Abstracts
Keynote Speakers:
|
|
Professor Birgit Strodel Peptides Under the Microscope: From Amyloid Aggregation Mechanics to Computational Design This talk bridges fundamental aggregation mechanics with computationally driven design strategies for therapeutic development. We begin by examining amyloid-beta (A尾) aggregation through the lens of free energy landscapes. In solution, A尾 monomers occupy a funnel-to-disorder state, but upon interaction with membranes or other A尾 molecules, the landscape transforms into a funnel-to-order regime that drives structured aggregation. To overcome the severe timescale limitations of simulating full-length A尾 assembly, we employ a fragment-based sliding window approach, systematically studying 7- and 10-residue homo- and heterodimers. This reductionist strategy successfully recapitulates experimentally observed PDB fibril structures, revealing how local sequence segments dictate global assembly. Building on these structural insights, we transition to therapeutic peptide design. We introduce an in silico saturation mutagenesis protocol that systematically explores sequence space to optimize peptide-protein interactions. This approach is demonstrated through the design of peptide inhibitors targeting SARS-CoV-2 proteases. Crucially, while this computational pipeline screens hundreds of candidate peptides, it narrows the selection so effectively that fewer than ten variants need to be synthesized and tested in the wet lab. The methodology exemplifies how a mechanistic understanding of molecular recognition directly accelerates and streamlines therapeutic development. |
 |
Professor of Fluid and Suspension Dynamics, University of Edinburgh Computer Simulations in Science and Engineering - What Can We Learn, and How Can They Help Us?The past decades have provided us with an impressive range of numerical algorithms and an enormous increase in available compute power. Computer simulations have since established themselves as a third research pillar beside theory and experiment. In this talk I will explore the pitfalls and opportunities of computer simulations. As a specific example, I will dive into the field of microfluidics which offers solutions to pressing problems in industry and healthcare. Due to the complexity and interplay of underlying mechanisms and geometries, it is notoriously difficult to reliably predict outcomes and design microfluidic devices. I will show how computer simulations can help generate understanding and facilitate the design process. I will conclude my talk with the argument that science works best when simulations go hand in hand with theory and experiments.
Bio
Timm Kr眉ger is Professor of Fluid and Suspension Dynamics and Head of the Institute for Multiscale Thermofluids in the School of Engineering at the University of Edinburgh. He obtained his PhD in Physics from Bochum University in 2011. After postdoctoral positions in Eindhoven and London, he became a Chancellor's Fellow at the University of Edinburgh in 2013 where has been working since. Timm enjoys research at the interface of fundamental understanding and application, bringing together experimentalists and modellers in the process. He is the lead author of a popular textbook about the lattice-Boltzmann method which has established itself as the standard reference in the field.
|
Student presentations:
| Matt Christensen | Cohort 5 Year 3 |
Quantum Algorithms for Plasma and Fusion Physics | Quantum computing presents a novel method for tackling the computational challenges inherent in kinetic plasma physics. Perhaps one of the most challenging systems is that of the Vlasov-Poisson system due to its high dimensional phase space. We employ a Fourier-Hermite expansion of the distribution function [1] to obtain a nonlinear ODE in terms of the coefficients of expansion, and examine the viability of Carleman embedding [2] to produce a quantum algorithm. We reformulate the convergence conditions of the embedding process in terms of the physical plasma properties, and show the regimes of convergence with respect to the collision frequency of the plasma. [1] Frederick C. Grant and Marc R. Feix. 鈥淔ourier-Hermite Solutions of the Vlasov Equations in the Linearized Limit鈥�. In: The Physics of Fluids 10.4 (Apr. 1, 1967), pp. 696– 702. issn: 0031-9171. doi: 10.1063/1.1762177. url: https://doi.org/10.1063/1.1762177 [2] Jin-Peng Liu et al. 鈥淓fficient quantum algorithm for dissipative nonlinear differential equations鈥�. In: Proceedings of the National Academy of Sciences 118.35 (Aug. 31, 2021), e2026805118. doi: 10 . 1073 / pnas . 2026805118. url: https://www.pnas.org/doi/10.1073/pnas.2026805118 |
| Nojus Plung臈 | Cohort 5 Year 3 |
Deep energy modelling of phase-field fracture |
This talk presents a variational physics-informed deep learning framework for phase-field fracture modelling. The work is motivated by the high computational cost of conventional phase-field solvers and by the broader aim of developing efficient neural approaches for complex fracture simulations. The study examines how physics-based neural networks learn fracture processes directly from the governing variational energy functional.
