Skip to main content

Module experimental_status

sci_form

Module experimental_status 

Source
Expand description

E4 – Experimental Module Validation & Retention Criteria

This module documents which experimental methods remain gated behind feature flags and the criteria required for eventual promotion to core.

ModuleCore pathJustification
EEQ chargescharges_eeqTopology-free partial charges, O(N²), validated vs Gasteiger
D4 dispersiondispersionBJ-damped DFT-D4 energy, complements UFF/MMFF94
ALPB solvationsolvation_alpbKlamt-corrected solvation, improves over plain GB
sTDA UV-Visspectroscopy::stda_uvvisCIS-like excitations from monopole charges, ~1000× faster than full CI
GIAO NMRspectroscopy::giao_nmrMulti-nucleus NMR shieldings, complements empirical HOSE

§Retained as Experimental (E4)

§EXP-201: Kernel Polynomial Method (KPM)

Feature flag: experimental-kpm Location: experimental/kpm/ (chebyshev.rs, density.rs)

KPM approximates the density of states via Chebyshev expansion of the spectral function. Useful for very large systems (10k+ atoms) but the current implementation lacks:

  • Adaptive polynomial order selection
  • Sparse-matrix backend for O(N) scaling
  • Benchmarks against direct diagonalization for <500 atoms

Promotion criteria:

  1. Demonstrate wall-clock advantage over direct DOS for >2000 atoms
  2. Validate spectral accuracy within 0.1 eV of full diagonalization
  3. Add sparse matrix support (nalgebra-sparse or sprs)

§EXP-202: Multi-Basin Hopping (MBH)

Feature flag: experimental-mbh Location: experimental/mbh/ (blocks.rs, hessian.rs)

Block-diagonal Hessian approach for exploring conformational basins. Intended for global optimization of flexible molecules.

Current limitations:

  • No energy evaluator integration (needs UFF/MMFF94 coupling)
  • Block decomposition heuristic is molecule-size dependent
  • Missing Metropolis acceptance criterion

Promotion criteria:

  1. Integrate with forcefield module for energy evaluation
  2. Reproduce known global minima for test molecules (alanine dipeptide, etc.)
  3. Compare basin-hopping yield vs multi-seed ETKDG

§EXP-203: Randomized Numerical Linear Algebra (RandNLA)

Feature flag: experimental-randnla Location: experimental/rand_nla/ (nystrom.rs, solver.rs)

Nyström approximation and randomized eigensolvers for large overlap/Fock matrices. Currently a backend alternative, not a user-facing feature.

Current limitations:

  • Rank selection requires manual tuning
  • No automatic accuracy estimation
  • Not yet wired into SCF or EHT solvers

Promotion criteria:

  1. Wire into EHT/SCF as optional backend with feature flag
  2. Show <1% error in eigenvalues for rank = N/4
  3. Demonstrate speedup for N>500 basis functions

§EXP-204: Spectral Dimensionality Reduction (SDR)

Feature flag: experimental-sdr Location: experimental/sdr/ (embedding.rs, projections.rs)

Diffusion-map and spectral embedding for chemical space visualization. Alternative to fingerprint-based Tanimoto similarity.

Current limitations:

  • Kernel bandwidth selection is fixed
  • No out-of-sample extension (Nyström embedding)
  • Limited to datasets fitting in memory

Promotion criteria:

  1. Validate embeddings preserve known SAR relationships
  2. Add adaptive bandwidth selection
  3. Compare clustering quality vs Butina (Tanimoto)

§EXP-205a: Conformal Geometric Algebra (CGA)

Feature flag: experimental-cga Location: experimental/cga/ (multivector.rs, motor.rs, conformer.rs, materials.rs)

CGA representation of molecular geometry: motors for rigid transforms, conformal points for distance geometry. Research-stage alternative to Cartesian coordinate manipulation.

Current limitations:

  • 5D multivector operations have overhead vs 3D rotations
  • Conformer pipeline not competitive with ETKDG
  • Materials assembly untested with real MOF structures

Promotion criteria:

  1. Demonstrate advantage in distance geometry constraint handling
  2. Benchmark CGA conformer vs ETKDG quality and speed
  3. Show utility for symmetry-constrained crystal assembly

§EXP-205b: Growing String Method (GSM)

Feature flag: experimental-gsm Location: experimental/gsm/ (string.rs, saddle.rs)

Reaction path finding via growing string interpolation between reactant and product geometries. Identifies transition states.

Current limitations:

  • No gradient source (needs coupled PES evaluation)
  • String convergence criteria incomplete
  • Saddle point refinement not validated

Promotion criteria:

  1. Couple with PM3 or xTB gradient evaluation
  2. Reproduce known barrier heights (SN2, Diels-Alder)
  3. Compare path energy profiles vs NEB reference

§EXP-205c: Continuous Phase Methods (CPM)

Feature flag: experimental-cpm Location: experimental/cpm/ (surface.rs, grand_potential.rs)

Phase-field and grand potential methods for crystal surface thermodynamics and phase boundary determination.

Current limitations:

  • Grand potential formulation needs thermodynamic data
  • Surface energy calculations untested for ternary systems
  • No integration with materials module

Promotion criteria:

  1. Validate surface energies against DFT reference for binary alloys
  2. Integrate with materials::UnitCell for consistent crystal handling
  3. Add Wulff construction for equilibrium crystal shapes