Abstracts

J.Nathan Kutz, University of Washington
Automated Learning of Physics from Data

Abstract: Sensing is a universal task in science and engineering. Downstream tasks from sensing include learning dynamical models, inferring full state estimates of a system (system identification), control decisions, and forecasting. These tasks are exceptionally challenging to achieve with limited sensors, noisy measurements, and corrupt or missing data. Existing techniques typically use current (static) sensor measurements to perform such tasks and require principled sensor placement or an abundance of randomly placed sensors. In contrast, we propose a SHallow REcurrent Decoder (SHRED) neural network structure which incorporates (i) a recurrent neural network (LSTM) to learn a latent representation of the temporal dynamics of the sensors, and (ii) a shallow decoder that learns a mapping between this latent representation and the high-dimensional state space. By explicitly accounting for the time-history, or trajectory, of the sensor measurements, SHRED enables accurate reconstructions with far fewer sensors, outperforms existing techniques when more measurements are available, and is agnostic towards sensor placement. In addition, a compressed representation of the high-dimensional state is directly obtained from sensor measurements, which provides an on-the-fly compression for modeling physical and engineering systems. Forecasting is also achieved from the sensor time-series data alone, producing an efficient paradigm for predicting temporal evolution with an exceptionally limited number of sensors. In the example cases explored, including turbulent flows, complex spatio-temporal dynamics can be characterized with exceedingly limited sensors that can be randomly placed with minimal loss of performance.

Bio: Nathan Kutz is the Boeing Professor of AI and Data-Driven Engineering in the Department of Applied Mathematics and Electrical and Computer Engineering and Director of the AI Institute in Dynamic Systems at the University of Washington, having served as chair of applied mathematics from 2007-2015. He received the BS degree in physics and mathematics from the University of Washington in 1990 and the Phd in applied mathematics from Northwestern University in 1994. He was a postdoc in the applied and computational mathematics program at Princeton University before taking his faculty position. He has a wide range of interests, including neuroscience to fluid dynamics where he integrates machine learning with dynamical systems and control.

Research Interests: Numerical methods and scientific computing, data analysis and dimensionality reduction (PCA, POD, etc) methods, dynamical systems, bifurcation theory, linear and nonlinear wave propagation, perturbation and asymptotic methods, nonlinear analysis, variational methods, soliton theory, nonlinear optics, mode-locked lasers, fluid dynamics, Bose-Einstein condensation, neuroscience, gesture recognition and video & image processing

Natalia Martsinovich, University of Sheffield

Computational modelling of materials: from fundamental properties to evaluation of photocatalytic efficiencies

Abstract:Computational modelling is firmly established as part of the research toolkit for investigating properties of materials, ranging from prediction of fundamental properties and structure-property relationships to evaluating the suitability of materials for various applications, such as energy storage, sensing and catalysis. In this talk I will present examples of our research, starting with the studies of fundamental properties of graphene-based materials, then addressing the role of graphene in improving the efficiency of TiO₂/graphene composite photocatalysts, and finally evaluating the mechanisms and effectiveness of TiO₂-based photocatalytic systems for methanol reforming.

Graphene’s excellent electronic and optical properties enable its use in a variety of applications, such as electronics and sensing. We collaborated with experimental partners to investigate oxygen-functionalized graphene-based materials, with the ultimate aim to design graphene-based sensors to detect phosphate, an essential plant nutrient. We used density-functional theory (DFT) calculations to investigate how chemical functionalization affects electronic properties of graphene. Oxygenation was found to have a significant effect on graphene’s properties, with band gap opening in functionalised graphenes with oxygen content above 6%. This tuning of graphene’s properties by chemical functionalization can lead to new applications in electronics and sensors [1].

One more application of graphene is in photocatalysis, where incorporation of graphene is known to increase the photocatalytic efficiency of TiO₂-based photocatalysts. Photocatalysts harvest sunlight to create photogenerated charges, which drive chemical reactions, such as degradation of pollutants and production of “green” hydrogen. We used DFT to investigate the electronic properties of composites of TiO₂ with graphene, graphene oxide (GO) and reduced graphene oxide (RGO). We found that TiO₂/RGO composites can form trap states that can prevent recombination of photogenerated electrons and holes [2]. Avoiding recombination of photogenerated charges results in improved utilisation of absorbed photons, and therefore in improved photocatalytic efficiencies.

