16 KiB

Raw Blame History

id	title	section	difficulty	estimated_time	prerequisites	objectives	tags
model-05	Resistance Optimization Methods	Advanced Modeling	advanced	45	[model-03 model-04]	[Master iterative resistance optimization algorithm with damping Apply position-dependent physical bounds to resistance values Understand circuit-determined resistance as simplified alternative Validate total resistance ranges and convergence behavior]	[optimization resistance iterative-algorithm convergence validation]

Resistance Optimization Methods

This lesson covers methods for determining the resistance values R[i] for each segment in a distributed spark model. We present two approaches: a rigorous iterative optimization and a simplified circuit-determined method.

The Optimization Problem

Goal and Challenges

Objective:

Find R[i] for i = 1 to n that maximizes total power dissipation:

  P_total = Σᵢ P[i]
  where P[i] = 0.5 × |I[i]|² × R[i]

Subject to physical constraints:
  R_min[i] ≤ R[i] ≤ R_max[i]

Challenge: Coupled optimization

Changing R[j] affects current in segment i:
  - Alters network impedance
  - Changes voltage distribution
  - Modifies all currents I[1], I[2], ..., I[n]

Cannot optimize each R[i] independently!
Must use iterative approach

Computational complexity:

For each R[i]:
  - Sweep through candidate values (20-50 points)
  - Run SPICE AC analysis for each
  - Calculate power P[i]

Total simulations per iteration: n × n_sweep
  n = 10, n_sweep = 20: 200 simulations
  Iterations: 5-10 typical
  Total: 1000-2000 AC analyses

Compare to lumped: 1 analysis
Trade-off: Accuracy vs computational cost

Iterative Optimization Algorithm

Initialization: Tapered Profile

Physical expectation:

Base: Hot, well-coupled → low R
Tip: Cool, weakly-coupled → high R

Monotonically increasing R[i] from base to tip

Initialize with gradient:

# Pseudocode
R_base = 10e3   # 10 kΩ (hot leader)
R_tip = 1e6     # 1 MΩ (cool streamer)

for i in range(1, n+1):
    position = (i-1) / (n-1)  # 0 at base (i=1), 1 at tip (i=n)
    R[i] = R_base + (R_tip - R_base) * position**2

Why quadratic taper?

Linear: R[i] = R_base + (R_tip - R_base) × position
  - Simple, but too gradual
  - Doesn't capture rapid rise near tip

Quadratic: position**2
  - Gentle rise at base
  - Steeper rise near tip
  - Better matches physics

Exponential: also valid, similar results

Example: n=5, R_base=10k, R_tip=1M

i=1: position=0.00 → R[1] = 10 + (1000-10)×0.00 = 10 kΩ
i=2: position=0.25 → R[2] = 10 + 990×0.0625 = 72 kΩ
i=3: position=0.50 → R[3] = 10 + 990×0.25 = 258 kΩ
i=4: position=0.75 → R[4] = 10 + 990×0.5625 = 567 kΩ
i=5: position=1.00 → R[5] = 10 + 990×1.0 = 1000 kΩ

Position-Dependent Bounds

Physical limits vary with position:

Minimum resistance R_min[i]:

Base can achieve low R (hot, well-coupled)
Tip unlikely to reach very low R (cool, weak coupling)

Formula:
  position = (i-1) / (n-1)
  R_min[i] = 1 kΩ + (10 kΩ - 1 kΩ) × position
           = 1 kΩ at base → 10 kΩ at tip

Maximum resistance R_max[i]:

Base unlikely to reach very high R (good power coupling)
Tip can reach very high R (streamer regime)

Formula:
  position = (i-1) / (n-1)
  R_max[i] = 100 kΩ + (100 MΩ - 100 kΩ) × position²
           = 100 kΩ at base → 100 MΩ at tip

