexist by ten orders of magnitude — implicit stepping is not a luxury but the only tractable option.\n\n**Solving Ax = b is where the compute actually goes.** For small problems a direct LU factorization is exact and reliable. For the million-to-billion-unknown systems of 3D process and device simulation, iterative Krylov methods win: conjugate gradient (CG) for symmetric positive-definite systems, GMRES and BiCGSTAB for the non-symmetric ones. Their convergence lives or dies on *preconditioning* — transforming the system so it converges in tens of iterations instead of thousands. Incomplete-LU, multigrid, and domain-decomposition preconditioners are the difference between a simulation that finishes overnight and one that never finishes at all.\n\n**When the continuum breaks down, go stochastic.** Some physics is too discrete or too rarefied for PDEs. Particle-in-cell (PIC) with Monte Carlo collisions tracks individual charges to model the plasma in an etch chamber. Direct Simulation Monte Carlo (DSMC) handles rarefied gas flow at high Knudsen number where the Navier-Stokes assumption fails. Kinetic Monte Carlo (KMC) advances atomic-scale events — an adatom hop, a surface reaction — one Poisson-timed event at a time, with the time increment drawn as $\Delta t = -\ln(r)/R_{\text{tot}}$. These methods trade smooth fields for statistical samples, and they are how atomistic reality gets injected into a manufacturing model.\n\n**Moving boundaries need level sets.** Etching and deposition literally move the material surface, changing the domain's topology as trenches merge or voids pinch off. The level-set method represents that surface as the zero contour of a field $\phi$ evolving by $\partial_t\phi + v_n|\nabla\phi| = 0$, so topology changes are handled automatically without remeshing — which is exactly why the feature-scale etch and deposition simulators on this site can watch a profile evolve.\n\n**Read numerical methods through a discretization-and-conditioning lens rather than a formula lens,** and the whole field snaps into focus: the equation's *type* dictates the discretization, the discretization sets the *matrix* structure, and the matrix's *conditioning* decides whether the solve is fast or hopeless. The reduced-order simulators here are exactly this pipeline compressed to run in a browser — trading a full 3D solve for a physically faithful surrogate. How Scharfetter-Gummel discretization keeps device currents stable at high bias, how multigrid achieves linear scaling, and how physics-informed neural networks are beginning to replace parts of the solve are the natural next layers to explore.