Predictive Modelling » Seminars

08/06 1pm-2pm: WCPM, Kevin Huang, Warwick

James Kermode — Thu, 21 May 2026 09:46:26 GMT

When: 1pm - 2pm, Mon, 08 Jun '26

Networking Lunch: Outside L5, from 12:30pm - 1pm.

Title: Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schrödinger Equation

Abstract: Incorporating group symmetries into neural networks has been a cornerstone of success in many AI-for-science applications. Diagonal groups of isometries, which describe the invariance under a simultaneous movement of multiple objects, arise naturally in many-body quantum problems. Despite their importance, diagonal groups have received relatively little attention, as they lack a natural choice of invariant maps except in special cases. We study different ways of incorporating diagonal invariance in neural network ansätze trained via variational Monte Carlo methods, and consider specifically data augmentation, group averaging and canonicalization. We show that, contrary to standard ML setups, in-training symmetrization destabilizes training and can lead to worse performance. Our theoretical and numerical results indicate that this unexpected behavior may arise from a unique computational-statistical tradeoff not found in standard ML analyses of symmetrization. Meanwhile, we demonstrate that post hoc averaging is less sensitive to such tradeoffs and emerges as a simple, flexible and effective method for improving neural network solvers.

Bio: Kevin is a postdoctoral research fellow funded by the Engineering and Physical Sciences Research Council (EPSRC) through the ProbAI Hub. They are currently based at the ��ӰƬ, working with Gareth Roberts, and collaborate with Boris Hanin at Princeton University.

They completed a PhD in machine learning at the Gatsby Computational Neuroscience Unit, University College London, under the supervision of Peter Orbanz and Morgane Austern. During this time, they were also a visiting researcher with the LIPS group at Princeton Computer Science, hosted by Ryan P. Adams. Prior to this, they completed both their undergraduate and master’s degrees in mathematics at the University of Cambridge.

Their research lies at the intersection of machine learning theory, probability, and statistics. They study the emergence of universal structures in large-scale stochastic systems, drawing on tools from random matrix theory, high-dimensional statistics, symmetry-based inference, and stochastic optimisation. Alongside this theoretical work, they increasingly engage with applied challenges, particularly around scaling laws in neural networks, AI for scientific discovery, and the robustness and safety of machine learning models.

For the 2025–2026 academic year, he is co-organising the ProbAI online seminar series and will lead the ProbAI Theory of Scaling Laws Workshop at Warwick in summer 2026.

15/06 1pm-2pm: WCPM, Fraser Birks

James Kermode — Thu, 04 Jun 2026 08:35:09 GMT

When: 1pm - 2pm, Mon, 15 Jun '26

Networking Lunch: Outside L5, from 12:30pm - 1pm.

Title: Understanding Plasticity in Amorphous Carbon with Machine-Learned Interatomic Potentials

Abstract: Amorphous carbon (a-C) is a coating material with many applications, commonly selected for its high hardness, low friction coefficient and high wear resistance. Generally, amorphous carbon is considered a brittle material, with cracks propagating at relatively low tensile stresses due to the presence of structural defects. However, when defect-free thin films are studied, experiments have shown that a-C can exhibit an unusual combination of high stiffness (~210 GPa) and unexpectedly large failure strain (~11%) [1]. At present, it is still unclear what structures and mechanisms in these films give rise to such anomalous mechanical behaviour, making this fertile ground for atomistic studies using realistic machine-learned interatomic potentials (MLIPs). In this talk, I will first give an overview of the history of atomistic modelling, introducing MLIPs and discussing their strengths and weaknesses. I will then present the results from two studies [2, 3] which combine MLIPs with molecular dynamics and arclength continuation to elucidate the underlying mechanisms governing plasticity in a-C. I will further show that, with careful structure preparation, it is possible to reproduce the experimental stress-strain response, opening the door to direct collaboration between theorists and experimentalists on this important class of materials.

[1] Yoon, J., Jang, Y., Kim, K., Kim, J., Son, S., & Lee, Z. (2022). In situ tensile and fracture behavior of monolithic ultra-thin amorphous carbon in TEM. Carbon, 196, 236–242. https://doi.org/10.1016/j.carbon.2022.04.062

[2] Birks, F., Ghanem, I., Pastewka, L., Kermode, J., & Buze, M. (2026). Resolving structural avalanches in amorphous carbon with arclength continuation. Physical Review Letters. https://doi.org/10.1103/6n5m-rxc1

[3] Birks, F., & Kermode, J. (2026). [Manuscript in preparation].

Bio: Fraser is a PhD researcher in Computational Materials Science at the ��ӰƬ and a member of the Warwick Centre for Predictive Modelling, where his work focuses on machine-learned interatomic potentials (MLIPs), atomistic simulations, and the mechanics of materials failure.