Example bounds: n=5

Segment 1 (base, pos=0.00):
  R_min[1] = 1.0 kΩ
  R_max[1] = 100 kΩ

Segment 2 (pos=0.25):
  R_min[2] = 3.25 kΩ
  R_max[2] = 6.3 MΩ

Segment 3 (pos=0.50):
  R_min[3] = 5.5 kΩ
  R_max[3] = 25.1 MΩ

Segment 4 (pos=0.75):
  R_min[4] = 7.75 kΩ
  R_max[4] = 56.3 MΩ

Segment 5 (tip, pos=1.00):
  R_min[5] = 10.0 kΩ
  R_max[5] = 100 MΩ

Rationale:

1. Prevents unphysical solutions
   - Tip with R = 1 kΩ: impossible (not hot enough)
   - Base with R = 100 MΩ: impossible (too much power)

2. Guides optimization
   - Narrows search space
   - Faster convergence
   - More stable

3. Based on physics
   - Leader regime: R ∝ 1/T, T high at base
   - Streamer regime: R very high, weakly coupled

Iterative Loop with Damping

Algorithm structure:

# Pseudocode: Iterative resistance optimization

# Initialize
R = initialize_tapered_profile(n, R_base, R_tip)
alpha = 0.3  # Damping factor
max_iterations = 20
tolerance = 0.01  # 1% convergence threshold

for iteration in range(max_iterations):
    R_old = R.copy()

    for i in range(1, n+1):
        # Sweep R[i] while keeping other R[j] (j≠i) fixed
        R_test = logspace(R_min[i], R_max[i], 20)  # 20 test points
        P_test = []

        for R_candidate in R_test:
            R[i] = R_candidate
            # Run SPICE AC analysis
            results = run_spice_ac(R, C_matrix, freq)
            I_i = results.current[i]
            P_i = 0.5 * abs(I_i)**2 * R_candidate
            P_test.append(P_i)

        # Find R that maximizes power in segment i
        idx_max = argmax(P_test)
        R_optimal[i] = R_test[idx_max]

        # Apply damping for stability
        R_new[i] = alpha * R_optimal[i] + (1 - alpha) * R_old[i]

        # Clip to physical bounds
        R[i] = clip(R_new[i], R_min[i], R_max[i])

    # Check convergence
    max_change = max(abs(R[i] - R_old[i]) / R_old[i] for i in range(1,n+1))
    print(f"Iteration {iteration}: max change = {max_change*100:.2f}%")

    if max_change < tolerance:
        print("Converged!")
        break

# Final result: optimized R[1], R[2], ..., R[n]

Key components:

1. Logarithmic sweep:

R_test = logspace(log10(R_min), log10(R_max), 20)

Why logarithmic?
  - R varies over orders of magnitude (1k to 100M)
  - Linear spacing: wastes points at low end
  - Log spacing: uniform coverage across decades

2. Power calculation:

P[i] = 0.5 × |I[i]|² × R[i]

AC steady-state: Factor of 0.5 for sinusoidal
RMS values: P = I_rms² × R (without 0.5)

Maximize power in segment i, not total power
  - Each segment optimized to extract maximum power
  - Self-consistent with hungry streamer physics

3. Damping factor α:

R_new[i] = α × R_optimal[i] + (1-α) × R_old[i]

α = 0.3 to 0.5 typical

Lower α (e.g., 0.2):
  - More stable (smaller steps)
  - Slower convergence (more iterations)
  - Use if oscillations occur

Higher α (e.g., 0.7):
  - Faster convergence (larger steps)
  - Risk of oscillation (overshooting)
  - Use if convergence slow

Start with α = 0.3, adjust if needed

4. Clipping:

R[i] = clip(R_new[i], R_min[i], R_max[i])

Ensures R stays within physical bounds
Prevents optimizer from exploring unphysical regions

Convergence Behavior

Well-coupled base segments:

Power curve P[i](R[i]) has sharp peak

Example: Segment 1
  R = 10k: P[1] = 5.2 kW
  R = 20k: P[1] = 8.1 kW
  R = 30k: P[1] = 9.4 kW ← maximum (sharp peak)
  R = 40k: P[1] = 8.9 kW
  R = 50k: P[1] = 7.8 kW

Characteristics:
  - Clear optimal R
  - Fast convergence (2-3 iterations)
  - Stable solution

Weakly-coupled tip segments:

Power curve P[i](R[i]) is FLAT

Example: Segment 5 (tip)
  R = 100k: P[5] = 0.82 kW
  R = 500k: P[5] = 0.85 kW
  R = 1M:   P[5] = 0.83 kW
  R = 5M:   P[5] = 0.81 kW

All values give similar power!

