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.

The use of ultrashort laser pulses to manipulate properties or investigate a materials response on femtosecond time-scales enables detailed tracking of charge, spin, and lattice degrees of freedom. When pushing the limits of experimental resolution, connection to theoretical modeling becomes increasingly important to infer causality relations. Weyl-semimetals are a particular class of materials of recent focus due to the topological protection of the Weyl-nodes, resulting in a number of fundamentally interesting phenomena. This work provides a first-principles framework based on time-dependent density-functional theory for tracking the distribution of Weyl-nodes in the Brillouin-zone following an excitation by a laser pulse. Investigating the prototype material TaAs, we show that residual shifts in the Weyl-Nodes' position and energy distribution are induced by a photo-excitation within femto-seconds through band-structure renormalization. Further, we provide an analysis of the relaxation pathway of the photoexcited band-structure through lattice vibrations.

Coherent Diffraction Imaging (CDI) is an experimental technique to image isolated structures by recording the scattered light. The sample density can be recovered from the scattered field through a Fourier Transform operation. However, the phase of the field is lost during the measurement and has to be algorithmically retrieved. Here we present SPRING, an analysis framework tailored to X-ray Free Electron Laser (XFEL) single-shot single-particle diffraction data that implements the Memetic Phase Retrieval method to mitigate the shortcomings of conventional algorithms. We benchmark the approach on data acquired in two experimental campaigns at SwissFEL and European XFEL. Results reveal unprecedented stability and resilience of the algorithm's behavior on the input parameters, and the capability of identifying the solution in conditions hardly treatable with conventional methods. A user-friendly implementation of SPRING is released as open-source software, aiming at being a reference tool for the CDI community at XFEL and synchrotron facilities.

The integration of machine learning (ML) models enhances the efficiency, affordability, and reliability of feature detection in microscopy, yet their development and applicability are hindered by the dependency on scarce and often flawed manually labeled datasets with a lack of domain awareness. We addressed these challenges by creating a physics-based synthetic image and data generator, resulting in an ML model that achieves comparable precision (0.86), recall (0.63), F1 scores (0.71), and engineering property predictions (  = 0.82) to a model trained on human-labeled data. We enhanced both models by using feature prediction confidence scores to derive an image-wide confidence metric, enabling simple thresholding to eliminate ambiguous and out-of-domain images, resulting in performance boosts of 5-30% with a filtering-out rate of 25%. Our study demonstrates that synthetic data can eliminate human reliance in ML and provides a means for domain awareness in cases where many feature detections per image are needed.

We introduce an automated, flexible framework (aiida-hubbard) to self-consistently calculate Hubbard and parameters from first-principles. By leveraging density-functional perturbation theory, the computation of the Hubbard parameters is efficiently parallelized using multiple concurrent and inexpensive primitive cell calculations. Furthermore, the intersite parameters are defined on-the-fly during the iterative procedure to account for atomic relaxations and diverse coordination environments. We devise a novel, code-agnostic data structure to store Hubbard related information together with the atomistic structure, to enhance the reproducibility of Hubbard-corrected calculations. We demonstrate the scalability and reliability of the framework by computing in high-throughput fashion the self-consistent onsite and intersite parameters for 115 Li-containing bulk solids with up to 32 atoms in the unit cell. Our analysis of the Hubbard parameters calculated reveals a significant correlation of the onsite values on the oxidation state and coordination environment of the atom on which the Hubbard manifold is centered, while intersite values exhibit a general decay with increasing interatomic distance. We find, e.g., that the numerical values of for the 3d orbitals of Fe and Mn can vary up to 3 eV and 6 eV, respectively; their distribution is characterized by typical shifts of about 0.5 eV and 1.0 eV upon change in oxidation state, or local coordination environment. For the intersite a narrower spread is found, with values ranging between 0.2 eV and 1.6 eV when considering transition metal and oxygen interactions. This framework paves the way for the exploration of redox materials chemistry and high-throughput screening of and compounds across diverse research areas, including the discovery and design of novel energy storage materials, as well as other technologically-relevant applications.

Maximally-localized Wannier functions (MLWFs) are widely employed as an essential tool for calculating the physical properties of materials due to their localized nature and computational efficiency. Projectability-disentangled Wannier functions (PDWFs) have recently emerged as a reliable and efficient approach for automatically constructing MLWFs that span both occupied and lowest unoccupied bands. Here, we extend the applicability of PDWFs to magnetic systems and/or those including spin-orbit coupling, and implement such extensions in automated workflows. Furthermore, we enhance the robustness and reliability of constructing PDWFs by defining an extended protocol that automatically expands the projectors manifold, when required, by introducing additional appropriate hydrogenic atomic orbitals. We benchmark our extended protocol on a set of 200 chemically diverse materials, as well as on the 40 systems with the largest band distance obtained with the standard PDWF approach, showing that on our test set the present approach delivers a success rate of over 98% in obtaining accurate Wannier-function interpolations, defined as an average band distance below 20 meV between the DFT and Wannier-interpolated bands, up to 2 eV above the Fermi level for metals or above the conduction band minimum for insulators (and a 100% success rate when including only bands up to 1 eV above these values).

