PhD Thesis · University of Liverpool · 2026

Electronic structure of conjugated molecules

From density-dependent machine-learned Fock matrices inside the self-consistent field cycle to ΔML-accelerated excited-state predictions.

The question

Most machine-learned Hamiltonian models predict a single static ground state from geometry alone. Can we instead learn a density-dependent Fock matrix F(D, S) that sits inside an iterative self-consistent field (SCF) cycle and responds physically to electrostatic perturbations?

The approach

Using extended π-conjugated molecules from ZINC calculated at the B3LYP level, I built blockwise neural models (MLP-H) mapping atomic-orbital density and overlap blocks to Fock blocks, using near-zero undersampling and training on densities distorted by external point charges to improve convergence.

Connection to published work

The second half of the thesis (Chapter 5) developed Message-Passing ΔML-ZINDO for fast excited-state screening, now published in the Journal of Chemical Theory and Computation (2026).

Thesis

Machine learning π-conjugated electronic structure: from density-dependent ML-Hamiltonians to ΔML-accelerated excited states
Adam Coxson · Department of Chemistry, University of Liverpool (2026)

Density-dependent ML-Fock self-consistent field loop schematic showing orbital blocks, PCD-trained MLP-H, and iterative diagonalisation
Blockwise density-dependent ML-SCF workflow developed in Chapters 3–4 of the thesis.

Point-Charge Distortion (PCD) vs. Gaussian Noise (GND)

Summary of verified thesis benchmarks across one-step block correlations (Tables 4.7 vs. 4.8), 33-molecule HCO ML-SCF trajectories, and near-zero matrix sparsity.

Three-panel PhD thesis summary comparing GND vs PCD block correlations, 33 HCO ML-SCF convergence trajectories, and Fock sparsity curves
Chapters 3–4 summary: PCD training lifts onsite block correlations and avoids the spurious out-of-manifold collapse seen under GND training.

Interactive ML-SCF & Fock-block explorer

Inspect real thesis test molecules (including ZINC000033835654 and ZINC000068719623), vary the near-zero magnitude cutoff on authentic Fock blocks, and compare DFT, PCD-trained, and GND-trained SCF trajectories step by step.

Interactive Authentic Thesis Fock Blocks & ML-SCF Trajectories Hover blocks · Drag cutoff slider · Switch molecules