Drug Crystallization Multi-Objective Optimization

What does Drug Crystallization Multi-Objective Optimization do?

Purpose

Advances a crystallization process optimization requirement from "having a set of parameters / a target process stage" to "obtaining an optimal process solution (complete Pareto front) that satisfies multiple constraints, is reproducible, and has a quantitative comparison with the baseline."

The core is an interpretable deterministic kinetic model (digital twin) that maps 6 process variables to 4 quality metrics (yield / purity / particle size D50 / dissolution similarity factor f2). This SKILL orchestrates: identifying the requirement → invoking the computational core → interpreting the Pareto front → outputting a before/after optimization comparison.

Core Method (Interpretable Kinetic Engine)

Module	Implementation	Literature Anchor
Solubility Curve	van't Hoff: C*(T)=4.48·exp[8.3/R·(1/293−1/T)]	Grade A: Biowaiver Monograph 2017 (20℃ 4.48 mg/mL)
Nucleation Rate	B = kb·max(σ−0.3,0)⁴ + kb2·G·μ2·f_mixing(stirring)	Primary classical theory + secondary growth coupling; threshold S=4.5 (Tian Binbin 2015)
Growth Rate	G = kg·σ^m, m=1.9955 (BCF screw dislocation growth)	Grade A: Tian Binbin et al., Chemical Engineering 2015 measured
Population Balance Equation (PBE)	Moment method μ0…μ4 recursion, concentration-moment fully coupled material balance conservation	Grade B: Randolph & Larson moment method

Optimization Method (Multi-Objective + Multi-Agent)

• Primary Output: NSGA-II 4-objective complete Pareto front (80 non-dominated solutions + knee point + max-f₂/max-yield extremes), objectives = yield↑ / purity↑ / f2↑ / particle size towards 30μm↑.
• Auxiliary: Gaussian Process (GP, RBF kernel + white noise, hyperparameters learned via log marginal likelihood LML) + Bayesian Optimization (EGO-style EI), for scalarization reference point cross-validation.
• Five-Agent Pipeline: Literature (literature anchor gatekeeper) → Kinetics Modeler (PBE solver + material balance self-check) → Optimizer (Pareto front) → Risk/Regulatory (regulatory criteria + CQA) → Reviewer (independent review, PASS/FAIL verdict).
• Sensitivity Analysis: Robustness assessment under temperature ±1°C / pH ±0.1 / cooling rate ±10% perturbations.

Trigger Conditions (When to Load This SKILL)

Triggered when any of the following scenarios are met:

• User provides crystallization process parameters and requests evaluation of quality metrics or whether targets are met.
• User requests to "optimize / search / find optimal parameters" for a crystallization process stage.
• User mentions quality objectives such as yield / purity / particle size D50 / dissolution f2 and it involves crystallization.
• User requests evaluation of dissolution similarity (f2) between generic and innovator drugs.
• User provides a reference dissolution curve or literature baseline and requests a comparison.

Trigger Keywords: crystallization process optimization, crystallization parameter optimization, multi-objective process optimization, generic drug crystallization, particle size window, dissolution f2, cooling curve, seeding strategy, supersaturation control, amoxicillin crystallization, Pareto front, population balance.

Input/Output Format

Input

After the SKILL is loaded by the AI in a session, the computational core is driven by running scripts/crystallization.py (passing JSON via stdin). CLI input protocol:

{
  "mode": "evaluate | optimize | sensitivity | agents | dissolution_plot | explain",
  "params": {                       // Required for evaluate mode, 6 process variables
    "T_cryst_C": 15,                // Crystallization temperature ℃, [5, 25]
    "r_cool_C_per_min": 0.2,        // Cooling rate ℃/min, [0.05, 1.0]
    "S0": 1.8,                      // Initial supersaturation (dimensionless), [1.2, 2.5]
    "seed_wt_pct": 2,               // Seed mass %, [0, 5.0]
    "stir_rpm": 300,                // Stirring speed rpm, [100, 600]
    "hold_t_min": 120               // Holding time min, [30, 240]
  },
  "seed": 42                        // Optional for optimize/agents mode, random seed (default 42 for deterministic reproducibility)
}

Output

The computational core outputs a single-line JSON to stdout. Using mode=optimize as an example, core fields:

