Files

T

Antigravity Agent f8b21af4be Wiki cleanup: error-doc removal, dedup merge, link normalization

10_Wiki/Topics 대규모 정리:
- 오류 캡처/미완성 stub 문서 227개 제거
- 교차폴더 중복 43클러스터 병합 (63파일 → redirect)
- 링크명 정규화: 깨진 링크 수정·redirect 직결·개념 매핑 ~2,400건
- 카테고리 MOC 6개 신규 생성
- Graph 섹션 미해결 related-keyword 링크 10,058건 제거

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-20 23:52:15 +09:00

6.3 KiB

Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

title

Partial Differential Equations

매 한 줄

"매 multivariable function 의 partial derivative relations". 매 PDE는 매 fluid (Navier-Stokes), heat, wave, elasticity, EM (Maxwell), QM (Schrödinger), finance (Black-Scholes) 의 universal language, 매 2026 numerical solving은 매 FDM/FEM/FVM/spectral + PINN/Neural Operator (FNO, DeepONet) 의 hybrid 시대.

매 핵심

매 분류 (2nd order linear)

B² - 4AC 로:
- Elliptic (<0): Laplace ∇²u=0, Poisson — equilibrium.
- Parabolic (=0): heat uₜ = α∇²u — diffusion.
- Hyperbolic (>0): wave uₜₜ = c²∇²u — propagation.

매 well-posed (Hadamard)

Existence, uniqueness, continuous dependence on data.
Boundary conditions: Dirichlet, Neumann, Robin.

매 numerical methods

FDM: structured grid, easy, low geometry flexibility.
FEM: weak form, complex geometry, h/p refinement.
FVM: conservation laws (CFD).
Spectral: smooth solutions, exponential convergence.
PINN (2026): NN minimizing PDE residual, mesh-free, inverse problems.
Neural Operator: FNO/DeepONet learn solution operator.

매 응용

CFD (aerospace, weather).
Heat transfer / thermal analysis.
Structural mechanics.
EM simulation (CST, COMSOL).
Option pricing (Black-Scholes PDE).
Diffusion models (LLM/image gen with score-PDE).

💻 패턴

1D Heat Equation — explicit FDM

import numpy as np
def heat_explicit(u0, alpha, dx, dt, T):
    r = alpha*dt/dx**2
    assert r <= 0.5, "CFL violated"
    u = u0.copy(); steps = int(T/dt)
    for _ in range(steps):
        u[1:-1] = u[1:-1] + r*(u[2:] - 2*u[1:-1] + u[:-2])
    return u

Crank-Nicolson (implicit, 2nd order)

from scipy.sparse import diags
from scipy.sparse.linalg import spsolve
def crank_nicolson(u0, alpha, dx, dt, T):
    n = len(u0); r = alpha*dt/(2*dx**2)
    A = diags([-r, 1+2*r, -r], [-1,0,1], shape=(n-2, n-2)).tocsc()
    B = diags([ r, 1-2*r,  r], [-1,0,1], shape=(n-2, n-2))
    u = u0.copy()
    for _ in range(int(T/dt)):
        u[1:-1] = spsolve(A, B @ u[1:-1])
    return u

2D Poisson via FEM (FEniCSx)

from dolfinx import mesh, fem
from ufl import TrialFunction, TestFunction, dx, grad, inner
import numpy as np
domain = mesh.create_unit_square(MPI.COMM_WORLD, 64, 64)
V = fem.FunctionSpace(domain, ("Lagrange", 1))
u, v = TrialFunction(V), TestFunction(V)
f = fem.Constant(domain, 1.0)
a = inner(grad(u), grad(v))*dx
L = f*v*dx
bc = fem.dirichletbc(0.0, ..., V)
problem = fem.petsc.LinearProblem(a, L, bcs=[bc])
uh = problem.solve()

1D Wave — leapfrog

def wave_leapfrog(u0, v0, c, dx, dt, T):
    r = c*dt/dx
    assert r <= 1, "CFL"
    u_prev = u0.copy()
    u = u0 + dt*v0  # half-step init
    steps = int(T/dt)
    for _ in range(steps):
        u_next = 2*u[1:-1] - u_prev[1:-1] + r**2*(u[2:] - 2*u[1:-1] + u[:-2])
        u_prev[1:-1] = u[1:-1]; u[1:-1] = u_next
    return u

PINN for Burgers' equation

import torch, torch.nn as nn
class PINN(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(2,64), nn.Tanh(), nn.Linear(64,64), nn.Tanh(),
            nn.Linear(64,64), nn.Tanh(), nn.Linear(64,1))
    def forward(self, xt): return self.net(xt)

def pde_residual(model, xt, nu=0.01/np.pi):
    xt.requires_grad_(True)
    u = model(xt)
    grads = torch.autograd.grad(u, xt, torch.ones_like(u), create_graph=True)[0]
    u_x, u_t = grads[:,0:1], grads[:,1:2]
    u_xx = torch.autograd.grad(u_x, xt, torch.ones_like(u_x), create_graph=True)[0][:,0:1]
    return u_t + u*u_x - nu*u_xx
# Loss = MSE(pde_residual) + MSE(IC) + MSE(BC); Adam optimize.

Fourier Neural Operator (FNO, 2026)

import torch, torch.nn as nn
class SpectralConv1d(nn.Module):
    def __init__(self, in_ch, out_ch, modes):
        super().__init__()
        self.modes = modes
        self.weight = nn.Parameter(torch.randn(in_ch, out_ch, modes, dtype=torch.cfloat)*0.02)
    def forward(self, x):  # x: (B, C, N)
        N = x.size(-1)
        x_ft = torch.fft.rfft(x)
        out_ft = torch.zeros(x.size(0), self.weight.size(1), N//2+1, dtype=torch.cfloat, device=x.device)
        out_ft[..., :self.modes] = torch.einsum("bcm,com->bom", x_ft[..., :self.modes], self.weight)
        return torch.fft.irfft(out_ft, n=N)

매 결정 기준

상황	Approach
Simple geometry, structured	FDM
Complex geometry / multi-physics	FEM
Conservation laws, shocks	FVM (CFD)
Smooth, periodic	Spectral / pseudo-spectral
Inverse / sparse data	PINN
Many similar PDEs (parametric)	Neural Operator (FNO/DeepONet)
Stochastic / high-dim	Deep BSDE / Monte Carlo

기본값: classical solving은 FEM (FEniCSx) or FVM (OpenFOAM); ML-side는 FNO; inverse problem은 PINN.

🔗 Graph

Adjacent: PINN · Diffusion-Models · Finite-Element-Method

🤖 LLM 활용

언제: PDE classification 설명, BC formulation help, weak form derivation, PINN architecture suggestion. 언제 X: actual numerical solving (FEniCSx/JAX/OpenFOAM 매 use).

❌ 안티패턴

CFL violation: explicit scheme에 dt 매 too large → blow up.
PINN as universal: PINN 매 hard problems 에 매 종종 fail (high-freq/turbulent) — 매 classical FEM 매 첫 baseline.
No mesh convergence study: 매 must show error vs h/p refinement.
Wrong BC: Neumann ↔ Dirichlet mistake → 매 entire solution wrong.
Ignoring stability: implicit ≠ unconditionally accurate (just stable).

🧪 검증 / 중복

Verified (Strikwerda "Finite Difference Schemes", Brenner & Scott "FEM", Karniadakis et al PINN review).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — classical methods + PINN/FNO (2026)

6.3 KiB Raw Blame History