Characteristics:
  - Optimal R poorly defined
  - Slow/no convergence to unique value
  - May oscillate between similar R values
  - Physical: weak coupling, low power anyway

Convergence criteria:

Base segments: Converge quickly (<5 iterations)
Middle segments: Moderate convergence (5-10 iterations)
Tip segments: May not converge fully

Solution:
  - Allow tip segments to remain at reasonable values
  - Check that change <5% for tip segments
  - Focus convergence on base/middle (where most power is)

Expected final distribution:

At 200 kHz, 2 m spark:

R[1] ≈ 5-20 kΩ    (base leader)
R[2] ≈ 10-40 kΩ
R[3] ≈ 20-80 kΩ
...
R[n-1] ≈ 50-200 kΩ
R[n] ≈ 100 kΩ - 10 MΩ  (tip streamer, wide range)

Total: R_total = Σ R[i] should be 50-500 kΩ at 200 kHz

Worked Example: n=3 Iterative Optimization

Given:

3 segments, f = 200 kHz
Capacitance matrix from FEMM (simplified example)
Initial: R[1]=50k, R[2]=100k, R[3]=500k
Damping: α = 0.4

Iteration 1:

Optimize R[1] (keeping R[2]=100k, R[3]=500k fixed)

Sweep R[1] in [1k, 100k] (20 points, log scale)

Results (example):
  R[1]=10k → P[1]=5.2 kW
  R[1]=20k → P[1]=8.1 kW
  R[1]=30k → P[1]=9.4 kW ← maximum
  R[1]=40k → P[1]=8.9 kW
  R[1]=50k → P[1]=7.8 kW (current value)
  ...

R_optimal[1] = 30 kΩ

Apply damping:

R_new[1] = 0.4 × 30k + 0.6 × 50k
         = 12k + 30k
         = 42 kΩ

Check bounds: 1k < 42k < 100k ✓
Update: R[1] = 42 kΩ

Optimize R[2] (with R[1]=42k, R[3]=500k)

Sweep R[2], find maximum at R_optimal[2] = 60 kΩ

Current: R[2] = 100 kΩ

R_new[2] = 0.4 × 60k + 0.6 × 100k
         = 24k + 60k
         = 84 kΩ

Update: R[2] = 84 kΩ

Optimize R[3] (with R[1]=42k, R[2]=84k)

Sweep R[3], power curve is FLAT:
  R[3]=200k → P[3]=0.80 kW
  R[3]=500k → P[3]=0.85 kW
  R[3]=1M   → P[3]=0.83 kW

Maximum at 500k, but very weak peak (±5%)
Tip segment: poorly coupled

R_optimal[3] = 500 kΩ (no change)
R_new[3] = 0.4 × 500k + 0.6 × 500k = 500 kΩ

Update: R[3] = 500 kΩ

Convergence check:

Changes:
  R[1]: 50k → 42k  (change = -16%)
  R[2]: 100k → 84k (change = -16%)
  R[3]: 500k → 500k (change = 0%)

Max change = 16% > 1% tolerance
→ Not converged, continue

Iteration 2:

Repeat with new R values... Typically base segments converge within 3-5 iterations

Final result (example):

After 5 iterations:

R[1] = 35 kΩ   (converged, change <1%)
R[2] = 75 kΩ   (converged, change <1%)
R[3] = 500 kΩ  (tip, flat curve, acceptable)