Beyond the electronic properties, mechanisms and rates of photocatalytic reactions are some of the key factors that determine the efficiency of photocatalytic processes. In a collaborative study, we combined DFT with microkinetic modelling to study the mechanism of methanol conversion to H₂ and oxidation products using a TiO₂ photocatalyst and transition metal co-catalysts. A reaction network was built to encompass all reaction steps, from reactants to multiple possible intermediates and products. Microkinetic modelling was used to evaluate the rates of these reaction steps and to assess the effectiveness of Pt, Pd, Au and Ag transition metal co-catalysts [3], showing strong predictive capability of this method.

References

1. Yong et al. ACS Appl. Nano Mater. 2024, 7, 18386−18397.
2. N. O. Gillespie, N. Martsinovich, ACS Appl. Mater. Interfaces 2019, 11, 31909−31922.
3. J. Wigglesworth et al. Chem Eng. J., submitted (2025).

Dr Natalia Martsinovich obtained her first degree in Chemistry from the Belarusian State University in 2000. She then obtained a PhD in Theoretical Chemistry from the University of Sussex in 2005, where she also worked as a temporary Lecturer in Physical Chemistry in 2003-04.

She was a postdoctoral researcher in the Department of Physics at King’s College London (2004-08), and in the Department of Chemistry at the University of Warwick (2008-13). In 2013 she was appointed Lecturer at the University of Sheffield.

Research interests: My research is focussed on studying the properties of materials and surface-adsorbate interfaces and processes taking place at these materials and interfaces. Important applications include photovoltaics and photocatalysis. We use a range of theoretical methods, mainly density-functional theory, and also charge transfer theory and molecular mechanics.

Cohort 4

Session 1

Computational atomic-level investigation of the Li-metal/Li-argyrodite interface stability

All solid-state batteries (ASSBs) have obtained increasing interest as next-generation battery technology due to their increased safety and the potential of enhanced energy density when combined with a Li-metal anode. However, chemical and electrochemical instabilities at the interfaces can limit the overall electrochemical performance of the battery. Computational modelling can be used to gain a better understanding of the diffusion behaviour at solid-solid interfaces and how electrolyte degradation affects Li-ion diffusion. In this talk, I will detail the methodology we used to create and analyse representative atomistic interface models of the Li-metal/Li-Argyrodite interface. To set up the models, we first found the three lowest energy surface structures, taking into account all symmetric surface terminations defined by low Miller indices (0,1), and then used our in-house code INTERFACER to create three interfaces with Li-metal. We conducted some initial analysis of the interfaces using ab-initio molecular dynamics (MD) simulations, showing that all three interfaces are unstable, as previously reported by Golov et al.1 and Cheng et al.2 The main reaction mechanism is the reduction of P (5+ to 3-) by Li-metal, leading to a breakdown of PS₄³- polyhedra and the formation of a solid-electrolyte interface layer (SEI) containing Li₃P, Li₂S and LiCl. To reach longer simulation lengths and consider larger models, consisting of ~2000 atoms, we trained machine learned force field (MLFF) models of our interfaces using MACE, a message passing neural network.3 The resulting trajectories allow for a more realistic and comprehensive insights into the SEI formation and growth, and its effect on the Li-diffusion across the interface, collected over a longer time scale. 1. Golov and J. Carrasco, ACS Applied Materials and Interfaces, 2021, 13, 43734−43745. 2. T. Cheng, B. V. Merinov, S. Morozov and W. A. Goddard, ACS Energy Letters, 2017, 2, 1454–1459. 3. I. Batatia, D. P. Kovacs, G. N. C. Simm, C. Ortner and G. Csanyi, Advances in Neural Information Processing Systems, 2022.

Machine Learning Excited State Potential Energy Surfaces of Solvated Nile Red with ESTEEM

Machine Learned Interatomic Potentials (MLIPs) offer a powerful combination of abilities for accelerating theoretical spectroscopy calculations utilising both ensemble sampling and trajectory post-processing for inclusion of vibronic effects, which can be very challenging for traditional ab initio MD approaches. We demonstrate a workflow that enables efficient generation of MLIPs for the solvatochromic dye nile red system, in a variety of solvents. We use iterative active learning techniques to make this process as efficient as possible in terms of number and size of DFT calculations. Additionally, we compare the efficacy of various methodologies: generating distinct MLIPs for each adiabatic state, using one ground state MLIP in combination with delta-ML of excitation energies, and using a three-headed multiheaded ML model. To evaluate the validity of the resulting models, we compare predicted absorption and emission spectra to experimental spectra.