{
  "params": {"T_cryst_C": 5.9, "r_cool_C_per_min": 0.11, "S0": 1.9, "seed_wt_pct": 4.58, "stir_rpm": 586, "hold_t_min": 238},
  "metrics": {"yield_pct": 67.75, "purity_pct": 99.44, "D50_um": 44.26, "D90_um": 116.0, "f2": 79.13},
  "score": 174.88,
  "pareto": {"n_pareto": 80, "knee_point": {...}, "extremes": {"max_f2": {...}, "max_yield": {...}}, "pareto_front": [...]},
  "agents": {"literature": {...}, "kinetics": {...}, "optimizer": {...}, "risk_regulatory": {...}, "reviewer": {"verdict": "PASS", ...}},
  "sensitivity": {"robustness": {"conclusion": "Robust: f2 remains ≥50 under three-factor ± perturbation", ...}},
  "surrogate_cv": {"r2_mean": 0.74, ...},
  "bo_trace": [...],
  "cqa_coverage": {...}
}

The AI interprets the JSON result into a human-readable comparison conclusion, following the format in this file's "Result Interpretation" section.

Execution Steps (Standard Workflow)

Execute in order; the result of the previous step determines the next action:

1. Identify Requirement: Confirm target API and process stage; determine if the user wants to "evaluate a single set of parameters" or "search for optimal parameters."
2. Establish Baseline: Read the classic process baseline from the literature (or user-provided); default amoxicillin baseline is in references/tuning-guide.md.
3. Multi-Objective Optimization (optimize): Invoke mode=optimize, primary output is the complete Pareto front (NSGA-II 80 non-dominated solutions).
4. Result Evaluation (evaluate): Call mode=evaluate on candidate solutions, verify checks.all_satisfied and PBE process variables.
5. Independent Review: Reviewer agent independently verifies Pareto non-dominance and feasibility, issues PASS/FAIL.
6. Output Conclusion: Provide Pareto front + knee point/extremes + four-metric before/after comparison + sensitivity analysis conclusions + boundary statement.

Dependencies

Type	Dependency	Description
Runtime	Python 3 (python3 / python / py whichever available)	Runs the computational core
Library	NumPy 2.5.4	GP Gaussian process and NSGA-II vectorized operations (pip install numpy==2.5.4)
Core File	scripts/crystallization.py	Interpretable kinetic core (PBE) + multi-objective optimization + five agents, the only executable logic of this SKILL
Reference Doc	references/install-and-usage.md	Installation, registration, usage examples
Reference Doc	references/tuning-guide.md	Physical meaning, value ranges, and impact of all tunable parameters
External Resources	None	Self-contained, no internet connection, no external API calls

Computational Core Invocation (CLI)

# evaluate: evaluate a single parameter set
echo '{"mode":"evaluate","params":{...}}' | python scripts/crystallization.py

# optimize: output complete Pareto front (primary) + GP-BO reference points
echo '{"mode":"optimize","seed":42}' | python scripts/crystallization.py

# sensitivity: sensitivity analysis (T±1°C/pH±0.1/cooling±10%)
echo '{"mode":"sensitivity","params":{...}}' | python scripts/crystallization.py

# explain: full explanation of what optimize ran (model card/search space/weights/constraints fully transparent)
echo '{"mode":"explain"}' | python scripts/crystallization.py

Result Interpretation

• pareto.n_pareto → Number of non-dominated solutions on the Pareto front; knee_point is the trade-off solution with the smallest distance to the ideal point.
• metrics.f2 >= 50 → Dissolution curve is "similar" to the innovator reference (FDA/CDE bioequivalence bridge criterion); f2 < 50 is considered dissimilar.
• agents.reviewer.verdict = PASS → Pareto non-dominance + feasibility + baseline comparison all pass independent review.
• sensitivity.robustness.f2_stays_above_50 → Whether f2 remains ≥50 under perturbation (operational robustness).
• bo_trace → 64 evaluations recorded step by step (parameters/scores/source/GP hyperparameters/EI), optimization process fully auditable.

Boundary Statement

This SKILL output is a process search and pre-screening tool based on an interpretable kinetic engine (not a GMP experimental substitute). Final bioequivalence conclusions must be supported by actual experiments and submission data. Yield Definition: This model's yield is based on material balance (precipitated solid / solution solid), which differs from the literature's "overall yield" definition; see references/tuning-guide.md for details.