The driving of vibrational motion by external electric fields is a topic of continued interest, due to the possibility of assessing new or metastable material phases with desirable properties. Here, we combine ab initio molecular dynamics within the electric-dipole approximation with machine-learning neural networks (NNs) to develop a general, efficient and accurate method to perform electric-field-driven nuclear dynamics for molecules, solids, and liquids. We train equivariant and autodifferentiable NNs for the interatomic potential and the dipole, modifying the model infrastructure to account for the multi-valued nature of the latter in periodic systems. We showcase the method by addressing property modifications induced by electric field interactions in a polar liquid and a polar solid from nanosecond-long molecular dynamics simulations with quantum-mechanical accuracy. For liquid water, we present a calculation of the dielectric function in the GHz to THz range and the electrofreezing transition, showing that nuclear quantum effects enhance this phenomenon. For the ferroelectric perovskite LiNbO, we simulate the ferroelectric to paraelectric phase transition and the non-equilibrium dynamics of driven phonon modes related to the polarization switching mechanisms, showing that a full polarization switch is not achieved in the simulations.

Machine learning and data-driven methods have started to transform the study of surfaces and interfaces. Here, we review how data-driven methods and machine learning approaches complement simulation workflows and contribute towards tackling grand challenges in computational surface science from 2D materials to interface engineering and electrocatalysis. Challenges remain, including the scarcity of large datasets and the need for more electronic structure methods for interfaces.

Understanding lattice dynamics and thermal transport mechanisms in cubic hybrid organic-inorganic perovskites remain challenging due to strong anharmonicity and phase transitions. Here, we investigate the thermal transport behavior in benchmark cubic hybrid perovskite FAPbI by coupling first principles-based anharmonic lattice dynamics with a linearized Wigner transport equation. Using the Temperature-Dependent Effective Potential (TDEP) method, we stabilize the negative soft modes, primarily dominated by organic FA cations. Our calculations predict an ultra-low thermal conductivity of ~ at 300 K, following a temperature dependence of . Contrary to common assumptions, we find that the [PbI] units, rather than FA cations, dominate thermal resistance. Furthermore, we demonstrate that anharmonic force constants are highly temperature-sensitive, relying on 0-K force constants significantly underestimates thermal conductivity. Our study not only elucidates the microscopic mechanisms governing thermal transport in FAPbI but also provides a robust framework for modeling heat conduction in hybrid organic-inorganic compounds.

Using first principles techniques, we show that infrared optical response allows us to discriminate between the topological and the trivial phases of 2D quantum spin Hall insulators (QSHI). We showcase germanene and jacutingaite, of recent experimental realization, as prototypical systems where the infrared spectrum is discontinuous across the transition, due to sudden and large discretized jumps of the Born effective charges (up to ~2). Our results, rationalized thanks to the low-energy Kane-Mele model, are robust with respect to dynamical effects, relevant when the electronic energy gap is comparable with the phonon frequency. In the small gap QSHI germanene, due to dynamical effects, the in-plane phonon resonance in the optical conductivity shows a Fano profile with remarkable differences in the intensity and the shape between different phases. Instead, the large-gap QSHI jacutingaite presents several IR-active phonon modes whose spectral intensities drastically change between different phases.

Infrared (IR) spectroscopy is a pivotal analytical tool as it provides real-time molecular insight into material structures and enables the observation of reaction intermediates in situ. However, interpreting IR spectra often requires high-fidelity simulations, such as density functional theory based ab-initio molecular dynamics, which are computationally expensive and therefore limited in the tractable system size and complexity. In this work, we present a novel active learning-based framework, implemented in the open-source software package PALIRS, for efficiently predicting the IR spectra of small catalytically relevant organic molecules. PALIRS leverages active learning to train a machine-learned interatomic potential, which is then used for machine learning-assisted molecular dynamics simulations to calculate IR spectra. PALIRS reproduces IR spectra computed with ab-initio molecular dynamics accurately at a fraction of the computational cost. PALIRS further agrees well with available experimental data not only for IR peak positions but also for their amplitudes. This advancement with PALIRS enables high-throughput prediction of IR spectra, facilitating the exploration of larger and more intricate catalytic systems and aiding the identification of novel reaction pathways.