|
| Philip Jones | Cohort 5 Year 3 |
What's that made of? Modelling muonic X-ray radiation for quantitative elemental analysis | Muonic x-ray emission spectroscopy is a technique typically used for identifying elemental composition of a given sample in a non-destructive manner. It is desirable to have a robust way of computationally modelling these experiments, to allow for simpler and more systematic identification of elemental X-ray intensities, particularly in samples consisting of multiple elements. Currently, muonic x-ray energies can be computed accurately using MuDirac, but the intensities are not correct, due to several pieces of missing physics. To get the intensities correct, several features of the cascade must be considered: the initial angular momentum distribution of the muon at capture, radiative transitions, Auger electron emission, and electron refilling. This work outlines current techniques for calculating intensities, and how modern computational and theoretical techniques can be applied to improve the intensities. |
| Roman Shantsila | Cohort 5 Year 3 |
From Automation to Application: Machine-Learned Potentials for Building and Modelling Battery Interfaces. | We examine interfaces formed between lithium thiophosphate (LPS) solid electrolytes and sulphur cathodes at the atomistic scale. A detailed description of these interfacial systems is a key step in realising All Solid State Batteries (ASSBs) and Lithium-Sulphur batteries (Li-S), which address safety and energy-density concerns of lithium-ion chemistries. The performance and stability of these batteries are dictated by the complex degradation pathways at these key interfaces. In this work, we demonstrate how modern Machine Learned Interatomic Potentials (MLIPs), specifically MACE, can be efficiently trained on high-accuracy Density Functional Theory (DFT) data to reliably model the interfacial chemistry and degradation seen at the LPS-S interface. We present how we automatically generate these datasets, efficiently sampling system-relevant simulations such as molecular dynamics (MD) and nudged elastic band (NEB) calculations, further showing how choices in dataset curation impact the accuracy of the resulting models in their ability to describe lithium diffusion and interface degradation. |
| Valdas Vitartas | Cohort 5 Year 3 |
Uncertainty-Aware Machine Learning Hamiltonians with On-Demand SCF Refinement |
The use of machine learning (ML) in computational sciences has grown in recent years, including machine learning interatomic potentials (MLIPs) that predict the energy as a function of the geometry and composition of materials. However, the prediction of the energy alone is insufficient in application scenarios where an explicit description of the electronic structure is required, such as in materials screening for desired electronic properties. In this context, ML models have increasingly shifted towards predicting lower-level, solver-intrinsic quantities, such as the self-consistent Kohn-Sham Hamiltonian. On the other hand, extracting observables from an ML Hamiltonian requires additional, often expensive postprocessing, such as band structure calculations. In this work, a workflow is presented that injects ML Hamiltonians back into the electronic structure solver, enabling the use of optimised diagonalisation and postprocessing routines, and the ability to iteratively refine the Hamiltonian with additional self-consistent field (SCF) iterations. Furthermore, the uncertainty of the Hamiltonian prediction is quantified and propagated to the eigenvalues, yielding uncertainty-aware quantities of interest. When the uncertainty exceeds a prescribed threshold, iterative SCF refinement is triggered to control the risk associated with ML inference. The workflow is applied to investigate the compositional dependence on the electronic properties of nitrogen-doped graphene. Additionally, the analysis quantifies the coverage of predicted uncertainties and the time-to-solution gains compared to conventional density functional theory. |
| YC Wong | Cohort 5 Year 3 |