Session 2

Metal-organofluorine interactions and their role in selective metal deposition for next-gen photovoltaics

Controlled deposition of metals onto a given surface is a slow and costly process, but is essential for electronics and photovoltaics. Recently, a novel method for selective deposition has been discovered by our experimental collaborators, the Hatton group, which uses a thin layer of specific organofluorine compounds to prevent metal atoms from being adsorbed onto a surface. We leveraged density functional theory to investigate the nature and strength of the interaction between various metals and organofluorines, and to determine the extent to which the direct interactions between a metal atom and an organofluorine molecule affect this process of selective deposition. In particular, we found that the condensation of zinc onto a surface can be made highly selective, as its interactions are extremely weak and dominated by dispersion forces.

Thermodynamically Informed Optimal Autonomous MLIP Database Generation

We present an optimal method of database generation for the training of machine learned interatomic potentials (MLIPs). Nested sampling is an unbiased Potential Energy Surface (PES) sampling technique that produces samples across all phases given no prior information. Since the accuracy of any MLIP depends on the underlying data it is trained on, and the data is required to undergo high cost ab-initio evaluation, selecting the fewest and most important data-points is a critical component in developing MLIPs efficiently. Samples generated by nested sampling form a sparse mesh of thermodynamically relevant points of the PES which creates a potent, low cost database that can be iteratively expanded through successive sampling runs. Based on the Atomic Cluster Expansion (ACE) we suggest a highly automated framework and, with our method, we reproduce fundamental properties of pure magnesium (vibrational and elastic properties, phase diagram, 0K enthalpy curves) with remarkably small databases.

Session 3

Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning

Machine-learned interatomic potentials are revolutionising atomistic materials simulations by providing accurate and scalable predictions within the scope covered by the training data. However, generation of an accurate and robust training data set remains a challenge, often requiring thousands of first-principles calculations to achieve high accuracy. Foundation models have started to emerge with the ambition to create universally applicable potentials across a wide range of materials. While foundation models can be robust and transferable, they do not yet achieve the accuracy required to predict reaction barriers, phase transitions, and material stability. This work demonstrates that foundation model potentials can reach chemical accuracy when fine-tuned using transfer learning with partially frozen weights and biases. For two challenging datasets on reactive chemistry at surfaces and stability and elastic properties of tertiary alloys, we show that frozen transfer learning with 10-20% of the data (hundreds of datapoints) achieves similar accuracies to models trained from scratch (on thousands of datapoints). Moreover, we show that an equally accurate, but significantly more efficient surrogate model can be built using the transfer learned potential as the ground truth. In combination, we present a simulation workflow for machine learning potentials that improves data efficiency and computational efficiency.

A Multiscale Approach to Modelling Dislocation Loops in Tungsten

In order to predict the long-term effects of irradiation on the material properties of tungsten, a continuum approach to simulating the interactions of dislocation loops, which arise from radiation damage is required. Contemporary continuum models of stress fields from dislocation loops are tricky to handle due to the presence of singularities near the core of the dislocations. We discuss a formula which handles such singularities whilst providing accurate predictions for the far-field interactions between loops. The model is verified using atomistic simulations to ensure that it is informed by lower-length scale phenomena and that the physics of the problem is correctly captured. We present the current model and its advantages, and show that predictions produced by atomistic simulations agree quite well with the far-field limit of the continuum model far enough away from boundary effects. In particular, we robustly demonstrate that the decay rate of atomistic results and continuum results coincide with one another as the size of the atomistic simulations approach the far-field limit.

