Algorithms for fitting Hansen Solubility Parameters in HSPiPy

[Gnpd]

Update (Current Status)

HSPiPy now uses the classic grid-based algorithm as the default fitting method, which provides deterministic results and good performance for most use cases. The library maintains support for multiple optimization approaches to accommodate different needs.

Context

HSPiPy fits Hansen Solubility Parameter (HSP) spheres using several algorithms:

1. Classic grid-based search (new default) - Deterministic, fast, and reproducible
2. Differential evolution from SciPy - Global optimization, robust but slower
3. Various SciPy minimize methods - Local optimization for specific use cases

Current Implementation

Available Algorithms:

• classic (default): Grid-based search with deterministic results
• differential_evolution: Global optimization for complex landscapes
• Nelder-Mead, Powell, BFGS, L-BFGS-B, TNC, COBYLA, SLSQP: Local optimization methods

Algorithm Characteristics:

Method	Type	Speed	Reproducible	Global Search
classic	Grid-based	Fast	✅ Yes	Limited
differential_evolution	Evolutionary	Slow	❌ No	✅ Yes
Local optimizers	Gradient-based	Medium	✅ Yes	❌ No

Questions for the Community

• Experience: Which optimizers have worked best for your specific HSP fitting problems?
• Reproducibility: How important is deterministic behavior in your workflow? Is it more a grid design issue?
• Benchmarks: Would anyone be interested in contributing datasets to compare algorithm performance?
• New Algorithms: Are there specific optimization approaches you'd like to see added?

---

💬 Feedback is very welcome! Please share your experiences, preferences, and feel free to include links to relevant publications.