My research

I use physics-informed machine learning to make calculations for molecules and materials more practical at scale.

Publications

Inverse Problems · 2022

Conductivity maps for graphene

20-contact 2D graphene sheet feeding ML Adaptive Electrode Selection (A-ESA, 11–41% lower RMSE) and GREIT conductivity reconstruction
Conductivity from edge measurements (11–41% lower RMSE with A-ESA).
Quick preview

We used machine learning to choose which electrodes to measure with, then tested conductivity reconstruction on a graphene laminate film.

J. Chem. Theory Comput. · 2026

Faster TDDFT-level predictions

JCTC 2026 Pearson correlation comparison: ZINDO 0.77 vs ZINDO plus ΔML 0.96 on 9,500 M06-2X test molecules and 0.88 to 0.99 on 24,000 QCDGE molecules
Published JCTC 2026 accuracy across 9.5k M06-2X and 24k QCDGE molecules.
Quick preview

A graph model learns a correction to fast ZINDO calculations. On 9,500 test molecules, the correction increased correlation with TDDFT from 0.77 to 0.96.

PhD thesis · Liverpool · 2026

Learning electronic structure

My thesis looks at two ways machine learning can speed up electronic-structure calculations for conjugated molecules.

  • Ground states. A learned Fock matrix updates as the electron density changes during the calculation.
  • Excited states. A learned correction brings fast ZINDO estimates closer to TDDFT reference energies.
Quick preview

The thesis explores how using electron densities distorted by external point charges helps the ground-state model respond to electrostatic changes.

Density-dependent orbital blocks F(D,S) passing through PCD-trained MLP-H (r above 0.98, 29.5 meV MAE) inside an iterative ML-SCF convergence loop
The learned Fock matrix in the self-consistent field cycle (Chapters 3–4).

Current work

Extending ΔML for excited states

I’m exploring how to use learned corrections to predict more excited-state properties in a single model, including singlet and triplet energies.

Hamiltonians for 2D materials

I’m exploring how learned Hamiltonians can describe electronic structure in materials where strain, defects, and spin–orbit effects matter.

Sneak peek · Ongoing development for Δ-ML · Papers inbound over this year

New descriptors, new architectures, lower errors

Extending our 2026 JCTC ΔML-ZINDO framework to joint multi-property excited-state prediction (singlet S₁–S₃, triplet T₁–T₂, frontier orbitals, and ΔEST) with guaranteed physical state ordering.

0.097 eVUnseen test MAE · E(S₁)
0.090 eVUnseen test MAE · E(T₁)
0.093 eVUnseen test MAE · ΔEST
0.00%Ordering violations (2,744 mols)
Compact 3-panel parity strip comparing Raw ZINDO against Next-Gen Δ-ML predictions vs TDDFT across E(S1), E(T1), and ΔE_ST for 2,744 unseen test molecules
ML-corrected vs. TDDFT parity across 2,744 unseen test molecules (M06-2X/3-21G*). Open interactive parity explorer →