Elemental Pr metal is unique among rare-earth elements in featuring a localized partially filled 4 shell without ordered magnetism. Experimental evidence attributes this absence of magnetism to a singlet crystal-field (CF) ground state of the Pr 4 configuration. Here, we construct an effective magnetic Hamiltonian for dhcp Pr, by combining density-functional theory with dynamical mean-field theory, in the quasiatomic Hubbard-I approximation. Our calculations fully determine the CF potential and predict singlet CF ground states for both inequivalent Pr sites. The intersite exchange interactions, obtained from the magnetic force theorem, are insufficient to close the CF gap to the magnetic doublets. Hence, ab-initio theory is demonstrated to explain the non-magnetic state of elemental Pr. We also find that the singlet ground state remains robust preventing conventional magnetic orders at the (0001) surface of Pr. Nevertheless, the gap between the ground state and the lowest excited singlet is significantly reduced at the surface, opening the possibility for exotic two-dimensional multipolar orders to emerge within this two-singlet manifold.

The design and understanding of oxide-ion and proton transport in solid electrolytes are pivotal to the development of fuel cells that can operate at reduced temperatures of <600 C. Atomistic modelling and machine learning are playing ever more crucial roles in achieving this objective. In this study, using passive and active learning techniques, we develop moment tensor potentials (MTPs) for two promising ionic conductors, namely, BaNbMoO and SrVO. Our MTPs accurately reproduce ab initio molecular dynamics data and demonstrate strong agreement with density functional theory calculations for forces, energies and stresses. They successfully predict diffusion coefficients and conductivities for both oxide ions and protons, showing excellent agreement with experimental data and ab initio molecular dynamics results. Additionally, the MTPs accurately estimate migration barriers, thereby underscoring their robustness and transferability. Our findings highlight the potential of MTPs in significantly reducing computational costs while maintaining high accuracy, making them invaluable for simulating complex ion transport mechanisms and supporting the development of next-generation solid oxide fuel cells.

Microstructural design is crucial yet challenging for thin-film semiconductors, creating barriers for new materials to achieve practical applications in photovoltaics and optoelectronics. We present the Daisy Visual Intelligence Framework (Daisy), which combines multiple AI models to learn from historical microscopic images and propose new synthesis conditions towards desirable microstructures. Daisy consists of an image interpreter to extract grain and defect statistics, and a reinforcement-learning-driven synthesis planner to optimize thin-film morphology. Using Ag-Bi-I perovskite-inspired materials as a case study, Daisy achieved over 120× and 87× acceleration in image analysis and synthesis planning, respectively, compared to manual methods. Processing parameters for AgBiI were optimized from over 1700 possible synthesis conditions within 3.5 min, yielding experimentally validated films with no visible pinholes and average grain sizes 14.5% larger than the historical mean. Our work advances computational frameworks for self-driving labs and shedding light on AI-accelerated microstructure development for emerging thin-film materials.

Training machine learning interatomic potentials that are both computationally and data-efficient is a key challenge for enabling their routine use in atomistic simulations. To this effect, we introduce franken, a scalable and lightweight transfer learning framework that extracts atomic descriptors from pre-trained graph neural networks and transfers them to new systems using random Fourier features - an efficient and scalable approximation of kernel methods. It also provides a closed-form fine-tuning strategy for general-purpose potentials such as MACE-MP0, enabling fast and accurate adaptation to new systems or levels of quantum mechanical theory with minimal hyperparameter tuning. On a benchmark dataset of 27 transition metals, franken outperforms optimized kernel-based methods in both training time and accuracy, reducing model training from tens of hours to minutes on a single GPU. We further demonstrate the framework's strong data-efficiency by training stable and accurate potentials for bulk water and the Pt(111)/water interface using just tens of training structures. Our open-source implementation (https://franken.readthedocs.io) offers a fast and practical solution for training potentials and deploying them for molecular dynamics simulations across diverse systems.

We have designed and implemented the Python package mag2exp, which enables researchers to perform a range of virtual experiments given a spatially resolved vector field for the magnetization, a typical result from computational methods to simulate magnetism such as micromagnetics. This software allows experimental measurements such as magnetometry, microscopy, and reciprocal space based techniques to be simulated in order to obtain observables that are comparable to those of the corresponding experimental measurement. Such virtual experiments tend to be more economic to carry out than actual experiments. There are many uses for virtual experiments, including (i) choosing the best experimental techniques and assessing their feasibility prior to experimentation, (ii) fine tuning experimental setup, (iii) guiding the experiment by conducting concurrent simulations of the measurement, and (iv) interpreting the experimental data at a later point though both qualitative and quantitative methods.