Before joining Warwick, Fraser studied Natural Sciences at the University of Cambridge, graduating with First Class honours and receiving multiple academic prizes, including the Wheatley Prize and the Part IB Physics Practical Prize.

His research combines computational physics, applied mathematics, and machine learning to better understand how materials behave at the atomic scale. Fraser has contributed to open-source scientific software development, published research in leading journals including Physical Review Letters, and presented award-winning work on machine learning for atomistic simulations at international conferences and national competitions such as STEM for BRITAIN, where he received a silver medal in Physics.

Alongside his research, Fraser is passionate about science communication and outreach, regularly speaking at public events, student conferences, and engagement activities exploring the role of AI and machine learning in modern physics and materials science.

22/06 1pm-2pm: WCPM, Ludovic Berthier, ESPCI

James Kermode — Mon, 01 Jun 2026 09:03:45 GMT

When: 1pm - 2pm, Mon, 22 Jun '26

Networking Lunch: Outside L5, from 12:30pm - 1pm.

Title: Fast equilibration of glassy systems: Where do we stand?

Abstract: Monte Carlo simulations are widely employed to measure the physical properties of glass-forming liquids in thermal equilibrium, thus offering an efficient alternative to molecular dynamics studies of the glass transition. In both approaches however, ensuring ergodicity and proper equilibrium sampling is a difficult challenge. I will provide a brief overview of Monte Carlo studies of glass-formers to illustrate the difficulty of the sampling task. I will then define and implement a series of enhanced Monte Carlo algorithms which display a much faster approach to the desired ergodic sampling of the configuration space for this family of complex systems. I will also discuss how generative models relying on machine learning models are starting to be developed to solve the same problem.

Bio: Ludovic Berthier is a Directeur de Recherche at the CNRS, based at the Laboratoire Gulliver at ESPCI Paris. He is an internationally recognised leader in statistical physics, specialising in the theory and simulation of complex, disordered systems.

His research spans a wide range of topics at the intersection of physics and materials science, including non-equilibrium statistical mechanics, soft matter and complex fluids, and the physics of supercooled liquids and glasses. He has made particularly influential contributions to understanding the glass transition, amorphous solids, and jamming phenomena, as well as emerging areas such as active and biological matter.

Ludovic’s work combines theoretical insight with advanced computational methods to uncover universal behaviours in high-dimensional and disordered systems. He has authored numerous high-impact publications in leading journals such as Nature Materials, Physical Review Letters, Physical Review X, and PNAS, and has contributed to major review articles shaping the field, including on yielding in amorphous solids and machine learning approaches to glassy systems.

Through his research, he continues to push the boundaries of how we understand and design complex materials, both in and out of equilibrium. Find out more here: https://ludovicberthier.github.io/

29/06 1pm-2pm: WCPM, Loïc Lannelongue, Cambridge

James Kermode — Wed, 22 Apr 2026 08:54:17 GMT

When: 1pm - 2pm, Mon, 29 Jun '26

Networking Lunch: Outside L5, from 12:30pm - 1pm.

Title: The (environmental) sustainability challenge of modern computing & AI

Abstract: From genetic studies and astrophysics simulations to AI, scientific computing has enabled amazing discoveries—and there's no doubt it will continue to do so. At the same time, the resource usage (energy, water) and environmental impacts of digital (research) infrastructures are becoming impossible to ignore given the urgency of the climate crisis. So what can we all do about it? And as scientists, should we even be thinking about this? We'll break down how computing activities impact the environment, debate our collective responsibility to tackle it, and discuss the latest efforts of the Cambridge Sustainable Computing Lab to empower researchers to understand and mitigate their environmental impacts. Through the lens of the GREENER principles for environmentally sustainable science, we'll explore the challenges the research community needs to overcome to create real change in this space. It will also be a chance to highlight how the Green DiSC certification framework can support scientists and institutions in making their research more sustainable.

Bio: Dr Loïc Lannelongue is an Assistant Research Professor in Computer Science at the University of Cambridge, where he also serves as Bye-Fellow and Director of Studies in Computer Science (Part II) at Jesus College Cambridge. His work sits at the intersection of computing, sustainability, and responsible innovation.

Dr Lannelongue specialises in environmentally sustainable computing, with a particular focus on understanding and reducing the environmental impact of modern computational practices, including artificial intelligence. His research takes a multi-faceted approach, combining technical development, behavioural insights, and policy engagement to drive more sustainable scientific workflows.

His academic interests include developing tools to monitor and reduce the carbon footprint of scientific computing, contributing to sustainability frameworks and policy, and exploring the ethical implications of modern science and AI. In parallel, he works in radiogenomics, applying machine learning to integrate genomics and medical imaging data to improve understanding of cardiovascular disease.

Through his research and teaching, Dr Lannelongue is committed to advancing a more sustainable and responsible future for computational science.

Webpage: https://www.jesus.cam.ac.uk/people/loic-lannelongueLink opens in a new window