We were delighted when Maths4DL members Teo Deveney, James Rowbottom and Georg Maierhofer were invited to attend the International Conference on Machine Learning (ICML) 2025 to present Maths4DL-based research titled “G-Adaptivity: Optimised Graph-Based Mesh Relocation for Finite Element Methods” as a spotlighted poster.
ICML is one of the three premier conferences in the field and is widely regarded as a top venue for publishing cutting-edge research in machine learning and artificial intelligence. This year’s conference drew over 8,000 registered participants and featured 3,333 accepted papers out of more than 12,000 submissions. The programme included tutorials on Sunday and Monday, the main conference track with invited talks and posters from Tuesday through Thursday, and specialist workshops on Friday and Saturday.
M4DL’s poster was extremely well received, both during the review phase – earning spotlight recognition as one of the top ~3% of submissions – and at the conference itself. Researchers from over 20 institutions expressed strong interest in the results, many highlighting the practical value of our openly available code and its potential for future application. Notably, the paper was among a small number accepted in the scientific machine learning (SciML) domain, underscoring the value of research undertaken within the Maths4DL programme.
The conference was dominated by many interesting topics at the forefront of machine learning research including large language models, generative-AI, foundation models, model distillation, alignment and scaling. Many topics overlapping with M4DL were also represented, such as neural operators, diffusion and generative models, graph neural networks, and optimisation.
Interestingly several works at the interface of language and scientific applications were presented, such as LLM-based foundation models for GNNs, and LLMs for dynamics discovery and prediction. Scientific Machine Learning had a visible presence at the conference, although it lacked dedicated sessions or workshops, instead appearing in pockets across the programme and many valuable conversations were had around the poster sessions.
Outside of the interesting talks and poster sessions, the conference was not short of extra-curricular activities with the conference venue being strategically positioned in Vancouver harbour. Highlights included watching cruise ships and sea planes dock and land in the harbour, walking to the spectacular Stanley Park as well as attending hosted waterside events by big tech and finance companies including Google DeepMind, Microsoft Research and Jane Street Capital.
In summary, ICML 2025 was an outstanding opportunity to showcase M4DL’s work and build new research connections that may lead to future collaborations. The level of investment in this space from academia, government, and industry is striking, and it is clear that Maths4DL is well positioned to contribute meaningfully to this rapidly evolving and increasingly important field.
We are very grateful to Teo, James and Georg for their excellent effort in showcasing some of our research,