Session 4

Ion transport via a travelling wave in a cylindrical guide

Electrostatic travelling waves have found many applications in ion transport instruments, including ion mobility separation in mass spectrometers. It is preferred over traditional direct current (DC) and multipole ion guides due to its capability of transport without high voltages to sustain an axial potential gradient. However, such transport mechanisms complicate ionic motion due to time-dependent forces, including de-focusing effects causing deposition of ions at the ion guide walls, leading to transport efficiency loss. It is therefore instructive to study the trajectories resulting from travelling carrier waves for understanding and refinement of ion transport. In this work, we analyse ionic motion in a confining travelling-wave based electric field in a cylindrical guide, demonstrating unseen effects on the trajectories due to geometry. We find de-focusing, surfing and slow-roll behaviour arising from the travelling wave, with the mean drift velocity profile varying in accordance to penetration distance of the field, which agrees on-axis with previous work.

Data-driven equation discovery for liquid film flows

Partial differential equation (PDE) discovery is an exciting alternative to standard first principles-based methodologies regularly used in mathematical modelling, particularly in regimes outside the reach of traditional approaches. This talk explores PDE discovery methods, with their application to liquid film flows as central motivation, drawing on synthetic data and direct numerical simulation results where appropriate. To begin with, we introduce the sparse identification of nonlinear dynamics (SINDy) equation discovery method, enhanced with ensemble learning techniques. We then focus our attention on the recovery of linear equations and nonlinear test cases that are commonly used to validate equation discovery frameworks. This is before looking to what we believe to be a novel application of equation discovery: a highly nonlinear, single equation thin-film model. We discuss challenges for other established thin film equations, highlighting important derivation aspects to build analytical understanding that enhances the data-driven process. Finally, we look to steer the developed framework towards new regimes of interest, such as thick liquid film flows, in which classical physical understanding is well complemented by the present approach.

Session 5

New Methods for Atomistic Simulations in Materials Science

This talk will be in two halves. The first half will be on ML-MIX, a package developed for the molecular dynamics software LAMMPS to accelerate simulations using QM/MM style force-mixing of machine learned interatomic potentials. The method itself will be presented, as well as a number of simple demonstration applications. The second half will be on using numerical continuation to model plasticity in amorphous carbon. In glasses, plasticity occurs through rearrangements of localized regions that flip discretely between states, so called `shear transformations’. In amorphous carbon specifically, shear transformations are the breaking or forming of single bonds, with much of the deformation manifesting as ‘avalanches’ of multiple of these single-bond breaking events. Via the application of a numerical continuation method, I will show that avalanche events, previously considered inseparable cascades, can be split into constituent smaller hops between basins, each with its own energy barrier. The insight now offered into the structure of these crucial processes has the potential to significantly contribute to the mechanistic understanding of plasticity in amorphous materials.

Hollow Multi-Slice Electron Ptychography for Simultaneous 3D Structural Imaging and EELS in 4D-STEM

Recent developments in 4D scanning transmission electron microscopy (4D-STEM) and electron ptychography have opened new frontiers in atomic-resolution imaging. However, integrating structural imaging with electron energy loss spectroscopy (EELS) remains a significant challenge due to detector geometry constraints. In this talk, we demonstrate that multi-slice electron ptychography using a hollow-cone pixelated detector enables simultaneous high-resolution 3D structural imaging and spectroscopic analysis from a single 4D-STEM acquisition. Using both experimental data and multi-slice simulations on PrScO₃, we show that up to 75% of the central diffraction signal can be blocked—creating angular space for EELS—without compromising picometer-level lateral or sub-nanometer depth resolution. We evaluate the effects of hollow radius, convergence angle, and dose, and demonstrate the robustness of this method across varying experimental conditions. Our results establish hollow multi-slice ptychography as a powerful, dose-efficient strategy for integrated structural and chemical characterization in advanced electron microscopy.

Predicting the committor for the 2D Ising model using graph and convolutional neural networks and optimal parallelisation strategies for flat histogram Monte Carlo sampling.

The committor, the probability that a microstate reaches one state before returning to another, enables direct calculation of key dynamic properties, notably the nucleation rate, but is often computationally expensive to obtain. This study uses the 2D Ising model as a tractable system for training Graph and Convolutional Neural Networks to predict committors. Due to the computationally inexpensive nature of the 2D Ising model it allows for high-throughput generation of microstates via Monte Carlo simulations used for nerual network training. In this study, neural network predictions are compared to the traditional collective variable, the size of the largest cluster, for estimating the committor, demonstrating the potential of machine learning to improve transition state characterization.

