Cyprien Bone

CrystaLLM-π

Sat, 03 Jan 2026 00:00:00 +0000

Coming soon...

Fri, 02 Jan 2026 00:00:00 +0000

Coming soon...

Thu, 01 Jan 2026 00:00:00 +0000

150k GPU-hours on Isambard-AI

Wed, 30 Sep 2026 00:00:00 +0000

Keith and I have been awarded 150k GPU-hours on Isambard-AI, the UK’s national AI supercomputer in Bristol, through UKRI’s AIRR AI Open Access call.

CrystaLLM-π workshop at the PSDI & Royce Materials Data Summit

Wed, 29 Jul 2026 00:00:00 +0000

Co-organised a lecture and hands-on workshop at the PSDI & Royce Materials Data Summit in Manchester, on how to use CrystaLLM-π to predict crystal structures. Notebooks are on github.

Conditional generative models enable targeted exploration of MAX phase design space

Thu, 23 Jul 2026 00:00:00 +0000

MAX phase paper out in J. Mater. Chem. C

Thu, 23 Jul 2026 00:00:00 +0000

Jamie’s paper is out in Journal of Materials Chemistry C. It uses CrystaLLM-π to explore the MAX phase design space.

Six Open Questions in Machine-Learned Interatomic Potential Foundation Models

Fri, 05 Jun 2026 00:00:00 +0000

Six open questions in MLIP foundation models

Fri, 05 Jun 2026 00:00:00 +0000

A perspective on foundation machine-learned interatomic potentials, written as a journal club as well as some other researchers in the field. We go through six questions we think will shape the area of MLIPs over the next few years.

Talk at AI4AM 2026 in Madrid

Tue, 19 May 2026 00:00:00 +0000

Gave a talk on CrystaLLM-π at the AI4AM 2026 conference in Madrid.

Poster at the IOP workshop on ML for experimental materials data

Thu, 02 Apr 2026 00:00:00 +0000

Presented CrystaLLM-π at the IOP workshop on Machine Learning for Experimental Materials Data in London.

Poster at the GeMMS 2026 Spring School

Mon, 23 Mar 2026 00:00:00 +0000

Presented a the CrystaLLM-π poster at the GeMMS 2026 Spring School in London. A week long of lectures, tutorials and research talks on deep generative models, including latent variable, diffusion, flow-based and autoregressive models.

CrystaLLM-π preprint is out

Wed, 26 Nov 2025 00:00:00 +0000

The CrystaLLM-π preprint is on arXiv! We inject continuous properties straight into the transformer’s attention, so the model keeps what it learned in pre-training while being steered towards a target. Example applications in the paper include recovering structures from X-ray diffraction patterns, and discover new photovoltaic candidates that we then checked with DFT.

Discovery and recovery of crystalline materials with property-conditioned transformers

Wed, 26 Nov 2025 00:00:00 +0000

Poster at the ML for Molecules and Materials school in Zadar

Mon, 09 Jun 2025 00:00:00 +0000

Attended the School on Machine Learning for Molecules and Materials Research in Zadar, Croatia. A week of lectures and tutorials, as well as a poster session where I presented CrystaLLM-Prop, predicting band gaps from CIFs with the original CrystaLLM.

Poster at the AIchemy annual conference

Mon, 17 Mar 2025 00:00:00 +0000

Brought the CrystaLLM-Prop poster to the first AIchemy annual conference, at Imperial’s White City campus in London. The project uses the original CrystaLLM to predict band gaps from CIFs.

Poster at the RSC Solid State Chemistry Christmas Meeting

Mon, 16 Dec 2024 00:00:00 +0000

My first conference poster of the PhD, at the 43rd RSC Solid State Chemistry Group Christmas Meeting in Birmingham. It was on CrystaLLM-Prop, where I fine-tuned the original CrystaLLM to predict a material’s band gap straight from its CIF, comparing next-token generation against a regression head on top of the model.

1st place at the Energy Idea Challenge 2023

Wed, 01 Feb 2023 00:00:00 +0000

Prescelli Annan, Niklas Jutting, Tom Burns and I won first place at the Energy Idea Challenge 2023, run by the Imperial College Energy Society and sponsored by Octopus Energy, KPMG and Guidehouse. Over a weekend, teams of students from Oxford, Cambridge, Imperial, UCL and other universities put together proposals to help the UK hit its net zero targets. Ours was a wastewater heat pump system, with machine learning for predictive maintenance.