Are we getting very close to good long range machine learning models? | Machine learning interatomic potentials (MLIPs) have advanced rapidly in recent years, yet long-range electrostatic interactions present a fundamental challenge for MLIPs, as most models are semilocal by construction and neglect Coulombic interactions entirely. In this talk, I present a systematic benchmark of three recently proposed long-range MLIPs — MACE-POLAR [1], MACE-LES [2], and LOREM [3], designed to isolate electrostatic performance from the many competing contributions in first-principles total energies. Using strontium titanate with Sr--O Schottky vacancy pairs as a physically motivated test system, I describe a benchmarking progression from idealised point-charge systems to DFT-derived data and discuss what the results reveal about the current capabilities and limitations of each approach. [1] Batatia, I. et al. "MACE-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry." arXiv 2602.19411 (2026). [2] Cheng, B. Latent Ewald summation for machine learning of long-range interactions. npj Comput Mater 11, 80 (2025). [3] Rumiantsev, E. et al. "Learning Long-Range Representations with Equivariant Messages." arXiv 2507.19382 (2025). |
| Yuji Go | Cohort 5 Year 3 |
Modelling transport phenomena in complex electronic structures |
This talk will be about modelling various forms of transport in complex materials. First, I will discuss electronic transport in thermoelectric materials. Using the Boltzmann transport formalism, I will demonstrate how complex electronic structures, working with various scattering mechanisms, can control its transport properties Next, I will introduce my ongoing work on Spintronic devices, looking into the transport of both spin and charge in candidate materials. This work, still in its initial stages, aims to model the magnetic transport of materials, using the Boltzmann transport formalism. I begin by considering MnSb, a ferromagnetic metal with applications as a hole-spin injector and a magnetic tunnel junction. |
| Zahra Bhatti | Cohort 5 Year 3 |
Modeling excited states for the simulation of UV/Vis spectra |
Photoactive molecules have a wide range of applications in photocatalysis, drug design, molecular sunscreens, and photovoltaics. Understanding how these molecules absorb light is therefore a key step in the discovery of new materials. The surrounding solvent can significantly influence the photochemical properties of a molecule and should therefore be included in spectroscopic simulations. Because solvent fluctuations stabilise a diverse range of solute geometries, accurately reproducing experimental conditions requires extensive sampling of solute–solvent configurations. Owing to both the size of these systems and the number of configurations required, machine-learning interatomic potentials (MLIPs) provide an efficient route to generating the large ensembles needed for simulation. This talk will explore the various physical effects that contribute to these challenges and discuss strategies for addressing them. We will examine how excited states can be classified as bright or dark, helping to identify which states contribute most strongly to a spectrum. We will also consider the role of electronic–nuclear coupling and how dynamical information can be incorporated into spectral simulations. Finally, we will look at how to develop training procedures that combine information from nearby states, to mimic how their properties are physically correlated, and guide our training to reproduce the key spectral features. |
| David Bewicke | Cohort 6 Year 2 |
Simulating Exoplanet Magnetospheres | Magnetospheres are commonly depicted as unconditionally protective. However, there is a weak spot in the form of an additional energy source in the upper atmosphere: Joule heating. Along with directly heating the thermosphere, this energy input may increase ionospheric outflow and lead to mass loss, affecting observable atmospheric properties. With the PLATO and Roman space observatories expected to launch within the next year, there is a new opportunity to consider this facet of habitability for exoplanets spanning Earth-to-Neptune masses and orbital separations. We develop a methodology for using 3D magnetohydrodynamic simulations to quantify Joule heating rates for exoplanets covered by these new observatories, and discuss preliminary results. |