Flat histogram methods like Wang-Landau sampling enable high-throughput calculation of phase diagrams for atomistic and lattice models. Various parallelization schemes, including static and dynamic energy domain decomposition, replica exchange, and multiple random walkers per domain, are benchmarked individually and in combination to determine their impact on efficiency. These strategies are applied to lattice-based models of AlTiVNb and AlTiCrMo refractory high-entropy alloys.

Cohort 5

Session 1

Machine Learning Prediction of Oscillator Strengths for the Generation of UV/Vis Spectra of Explicitly-Solvated Systems

We present an approach for generating UV/Vis spectra with peaks heights corresponding to probabilities of transitions to different states after excitation, as seen experimentally. The MACE architecture was used within an active learning cycle to predict energies, forces, and transition dipole moments (TDMs) of caffeine in water. The forces allow propagation of machine-learned molecular dynamics, from which caffeine-water clusters are sampled. The oscillator strength is calculated from the TDMs to produce a final weighted KDE plot.

Learning TDMs necessitated introducing a phaseless™ loss function withing MACE. This led to extremely successful training for the first excited state and qualitative agreement with experiment. However, higher-lying excited states proved more difficult to train force and TDM predictions for; analysis of training data uncovered multiple reasons for this. Therefore, current work involves training multi-headed models with an additive term in the loss function for TDMs when states lie close in energy.

(Inter)facing the Bitter Truth: How to Design Better Interfaces in Next-Gen Batteries using Atomistic Simulations Assisted by Machine-Learning

Lithium - Sulphur and solid state batteries as a whole, are key targets in the search for new chemistries to support the expanding energy storage sector. While recently industry has begun to incorporate these materials in recent months, widespread adoption is limited by the lack of understanding of the degradation seen at solid-solid interfaces in these systems. This is especially true when looking at systems including a Sulphur cathode. Addressing this would provide a key piece of the puzzle. Machine Learned Interatomic Potentials (MLIPs), have emerged as a key tool in combining the accuracy of Density Functional Theory (DFT), and approaching the speed of classical potentials. To be of use in explaining degradation in our systems, the MLIPs must be trained and validated with caution to ensure accurate results. We examine several MLIPs,including foundational models like MACE-MP, SevenNet, and CHGNET as starting points. We also compare other purpose trained models such as VASP on the fly, and MACE potentials trained with new proposed methodologies. We will demonstrate their performance in modelling novel Lithium Thiophosphate (Li-P-S) solid electrolytes. We compare the various MLIPs, and training strategies, on key properties of interest such as thermal properties and their ability to predict energy barriers. We show that our new training strategies are competitive and show good performance in the broad range of calculations for the bulk solid electrolytes.

Surrogate Hamiltonians for Gold: From Linear Models to Deep Learning

Over recent years, many atomistic machine learning (ML) models have been proposed to predict material properties from Density Functional Theory (DFT) as a function of configuration and composition. Learned quantities include electronic properties such as the Fermi level and band gap, and atomic properties such as local energies and forces, the latter being predicted with machine-learned interatomic potentials (MLIPs). However, such models do not fully capture electronic degrees of freedom, making them less universal and only tailored to a specific property. The self-consistent electronic Hamiltonian matrix from DFT can be used to compute the ground-state charge density of the system, making it a suitable target to learn all electronic properties with a single ML model. In this work, we train Hamiltonian models on high-temperature bulk gold data and present a preprocessing strategy to reduce the dimensionality of the Hamiltonian to only include energetically relevant valence states. The surrogates were trained using a linear ACEhamiltonians model and a message-passing neural network MACE-H which we have recently developed. We discuss the inherent strengths and weaknesses of both models by illustrating their accuracy, performance, data efficiency, and their ability to predict electronic quantities of interest for out-of-distribution configurations.

Session 2

Muon Cascade Calculations

Muonic X-ray Emission Spectroscopy is a non-destructive method of elemental analysis, which has recently been adopted several fields, such as cultural heritage. 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 [1], but the intensities are not correct. In order to do this, muonic cascade calculations are performed, which is typically done using the Akylas-Vogel [2] cascade code.

There are several important features of a muonic cascade calculation: the initial angular momentum distribution of the muon at capture, radiative transitions, Auger electron emission, and electron refilling. All of these must be treated correctly to obtain correct intensities. The Akylas-Vogel cascade code treats these mechanics using techniques appropriate at the time of writing, at the expense of having several input degrees of freedom.