Scanning Electron Microscopes (SEMs) are widely used in experimental science laboratories, often requiring cumbersome and repetitive user analysis. Automating SEM image analysis processes is highly desirable to address this challenge. In particle sample analysis, Machine Learning (ML) has emerged as the most effective approach for particle segmentation. However, the time-intensive process of manually annotating thousands of SEM images limits the applicability of supervised learning approaches. Self-Supervised Learning (SSL) offers a promising alternative by enabling knowledge extraction from raw, unlabeled data. This study presents a framework for evaluating SSL techniques in SEM image analysis, focusing on novel methods leveraging the ConvNeXtV2 architecture for particle detection. A dataset comprising 25,000 SEM images is curated to benchmark these proposed SSL methods. The results demonstrate that ConvNeXtV2 models, with varying parameter counts, consistently outperform other techniques in particle detection across different length scales, achieving up to a 34% reduction in relative error compared to established SSL methods. Furthermore, an ablation study explores the relationship between dataset size and SSL performance, providing actionable insights for practitioners regarding model selection and resource efficiency. This research advances the integration of SSL into autonomous analysis pipelines and supports its application in accelerating materials science discovery.

We present NeuralMag, a flexible and high-performance open-source Python library for micromagnetic simulations. NeuralMag leverages modern machine learning frameworks, such as PyTorch and JAX, to perform efficient tensor operations on various parallel hardware, including CPUs, GPUs, and TPUs. The library implements a novel nodal finite-difference discretization scheme that provides improved accuracy over traditional finite-difference methods without increasing computational complexity. NeuralMag is particularly well-suited for solving inverse problems, especially those with time-dependent objectives, thanks to its automatic differentiation capabilities. Performance benchmarks show that NeuralMag is competitive with state-of-the-art simulation codes while offering enhanced flexibility through its Python interface and integration with high-level computational backends.

Graph Neural Network (GNN) potentials relying on chemical locality offer near-quantum mechanical accuracy at significantly reduced computational costs. Message-passing GNNs model interactions beyond their immediate neighborhood by propagating local information between neighboring particles while remaining effectively local. However, locality precludes modeling long-range effects critical to many real-world systems, such as charge transfer, electrostatic interactions, and dispersion effects. In this work, we propose the Charge Equilibration Layer for Long-range Interactions (CELLI) to address the challenge of efficiently modeling non-local interactions. This novel architecture generalizes the classical charge equilibration (Qeq) method to a model-agnostic building block for modern equivariant GNN potentials. Therefore, CELLI extends the capability of GNNs to model long-range interactions while providing high interpretability through explicitly modeled charges. On benchmark systems, CELLI achieves state-of-the-art results for strictly local models. CELLI generalizes to diverse datasets and large structures while providing high computational efficiency and robust predictions.

Many engineering alloys originally designed for conventional manufacturing lack considerations for additive manufacturing (AM), presenting opportunities for novel alloy designs. Evaluating alloy printability requires extensive analysis of chemical composition and processing conditions. The complexity of experimental exploration drives the need for high-throughput computational frameworks. This study introduces a framework that integrates material properties, processing parameters, and melt pool profiles from three thermal models to assess process-induced defects, such as lack-of-fusion, balling, and keyholing. A deep learning surrogate model accelerates the printability assessment by 1000 times without losing accuracy. We validate the framework with printability maps for the equiatomic CoCrFeMnNi system and apply it to explore printable alloys in the Co-Cr-Fe-Mn-Ni high-entropy alloy space. Ensemble probabilistic printability maps further provide insights into defect likelihood and uncertainty, enhancing alloy design for AM by efficiently navigating vast design spaces.

Rising atmospheric CO levels threaten climate stability, demanding transformative solutions in carbon capture, utilization, and storage. Porous activated carbons (ACs) derived from sustainable waste sources offer a promising route for cost-effective and eco-friendly carbon capture, thanks to their tunable surface chemistry and high surface areas. However, optimizing ACs for peak CO uptake is often hindered by complex, resource-intensive experimental workflows and the scarcity of high-quality data. This study presents a machine learning-driven framework that combines a multi-headed one-dimensional convolutional neural network (MH1DCNN) with multi-fidelity Bayesian optimization (MFBO) to efficiently navigate large design spaces by balancing exploration of uncertain regions with exploitation of known high-performing candidates. The MH1DCNN captures nonlinear relationships between physicochemical properties and CO uptake, serving as a deployable low-fidelity model. Using 841 literature-reported samples as high-cost, high-fidelity data and MH1DCNN-generated predictions as low-cost, low-fidelity evaluations, MFBO fuses these information sources through a probabilistic surrogate model, enabling rapid and cost-effective optimization. This approach reduces high-fidelity evaluation requirements by over 75% and identifies top-performing candidates using only 13 high-fidelity acquisitions. This scalable, data-driven strategy supports the development of closed-loop experiment-analysis-planning systems for future autonomous laboratories and accelerates sustainable materials discovery.