| Facundo Costa | Cohort 6 Year 2 |
Data Driven Alpha Titanium Dislocation Glide |
Titanium and its alloys have exceptional strength to weight ratio and corrosion resistance, making them important for turbine blades, airframes and biomedical implants. Reliable design depends on understanding the underlying plastic deformation mechanisms. In this presentation, we cover the current issues facing of state of the art methods to capture the correct 1/3<11-20> screw dislocation core spreading, which is the most active form of plastic deformation in the hexagonal closed packed phase. The Titanium 1/3<11-20> screw has been shown to have two competing core structures, a pyramidal spread ground state, and a metastable prismatic core. The prismatic core has a much lower glide barrier, resulting in a jerky unlocking-locking mechanism in dislocation glide. We use CASTEP to generate a defect database consisting of gamma surfaces and quadrupolar cells to train ACE models to study these competing structures. We achieve similar energetics differences on core spreading with our potential to established DFT benchmarks. Furthermore, we find that the metastable structure becomes unstable in our ACE under tight relaxation. To address this, we introduce a leverage-based loss function that targets geometrically sparse atomic regions to reduce force errors in defect configurations. We share preliminary results of this approach applied to silicon point defects, demonstrating it as a promising pathway to improve the structural stability of the titanium dislocation cores. |
| George McKay | Cohort 6 Year 2 |
Error bounds in weakly symmetric quantum trajectories |
Quantum Jump Monte Carlo (QJMC) provides a stochastic representation of Lindblad dynamics in open quantum systems, enabling connections between non-equilibrium quantum dynamics and large-deviation structures analogous to equilibrium statistical mechanics. Although QJMC avoids the full density-matrix scaling of Lindblad evolution, the underlying Hilbert space still grows exponentially with system size, limiting the exact simulation of many-body systems. In this talk, we will discuss recent symmetry-adapted approaches to advance QJMC simulation based on weak symmetries of Lindblad dynamics. After reviewing the abelian case, we will describe ongoing work extending these constructions to non-abelian weak symmetries, in particular, to permutation symmetry. Finally, we derive error-bounds for both local and steady-state uncertainty of weakly symmetric unravellings. |
| George Simmons | Cohort 6 Year 2 |
Physics Informed Priors for Machine Learning Interatomic Potentials |
Machine learning interatomic potentials (MLIPs) give us access to longer timescales and larger scale atomistic simulations compared to ab initio molecular dynamics. This is achieved by learning the potential energy surface (PES) across configuration space as predicted by DFT, which is commonly done using linear models, and in our case using the Atomic Cluster Expansion framework employed in ACEpotentials.jl. However, accuracy on test data is not a guarantee for physical predictions for quantities of interest. To combat unphysical predictions and holes in the PES, active learning is a popular approach to improve configuration space coverage. But in the context of linear machine learning models, priors can be exactly enforced and so we can use indicator functions in parameter space as our priors to enforce physical constraints during model training.
We have formulated our problem as a quadratic programming (QP) problem with linear constraints, where the QP part is the analytical least squares minimisation of our model and the linear constraints correspond to our priors (e.g. Born elastic stability constraints, repulsive core interactions). This improves the predictive power of MLIPs as we can ensure stability across a wider range of conditions without having to evaluate new and expensive atomic configurations using DFT.
