Skill: Design Flash Benchmark

Purpose

Create a structured test matrix for comparing flash algorithm performance across fluid types, thermodynamic conditions, and difficulty levels.

When to Use

• Starting a new flash algorithm comparison study
• Extending an existing benchmark to new fluid families
• Designing stress tests for near-critical or multi-phase regions

Benchmark Design Procedure

Step 1: Select Fluid Families

Choose from these standard families:

Step 2: Define PT Space

For each family, define the pressure-temperature sampling grid:

import numpy as np

def generate_pt_grid(T_min_K, T_max_K, P_min_bara, P_max_bara, n_T=20, n_P=20):
    """Generate a regular PT grid."""
    T_values = np.linspace(T_min_K, T_max_K, n_T)
    P_values = np.logspace(np.log10(P_min_bara), np.log10(P_max_bara), n_P)
    cases = []
    for T in T_values:
        for P in P_values:
            cases.append({"T_K": float(T), "P_bara": float(P)})
    return cases

Standard ranges by family:

Step 3: Add Stress Cases

Add cases specifically designed to challenge the algorithm:

1. Near bubble point: T slightly above Tbub at given P
2. Near dew point: T slightly below Tdew at given P
3. Near critical: T ≈ Tc ± 5K, P ≈ Pc ± 5 bar
4. Very low vapor fraction: β ≈ 0.001
5. Very high vapor fraction: β ≈ 0.999
6. Single phase (verification): Known single-phase conditions
7. Trace components: One component at < 1e-6 mole fraction

Step 4: Define Composition Perturbation

Use Dirichlet sampling to generate composition variants:

from numpy.random import dirichlet

def perturb_composition(base_comp, n_variants=10, concentration=50):
    """Generate composition variants around a base composition.

    Higher concentration = less perturbation.
    """
    names = list(base_comp.keys())
    alpha = np.array([base_comp[n] for n in names]) * concentration
    variants = []
    for _ in range(n_variants):
        x = dirichlet(alpha)
        variants.append(dict(zip(names, x.tolist())))
    return variants

Step 5: Define Metrics

Every benchmark case must record:

Step 6: Estimate Total Cases

Target: 500–2000 cases per algorithm version.

Step 7: Generate Config File

Output benchmark_config.json:

{
  "benchmark_id": "tpflash_2026_01",
  "created": "2026-03-31",
  "algorithms": ["baseline", "candidate_eigenvalue_switch"],
  "eos_models": ["SRK"],
  "timing_repeats": 3,
  "families": [
    {
      "name": "lean_gas",
      "base_composition": {"methane": 0.90, "ethane": 0.05, "propane": 0.03, "nitrogen": 0.01, "CO2": 0.01},
      "n_composition_variants": 10,
      "dirichlet_concentration": 50,
      "T_range_K": [200, 400],
      "P_range_bara": [1, 200],
      "n_T": 20,
      "n_P": 20
    }
  ],
  "stress_cases": {
    "near_critical": 20,
    "near_bubble": 10,
    "near_dew": 10,
    "trace_component": 10
  }
}

Checklist

• All components available in NeqSim database
• PT ranges cover the two-phase region for each family
• Stress cases are well-defined
• Metrics list is complete
• Total case count is between 500 and 2000
• Random seed is fixed for reproducibility