This talk will discuss muon cascade calculations in their current form, and how modern computational and theoretical techniques can be applied to the problem. This will include how the theory can be incorporated into a relativistic framework, in particular for radiative and Auger transitions.

[1] S. Sturniolo and A. Hillier, Mudirac: A Dirac equation solver for elemental analysis with muonic X-rays, X-Ray Spectrometry 50, 180 (2021).

[2] V. R. Akylas and P. Vogel, Muonic atom cascade program, Computer Physics Communications 15, 291 (1978).

Machine Learning Multiscale Simulation for Defects in Strongly Correlated Oxides

Strongly correlated oxides such as strontium titanate (SrTiO3), exhibit emergent phenomena including ferroelectricity, magnetism, and superconductivity, making them important to condensed matter research and energy technologies. Defects also play a key role in tuning their electronic properties, such as photoconductivity. However, modelling these defects using density functional theory (DFT) is computationally demanding. To address this, we develop surrogate machine learning interatomic potentials for SrTiO3 based on the MACE framework, capable of accurately predicting energies, forces, and stresses in both bulk and defective structures. In addition, we demonstrate task-specific models that use a sparse representation to predict Schottky defect formation energies. These defect-aware models outperform general-purpose MLIPs in both accuracy and efficiency. Beyond SrTiO3, our approach shows promise for generalising to other correlated oxides, enabling scalable and accurate simulation of defect-driven phenomena.

Quantum Carleman Linearisation of the Vlasov-Maxwell System

Quantum computing presents a novel method for tackling the computational challenges inherent in kinetic plasma physics. Perhaps one of the most challeng- ing systems is that of the Vlasov-Maxwell system due to its 3+3+1 dimensional phase space. Here we employ a Fourier-Hermite expansion of the distribution function to obtain a nonlinear PDE in terms of the coefficients of expansion, and examine the viability of Carleman linearisation to produce a quantum algorithm. We find that the dynamics of the Fourier-Hermite coefficients do not satisfy the dissipative and weakly nonlinear conditions required by Carleman linearisation to ensure error bounds in the infinite-time limit, in both the col- lisional and collisionless regimes. Nonetheless, we derive an expression for the finite time over which the errors remain bounded, as a function of the number of Fourier and Hermite modes used in the expansion.

Session 3

Global MHD and Test-Particle simulations of outer radiation belt flux drop-out events

The radiation belt electron flux can exhibit dramatic variations across a range of spatial and temporal scales, including global‐scale radial transport, mesoscale injections, and local‐scale wave‐particle interactions. Long-term radiation belt variability has been successfully captured by solving the Fokker Planck diffusion equation (e.g., BAS-RBM), incorporating radial, pitch-angle and energy diffusion and imposed upon semi-empirical Tsyganenko magnetic models. However, during geomagnetic storms, non-diffusive processes become significant. Enhancements in the partial ring current and dawn–dusk electric fields, and associated magnetic field distortions can lead to violation of the third adiabatic invariant and rapid outward radial transport. Electron flux dropouts often occur in the outer radiation belt during geomagnetic storms across several orders of magnitude over a wide range of L shells. These irreversible electron losses occur either by precipitating into atmosphere or by escaping through magnetopause. In this study, we employ global magnetohydrodynamic and test-particle (MHD-TP) simulations to investigate the dropout mechanisms by introducing an ensemble of test particles into the global MHD fields and tracking their trajectories. We aim to distinguish the relative contributions of magnetopause shadowing and wave–particle interactions in producing the observed rapid changes in electron flux.