|
| Keyi Wei | Cohort 6 Year 2 |
Structural and Mechanistic Insights into Gram-Negative Bacterial Pathways | The outer membrane of Gram-negative bacteria forms a major permeability barrier to antibiotics, making its assembly and biogenesis important targets for antimicrobial development. The enterobacterial common antigen (ECA) is a conserved surface polysaccharide in Enterobacterales, composed of lipid-linked repeating units of three amino sugars. ECA is produced in three forms: two surface-associated variants, ECA_PG and ECA_LPS, and a periplasmic cyclic form, ECA_CYC. Despite its long-standing identification, the membrane protein machinery responsible for ECA biosynthesis remains incompletely defined. In particular, the molecular mechanisms governing the initiation, membrane translocation, and assembly of ECA remain poorly understood. Emerging evidence suggests that ECA plays important roles in bacterial physiology and envelope integrity, highlighting the need for a mechanistic understanding of its biogenesis. This work currently focuses on the membrane proteins involved in the early stages of ECA biosynthesis, with particular emphasis on WecA, the initiating glycosyltransferase, and WzxE, the flippase responsible for translocating the lipid-linked trisaccharide from the cytoplasmic to the periplasmic side of the inner membrane. By examining their interactions, dynamics, and substrate-binding determinants, this study aims to provide structural and mechanistic insight into the initiation and membrane translocation steps of ECA assembly. The talk will also introduce how nested sampling can be applied to biological systems to explore conformational landscapes and mechanistic pathways. |
| Luca Gasparro | Cohort 6 Year 2 |
Computing the Glass Transition Temperature of Hydrophobic and Hydrophilic Painkillers via Molecular Dynamics Simulations |
Poorly water-soluble make up 40% of the current pharmaceutical market. The effectiveness of water-soluble drugs is severely limited by their low bioavailability. A solution to this problem is formulating such drugs as amorphous solids, as the amorphous state exhibits enhanced solubility with respect to its crystalline counterpart [1]. However, amorphous solids are metastable with respect to the crystal, thus tending to re-crystallise over pharmaceutically relevant time-scales, causing their storage to be problematic. The physical stability of amorphous formulations is influenced in a non-trivial fashion by their glass transition temperature, Tg [2]. Our work utilises molecular dynamics simulations to identify and understand the Tg of naproxen as well as paracetamol – two prototypical painkillers, the former being a hydrophobic molecule and the second being a hydrophilic molecule instead. In particular, we wanted to investigate the impact of different water contents, mimicking the impact of real life humidity in terms of storage conditions, on Tg. We calculated the glass transition temperatures of these drugs by means of a thorough computational approach, finding that the Tg of naproxen exhibits minima with respect to the weight percentage of water. Paracetamol on the other hand exhibits a monotonically decreasing Tg with water weight content, consistent with the expected plasticising effect of water. Through dynamical and structural analysis, we attribute the non-standard dependence of naproxen鈥檚 Tg with water content to hydrophobically driven water clustering. This work lays the foundation for future studies on the physical stability of amorphous solid dispersions, an attractive alternative to pure amorphous formulations. [1] N.S. Krishna Kumar, R. Suryanarayanan, Mol. Pharm. 19 (2022) 472–483. [2] Roudaut G, Simatos D, Champion D. Roudaut G, Simatos D, Champion D, et al. Innov Food Sci Emerg Technol. 5 (2004) 127–134. |
| Luca Seaford | Cohort 6 Year 2 |
Learning Grand-Canonical Free-Energy Surfaces for Au Electrodeposition on Graphene Bilayers |
Electrochemical nucleation and growth underpin processes ranging from electrocatalyst synthesis and battery degradation to corrosion, sensing, and nanoscale materials fabrication. However, understanding of the exact mechanisms governing nucleation, cluster growth, and morphological changes at the atomistic scale remains elusive. Ab initio electronic structure methods, like the Kohn-Sham formulation of density functional theory (KS-DFT), are often poorly suited for modelling fixed-potential electrochemical interfaces and are typically limited to length and time scales several orders of magnitude smaller than those accessible using existing in-situ and ex-situ S-TEM experiments. Recent work has, therefore, sought to augment existing electronic structure approaches with machine-learned interatomic potentials (MLIPs) and improved solvation techniques to bridge the gap between in silico computations and real-world experiments.
