Research
1 entriesMachine learning potentials for carbon
Method Gaussian approximation potentials fitted to van der Waals-inclusive DFT and run in LAMMPS System Carbon from graphene and diamond to the liquid, amorphous networks and free clusters Result Density, not temperature, decides whether an annealed carbon network graphitisesProjects
5 entriesGenerative language models for small molecules
Method Character-level LSTMs over SMILES, trained from scratch, conditioned on indication System ChEMBL, late-stage and approved drugs across the 50 most common indications Result Conditional generation from an indication label, with the failure modes separated out 2022–24Antibody selection and engineering workflows
Method Structure-based simulation and ML scoring over high-throughput screening data System Therapeutic antibodies, from early screening through to candidate delivery Result Deployed across more than thirty clinical targets 2025Atomic orbitals, computed and drawn properly
Method Vectorised evaluation of hydrogenic wavefunctions on an xarray grid, marching cubes for isosurfaces System One-electron atomic orbitals at arbitrary n, l and m Result Coarse evaluation, interpolation, then a smooth isosurface. Pure Python, no viewer 2025Infinite monkeys, or character-level language models
Method Character-level RNNs and LSTMs in PyTorch, trained from scratch System The complete works of Shakespeare, one character at a time Result Coherent-looking pastiche from a model that never saw a word boundary 2025Continuum Clock Live demo
Method Vanilla HTML, CSS and JavaScript; conic gradients, no framework, no build step System A wall clock, if the hands were the size of the wall Result Live and self-hosted here; also running at harishpersad.comPublications
13 entries- 2023
V. de Puyraimond, P. Rowe, J. Mai, …, B. C. Barnhart. A rational approach for selecting CD3-binding antibodies for T-cell engager development. J. Immunother. Cancer 11 (Suppl 1), A1523. Poster
- 2023
D. Tortora, P. Bergqvist, T. Jacobs, …, P. Rowe (25th of 31), …, B. C. Barnhart. Discovery and development of functional and specific T-cell engagers against a MAGE-A4 pMHC. J. Immunother. Cancer 11 (Suppl 1), A1550. Poster
- 2023
- 2022
B. Karasulu, J.-M. Leyssale, P. Rowe, C. Weber, C. de Tomas. Accelerating the prediction of large carbon clusters via structure search: evaluation of machine-learning and classical potentials. Carbon 191, 255–266.
GAP-20 matched the DFT minima, energies, rings and coordination of random-structure-searched carbon clusters better than every classical potential in a seven-way test. It over-bound C₆₀ by 0.10 eV per atom where the classical potentials under-bound it by 0.30 to 0.91, and then carried the search beyond DFT's reach, to 720 atoms. Third author.
- 2022
- 2021
F. L. Thiemann, P. Rowe, A. Zen, E. A. Müller, A. Michaelides. Defect-dependent corrugation in graphene. Nano Lett. 21, 8143–8150.
Point defects corrugate free-standing graphene even at a concentration of 0.2%, and the type matters. At 3% divacancies make the sheet almost five times as corrugated as pristine graphene and Stone–Wales defects almost three, and only the machine-learned potential found the divacancy reconstruction unprompted; the two classical potentials tested had to be told. Second author.
- 2021
- 2020
P. Rowe, V. L. Deringer, P. Gasparotto, G. Csányi, A. Michaelides. An accurate and transferable machine learning potential for carbon. J. Chem. Phys. 153, 034702.
One potential for diamond, graphite, graphene, nanotubes, fullerenes, the liquid and amorphous carbon, with lattice parameters within 0.2% of DFT on average. Covering all of it from a database of about 17,000 DFT configurations cost a factor of four in graphene phonon accuracy, 4 meV against the graphene-only model's 1, and a 1,000-structure random search said the price was right. First author.
- 2020
K. Huang, P. Rowe, C. Chi, …, R. R. Nair. Cation-controlled wetting properties of vermiculite membranes and its promise for fouling resistant oil–water separation. Nat. Commun. 11, 1097.
- 2020
- 2018
P. Rowe, G. Csányi, D. Alfè, A. Michaelides. Development of a machine learning potential for graphene. Phys. Rev. B 97, 054303.
A Gaussian approximation potential for graphene with an in-plane force error of 0.028 eV Å⁻¹ against DFT. That is twenty times below the best empirical potential, and the paper named the catch itself: the model knew one form of carbon and could not describe diamond. First author.
- 2017
- 2017