Total: 610 kΩ at 200 kHz
✓ Within expected range (50-500 kΩ, high end due to tip)

Simplified Method: Circuit-Determined Resistance

Key Insight

Hungry streamer physics:

Plasma adjusts diameter to seek R_opt_power for maximum power
R_opt = 1 / (ω × C_total)

For each segment:
  - Segment sees total capacitance C_total[i]
  - Adjusts to R[i] = 1 / (ω × C_total[i])
  - Self-consistent with power optimization

Capacitance weakly depends on diameter:

C ∝ 1 / ln(h/d)

Logarithmic dependence:
  - 2× diameter → ~10% capacitance change
  - R_opt also changes ~10%
  - Small error from assuming fixed C

Formula

For each segment i:

C_total[i] = Σⱼ₌₀ⁿ |C[i,j]|

Sum of absolute values of all capacitances involving segment i

Then:
  R[i] = 1 / (ω × C_total[i])
  R[i] = clip(R[i], R_min[i], R_max[i])

Why this works:

1. Hungry streamer: Seeks R_opt = 1/(ωC)
2. Diameter self-adjusts: Matches R to C
3. Logarithmic C(d): Error ~10-15% (small)
4. Other uncertainties: FEMM ±5-10%, physics ±30-50%
5. Diameter error is SMALL compared to total uncertainty

When to Use

Simplified method good for:

1. Standard cases
   - Typical geometries (vertical spark, toroid topload)
   - Typical frequencies (100-300 kHz)
   - Typical lengths (1-3 m)

2. First-pass analysis
   - Initial design evaluation
   - Quick parameter studies

3. Engineering estimates
   - ±20% accuracy sufficient
   - Fast turnaround needed

4. Educational purposes
   - Understanding physics
   - Building intuition

Iterative method when:

1. Research/validation
   - Publication-quality results
   - Detailed physics studies

2. Extreme parameters
   - Very long sparks (>5 m)
   - Very short sparks (<0.5 m)
   - Very low frequency (<50 kHz)

3. Measurement comparison
   - Highest accuracy required
   - Factor of 1.5 differences matter

4. Unusual geometries
   - Horizontal sparks
   - Branched discharge
   - Non-uniform diameter

Computational savings:

Iterative:
  5-10 iterations × 20 points × n segments
  = 1000-2000 AC analyses
  Time: 100-500 seconds for n=10

Simplified:
  1 AC analysis (after R calculation)
  Time: <1 second

Speedup: 1000-5000× faster!

Use simplified unless specific need for iterative

Worked Example: Simplified Calculation

Given (same matrix as before):

f = 190 kHz
ω = 2π × 190×10³ = 1.194×10⁶ rad/s

Capacitance matrix (n=5):
         [0]     [1]     [2]     [3]     [4]     [5]
[0]  [  32.5    -9.2    -3.1    -1.2    -0.6    -0.3  ]
[1]  [  -9.2    14.8    -2.8    -0.9    -0.4    -0.2  ]
[2]  [  -3.1    -2.8    10.4    -2.1    -0.7    -0.3  ]
[3]  [  -1.2    -0.9    -2.1     8.6    -1.8    -0.5  ]
[4]  [  -0.6    -0.4    -0.7    -1.8     7.4    -1.4  ]
[5]  [  -0.3    -0.2    -0.3    -0.5    -1.4     5.8  ] pF

Calculate R[i] for each segment:

Segment 1 (base):

C_total[1] = |C[1,0]| + |C[1,2]| + |C[1,3]| + |C[1,4]| + |C[1,5]|
           = 9.2 + 2.8 + 0.9 + 0.4 + 0.2
           = 13.5 pF

R[1] = 1 / (ω × C_total[1])
     = 1 / (1.194×10⁶ × 13.5×10⁻¹²)
     = 1 / (1.612×10⁻⁵)
     = 62.0 kΩ

Check bounds: 1k < 62k < 100k ✓

Segment 2:

C_total[2] = 3.1 + 2.8 + 2.1 + 0.7 + 0.3 = 9.0 pF

R[2] = 1 / (1.194×10⁶ × 9.0×10⁻¹²)
     = 93.0 kΩ ✓

Segment 3:

C_total[3] = 1.2 + 0.9 + 2.1 + 1.8 + 0.5 = 6.5 pF

R[3] = 1 / (1.194×10⁶ × 6.5×10⁻¹²)
     = 129 kΩ ✓

Segment 4:

C_total[4] = 0.6 + 0.4 + 0.7 + 1.8 + 1.4 = 4.9 pF

R[4] = 1 / (1.194×10⁶ × 4.9×10⁻¹²)
     = 171 kΩ ✓

Segment 5 (tip):

C_total[5] = 0.3 + 0.2 + 0.3 + 0.5 + 1.4 = 2.7 pF

R[5] = 1 / (1.194×10⁶ × 2.7×10⁻¹²)
     = 310 kΩ ✓

Summary:

R[1] = 62 kΩ   (base, lowest)
R[2] = 93 kΩ
R[3] = 129 kΩ
R[4] = 171 kΩ
R[5] = 310 kΩ  (tip, highest)

✓ Monotonically increasing
✓ All within position-dependent bounds

Total: R_total = 765 kΩ

Validation

Total resistance check:

Expected at 190 kHz for 2 m spark:
  Lower bound: ~50 kΩ (very hot, efficient)
  Typical: 100-300 kΩ
  Upper bound: ~500 kΩ (cool, streamer-dominated)

Result: 765 kΩ

Higher than typical, but reasonable because:
  1. Long spark (2 m)
  2. Distributed model (tip high R, 310 kΩ)
  3. Tip weakly coupled (high R expected)

Within factor of 2-3 of typical: Acceptable

If result very different:

R_total < 20 kΩ:
  - Check formula (missing units conversion?)
  - Check C values (pF vs F?)
  - Too low for plasma physics

R_total > 2 MΩ:
  - Check frequency (Hz vs kHz?)
  - Tip resistance very high (check tip coupling)
  - May need iterative method to find lower solution

Total Resistance Validation Ranges

Frequency dependence:

At 100 kHz:
  Typical: 100-600 kΩ (higher R at lower f)

At 200 kHz:
  Typical: 50-300 kΩ

At 400 kHz:
  Typical: 25-150 kΩ (lower R at higher f)

Rule: R_total ∝ 1/f  (approximately)

Length dependence:

At 200 kHz:
  0.5 m: 30-100 kΩ
  1.0 m: 50-150 kΩ
  2.0 m: 100-300 kΩ
  3.0 m: 150-500 kΩ

Rule: R_total ∝ L (approximately, distributed effects complicate)

Operating mode:

QCW (long ramp):
  - Lower R (hot channel)
  - Factor 0.5-1× above estimates

Burst (short pulse):
  - Higher R (cooler channel)
  - Factor 1-2× above estimates

Key Takeaways

Iterative optimization maximizes power per segment, uses damping (α ≈ 0.3-0.5) for stability, 5-10 iterations typical
Position-dependent bounds: R_min increases 1k→10k, R_max increases 100k→100M from base to tip (quadratic)
Convergence: Base segments converge fast (sharp power peak), tip segments slow (flat curve, weakly coupled)
Simplified method: R[i] = 1/(ω × C_total[i]) from circuit theory, 1000× faster, ±20% accuracy
When simplified: Standard cases, first-pass analysis, engineering estimates, educational use
When iterative: Research, extreme parameters, measurement comparison, publication quality
Validation: R_total should be 50-500 kΩ at 200 kHz for 1-3 m sparks, monotonic increase base→tip
Total resistance: Scales as R ∝ 1/f and R ∝ L approximately, QCW lower than burst mode

Practice

{exercise:model-ex-05}

Next Lesson: Part 4 Review and Comprehensive Exercises

16 KiB Raw Blame History