This work outlines a framework for training bias-aware MACE MLIPs using data obtained from grand-canonical linear-scaling KS-DFT calculations applied to Au electrodeposition and cluster aggregation on graphene bilayers. After briefly covering the existing grand-canonical ensemble-DFT (GC-eDFT) capability in the ONETEP linear-scaling electronic structure package, I will describe the construction of potential-dependent training datasets from Au cluster configurations labelled with vacuum GC-eDFT energies and forces. Selected results from the training and evaluation of the models will be presented, including Pareto-front investigations of accuracy, computational cost, and dynamical performance, alongside benchmarking against current MACE foundation models.
|
| Maksymilian Diakiewicz | Cohort 6 Year 2 |
Modelling supercooling dynamics of phase change materials. | Heat batteries store and release thermal energy efficiently, helping to balance temporal mismatches between heat supply and demand in civil and industrial applications. At their core are phase change materials (PCMs), which absorb and release energy during the melting and solidification processes. During discharge, the PCM can become supercooled; liquid remains below its melting temperature and enters a metastable thermodynamic state that can delay or disrupt solidification and thus impair system performance. In this talk, we present an overview of heat battery devices and study a reduced toy model of a finite supercooled PCM system across different parameter regimes. We compare exact numerical solutions with asymptotic predictions for key quantities, including solidification speed, total freezing time, and energy release. |
| Rumesh Sudhaharan | Cohort 6 Year 2 |
On thin ice: modelling and simulation of an unstable interface | Understanding the dynamics of ice formation is important in many real life applications. An example is ice accretion on aeroplane wings, where liquid droplets below the melting point freeze to form ice crystals on the wings. Modelling ice growth in such conditions presents a challenge, as we have a Stefan-type problem, with a moving boundary and an unstable ice-water interface. The ice crystals form dendritic shapes due to an instability at this interface, called the Mullins-Sekerka instability. We attempt to mathematically model and build an associated numerical simulation to study this instability by solving the two-dimensional heat equation using the Finite Element Method. We use an interface tracking method to investigate the shape of the interface and study its evolution with time. We verify our numerical results to ensure robustness and mesh independence, as well as validate against analytical solutions derived based on a linear stability analysis procedure. |
| Sean Wu | Cohort 6 Year 2 |
Towards Correlative 4D-STEM Ptychography and EELS for Cryogenic Single-Particle Analysis | Cryogenic electron microscopy has enabled atomic resolution structural determination of biological macromolecules. However, phase contrast imaging alone does not reveal elemental composition or other valuable chemical information. Electron energy loss spectroscopy (EELS) provides rich vibrational and chemical details, integrating it with high-resolution structural imaging for correlative analysis of biological specimens remains a major challenge, as conventional acquisition modes are often mutually exclusive. In this study, we focused on a simulated hollow detector geometry designed to allow simultaneous acquisition of diffraction patterns and EELS spectra within a single dose budget. Using a low-dose apoferritin dataset, we demonstrate sub-nanometre 3D resolution in single particle analysis, even with a simulated hollow angle of 55% of the convergence angle and approximately 30% of the total dose allocated to EELS. Finally, we propose two theoretical optical geometries to enable correlative 4D-STEM with EELS: (1) sequential acquisition via a post-specimen beam deflector, and (2) simultaneous acquisition using a hollow detector. Furthermore, we outline a data-processing workflow that integrates ptychography with EELS for correlative, energy-resolved 3D mapping of biological macromolecules, paving the way toward unified structural, chemical, and vibrational analysis of biological specimens. |
| Swathi Mahashetti | Cohort 6 Year 2 |
Modelling Point Defects and their properties in Perovskite bulks, surfaces and interfaces |
Perovskite-based solar cells have emerged as a leading candidate for next-generation clean energy technologies; however, persistent stability concerns continue to impede their commercialisation. FAPbI鈧� exhibits exceptional thermal stability and a near-optimum bandgap, positioning it as a prime candidate for high-efficiency single-junction photovoltaic devices. Nevertheless, ionic migration driven by point defect populations in the bulk material remains a critical bottleneck for long-term device durability. This work presents a comprehensive ab initio and atomistic investigation of defect formation energetics in bulk FAPbI鈧�, elucidating the role of intrinsic point defects in facilitating ionic diffusion and quantifying the migration energy barriers of the most prevalent defect species. In addition to this, the surface properties of FAPbI鈧� are examined with the goal of forming stable interfaces with charge transport layers, as encountered in operational solar cell architectures. |