Saaed Amirjalayer
Heidelberg University, Germany
Vikram Deshpande
University of Cambridge, United Kingdom
Capturing high-rate spatiotemporal deformation of materials in three dimensions (3D) remains a significant challenge with current X-ray imaging techniques. We present a methodology that combines advances in neural rendering techniques with volume correlation methods to accurately reconstruct complex, high-rate 3D spatiotemporal structural evolutions. The fidelity and versatility of the method, which requires no pre-training, are demonstrated for a diverse set of intricate 3D-printed micro-architected solids. Using laboratory-based X-ray tomography, we capture the 3D growth of a dynamic crush band on a timescale of less than 100 milliseconds. By broadening this idea to a stereo X-ray concept, we eliminate the need to rotate the image object, thereby extending the technique to significantly faster timescales. Our neural rendering framework opens new possibilities for studying numerous poorly understood dynamic processes, such as the runaway failure of batteries and the temporal evolution of 3D shock microstructures under impact loading, all using laboratory X-ray systems.
Pascal Friederich
Karlsuhe Institute of Technology (KIT), Germany
Machine learning can accelerate the screening, design, and discovery of new molecules and materials in multiple ways, e.g. by virtually predicting properties of molecules and materials, by extracting hidden relations from large amounts of simulated or experimental data, or even by interfacing machine learning algorithms for autonomous decision-making directly with automated high-throughput experiments. In this talk, I will focus on our research activities on materials property prediction.
[1,2] and inverse design of materials [3], materials foundation models [4], as well as self-explaining graph neural networks [5,6] to foster scientific understanding in chemistry and materials science. [1] Ruff et al., Digital Discovery 3, 594-601, 2024.[2] Reiser et al., Communications Materials 3, 93, 2022. [3] Ruple et al., arXiv:2502.03146, under review. [4] Mirza et al., ICLR AI4MAT Workshop, 2025. [5] Teufel et al., Explainable Artificial Intelligence 2023 1902, 338-360 [6] Sturm et al., Angewandte Chemie 2025.
Stefanie Gräfe
Friedrich Schiller University Jena, Germany
The excitation of collective electron dynamics inside the metallic nanoparticles induced by external light fields leads to strongly re-shaped electromagnetic nearfields with a complex spatial and temporal profile. The interaction of these modified and enhanced nearfields with systems located in close vicinity to the metallic nanoparticle is the origin of many astonishing physical and chemical phenomena, such as the formation of new quasiparticles, new mechanisms for chemical reactions or the ultra-high spatial resolution and selectivity in molecular detection. For the theoretical description of such plasmonic hybrid systems in external light fields, it is necessary to describe both the electromagnetic interaction and the more chemical effects equally. In this talk, I will introduce our recent results on the theoretical description of these systems, with particular emphasis on spectroscopic applications, e.g., in the context of tip-enhanced Raman scattering spectroscopy and/or plasmon-induced catalysis [1-5]. Our calculations show pronounced changes of the Raman spectrum under non-resonant and resonant conditions and support the possibility of sub-nanometer spatial resolution.
Mike Hagan
Brandeis University, USA
Active matter is composed of particles that generate forces or motion, which leads to spectacular emergent dynamics. In principle, active matter could form the basis for a new class of materials with lifelike properties of biological organisms. Yet, active materials exhibit diverse dynamical states, most of which have chaotic dynamics that do not produce work or other useful functions. Thus, a robust control strategy is needed to drive active materials into an emergent state that corresponds to a desirable function. However, designing control protocols requires accurate dynamical models, which are not available for most active matter systems. Developing quantitative models using traditional statistical physics approaches is challenging because active materials lack the scale separation characteristic of equilibrium systems.
In this presentation, I will discuss efforts to combine machine learning, other data-driven techniques, and physics-based models with control theory to address this challenge in the context of a widely-used active material, microtubule-based active nematics. Recent advances in optogenetic motors have enabled constructing the light-activated active nematics, in which the activity can be spatiotemporally controlled by shining light on the sample. The challenge is to determine the spatiotemporal light sequence required to drive the system into a desired behavior.
I will describe two complementary approaches to computationally determine an optimal light sequence. In the first, we have adapted a method to discover optimal physics-based continuum models directly from spatiotemporal data, using sparse regression. We have identified several approaches to mitigate measurement errors in the data. We find that the method can reveal the relative contributions of different physical mechanisms, and quantitatively estimates key experimental parameters. Then, we have developed a framework to combine the optimal physics-based model with optimal control theory to solve for the spatiotemporal activity profile that drives the system into a desired state. We demonstrate that active materials can be driven into arbitrary behaviors, including those which do not correspond to dynamical attractors and thus cannot be accessed without control.
Since no model is perfectly accurate for a specific system, in the second approach we develop a deep reinforcement learning (DRL) based controller to enable model-free control of active materials. The controller discovers and implements spatiotemporal sequences of activity to drive a 2D active nematic system toward a prescribed dynamical steady-state. This framework does not require a detailed physics model, making it ideal for complex active materials that lack quantitative theoretical descriptions. Furthermore, the approach is extremely robust to noise and experimental measurement error. We compare the performance of the physics-based and DRL-based controllers for active nematics.
This work was supported by DE-SC0022291. Preliminary work was supported by NSF DMR-1855914 and DMR-2011846. Computing resources were provided by XSEDE TG-MCB090163 and the Brandeis HPCC (DMR-MRSEC 2011846 and OAC-1920147).
Antonio Calà Lesina
Leibniz University Hannover, Germany
Inverse design methods based on topology optimization can uncover nanophotonic structures in 3D with free-form shapes beyond human intuition, and optical functionalities not
obtainable with conventional design methods. This talk highlights the recent achievements of my team on large-scale topology optimization for metaphotonics and integrated optics. Some of the topics include the inverse design of nanostructures made of arbitrary dispersive optical materials, the broadband optimization of absorption in metallic and dielectric nanostructures, anapole effects in plasmonic meta-atoms for transparent metasurfaces and metamaterials, and nanoantennas with desired multipolar response for scattering engineering.
Karsten Reuter
Fritz Haber Institute of the Max Planck Society, Germany
Yair Shokef
Tel Aviv University, Israel
Martin Wegener
Karlsruhe Institute of Technology (KIT), Germany