AuNP Speciation — UV-Vis broadening from monomer/dimer/trimer equilibria

Part of Speciate — Built with Claude: Life Sciences (Builder Track). An automated pipeline that turns a TEM micrograph of gold nanoparticles into mean size, size distribution, and aggregation state, then predicts the UV-Vis spectrum with zero fitted parameters — replacing a half-day of manual counting. This repo is the optics + inversion half; stage-one automated TEM sizing lives in tem-particle-metrics.

📄 Read the mid-project progress report (PDF) — full writeup: motivation, validation, and findings.

1 · Foundation-model segmentation handles aggregates

Classical thresholding merges touching particles into oversized blobs (left); an NP-SAM / Segment-Anything model separates them correctly (right). Automated sizes match hand-counting to within ±6% across all eight samples, from up to 6× more particles.

2 · Zero-parameter UV-Vis prediction from TEM size alone

Predicted (red) vs measured (black), normalized to peak. Well-dispersed samples match within a few nm with nothing fitted; where the prediction breaks (right, 55 nm), that 46 nm gap is the aggregation signal.

3 · Aggregation, quantified

Exact multipole T-matrix fit (red) vs fast CDA (dotted): the aggregated gold fraction per sample — 0% for the pure control, ~⅓ to ⅘ for the aggregated batches. Only the total fraction is identifiable, so size comes from TEM.

---

Prototype model for the hypothesis (Dragnea group, Indiana University) that the anomalously broad, red-tailed UV-Vis of 12 nm gold colloids is caused by reversible cluster formation (monomer ⇌ dimer ⇌ trimer …), not size polydispersity alone.

See CLAUDE.md for the scientific background, architecture, and — importantly — the known limitations. MODELLING_STACK.md documents each modelling choice and its alternatives: production optics use tabulated Johnson & Christy gold with an exact multi-sphere T-matrix backend for clusters (the fast coupled-dipole model under-couples at near-contact and serves only as a lower bound).

Install

pip install -r requirements.txt              # system env: CDA/Mie/fits
# exact T-matrix backend needs a pinned venv:
python -m venv mstm-env
mstm-env/bin/pip install "numpy==1.26.4" "scipy==1.11.4" treams matplotlib

Run

python scripts/verify.py   # physics sanity checks
python scripts/demo.py     # writes figures to outputs/

What the demo shows

• fig1_species_spectra.png — per-particle extinction of monomer / dimer / linear & triangular trimer (coupling adds red-side intensity).
• fig2_broadening.png — monomer-only (4% polydispersity) vs a realistic speciation mixture; the peak barely moves but a red tail appears (quantified by a red-tail index, not FWHM).
• fig3_isosbestic.png — temperature series (van 't Hoff equilibrium) showing an isosbestic point, the thermodynamic fingerprint of the two-species mix.
• fig4_gap_sensitivity.png — interparticle gap is the dominant lever on the red tail.
• fig5_inverse_fit.png — Layer 3 single-spectrum fit: recovers the aggregated gold fraction; flags size/polydispersity/split as under-determined.
• fig6_mstm_vs_cda.png — exact T-matrix vs point-dipole CDA: CDA under-couples (exact gives 1.6×/2.8× more red-tail for dimer/trimer). Needs the mstm-env venv.
• fig7_global_fit.png — global multi-temperature fit: recovers size, thermodynamics (ΔH, ΔS) and the aggregated-fraction-vs-T curve jointly.
• fig8_fit_real.png — real-data driver: loads a UV-Vis CSV and fits it with the cached EXACT optics. On the self-consistent example it recovers D=12.0 nm, ΔH₂=-25.4 kJ/mol, and the aggregated-vs-T curve to the third decimal.

Package layout

src/aunp_speciation/
  dielectric.py   gold ε(λ)  [analytic — swap for J&C]
  mie.py          single-sphere Mie + dipole polarizability
  clusters.py         coupled-dipole (CDA) clusters — fast, lower-bound coupling
  clusters_tmatrix.py exact multipole T-matrix backend (treams; mstm-env venv)
  equilibrium.py      monomer/dimer/trimer association vs temperature
  spectra.py          polydispersity + mixing; backend='cda'|'tmatrix'
  fitting.py          Layer 3 single-spectrum inverse fit
  fit_global.py       Layer 3 global multi-temperature fit
  basis_cache.py      precomputed exact-optics basis + interpolator
  io_data.py          load experimental UV-Vis CSV/TSV (single or T-series)
scripts/
  verify.py           regression / physics checks
  demo.py             figs 1-4
  fit_demo.py         single-spectrum fit self-test (fig5)
  validate_mstm.py    exact-vs-CDA validation (fig6; mstm-env/bin/python)
  fit_global_demo.py  global multi-temperature fit self-test (fig7)
  build_tmatrix_basis.py  precompute exact basis cache (mstm-env/bin/python)
  make_example_data.py    write data/example_series.csv
  fit_real.py             load & fit a real UV-Vis file (fig8)
data/
  example_series.csv  synthetic example temperature series for fit_real.py

Status

Layers 1-3 implemented, plus an exact T-matrix optics backend and a global multi-temperature fit. Single-spectrum fit gives the aggregated fraction; the global T-fit adds size + thermodynamics (ΔH, ΔS) + speciation-vs-T. Remaining: wire tabulated ε + size-damping into the size-integral, precompute a T-matrix basis for quantitative fits, add the Haiss ratio, and run on real data. See CLAUDE.md.