A mathematical model for rocking bioreactors using feedback control

Cultivated meat, also known as cell-based or artificial meat, is an alternative protein source created from animal cells grown outside of their natural environment. Cells are initially grown on a very small scale in so-called seed trains. Once a sufficiently large number of cells with the desired properties have formed, they are then transferred to larger vessels known as bioreactors for proliferation (growth and multiplication). This stage, combining multi-phase fluid mechanics with advection-diffusion processes, is the focus of my research and this talk. The most common bioreactors are stirred bioreactors, where the media is agitated by stirring with propellers. Previous research has suggested that this type of reactor may create high shear stresses, which have a particularly noticeable impact for the production of cultivated meats, as mammalian cells are particularly shear sensitive. To mediate these issues, an alternative reactor design is the rocking-wave bioreactor. Here, the whole reactor, a small 5-10L bag half-filled with liquid, is gently rocked back and forth, generating waves and vortical structures in the flow which induce mixing and oxygen transfer at the surface, allowing cells to proliferate while reducing shear. I will present our mathematical model and dedicated direct numerical simulation (DNS) framework of rocking bioreactors, which is then used to develop a feedback control strategy to optimise mixing while maintaining a target value for the shear stress, aiding cell growth

Characteristic scattering exponents for electronic transport in complex thermoelectric materials from ab initio calculations

A typical way to understand the electronic transport mechanism that dominates the properties of a material, is to fit experimental data into transport models where the scattering times are treated by a simple exponential energy dependence as τ = τ0 (E/E0)r. Here, r, is a characteristic exponent which depends on the dominant scattering mechanism. For example, transport processes dominated by acoustic deformation potential (ADP), optical deformation potential (ODP), polar optical phonon (POP) and ionized impurity scattering (IIS), are considered to follow r = -0.5, -0.5, 0.5, and 1.5, respectively [1, 2]. Although these values are based on single band considerations, they are often used without sufficient justification. Modern thermoelectric (TE) materials, however, have complex electronic structures interacting via complex scattering mechanisms. To better characterize transport measurements, this scattering exponent approach needs to be re-examined, validated, and corrected. In this work we use Boltzmann Transport theory and start with re-evaluating the exponent approximation for a single parabolic band material. We then evaluate a two-band system, where the effect of the energy separation and the effective mass ratio between the two bands is examined. Finally, we investigate the exponent validity for many real materials from the half-Heusler group using ab initio DFT electronic structures and Boltzmann transport with full energy/momentum/band dependent scattering, as implemented in the code ElecTra [3]. In the case of a single band, we find that the standard literature scattering exponent values are in good agreement for ADP, only at high energies for ODP and POP, and only at low carrier densities for IIS. In the case of a two-band material, significant deviations are seen for all scattering mechanisms, which increase with band mass ratio and band energy separation. In the case of half-Heusler materials at high carrier densities, where carrier screening is significant, the scattering exponents vary at such a degree, that it is nearly impossible to link transport trends (i.e. mobility) to a specific scattering mechanism. From a statistical analysis for 11 different half-Heusler materials, we find that the scattering exponents follow a gaussian distribution with μ = −0.102 and σ = 0.175, which does not fit any of the standard exponents. In conclusion, our study will help improve the characterization of TE measurements, leading to better understanding and optimization. Alongside my presentation, I will be bringing a poster which will describe other work done using our code ElecTra.

Progress towards phase-field modeling of anisotropic crack propagation using physics-informed deep learning

In this study, we investigate the ability of variational physics-informed neural networks (VPINNs) to learn complex fracture processes in anisotropic media. VPINNs have recently been explored in the context of crack propagation in isotropic brittle solids, demonstrating key fracture mechanisms such as crack nucleation, propagation, kinking, branching, and coalescence through phase-field damage modelling, which represents the current state of the art [1]. However, their extension to crack propagation in anisotropic materials remains an open research challenge, which this work seeks to address. Unlike the second-order approximations used for isotropic cases, modelling anisotropic crack propagation in the phase-field framework requires fourth-order approximations of fracture energy [2]. This is accomplished by using NURBS elements to compute higher-order gradients. A neural network is then trained to minimise the system’s variational energy computed via finite element (FE) calculations. The proposed methodology is applied to several benchmark problems, such as a 2D square plate under pure tension and 2D L-shaped section with the right side forced upward. [1] Manav M., Molinaro R., Mishra S., De Lorenzis L., (2024). Phase-field modeling of fracture with physics-informed deep learning, Computer Methods in Applied Mechanics and Engineering, 429, 117104, DOI: https://doi.org/10.1016/j.cma.2024.117104 [2] Kakouris E.G, Triantafyllou S.P. (2018), Material point method for crack propagation in anisotropic media: a phase field approach, Archive of Applied Mechanics, 88 (1), 287-316, DOI: https://doi.org/10.1007/s00419-017-1272-7