arXiv subjects

Jonas Hänseroth

Publications and source records attributed to Jonas Hänseroth.

12 recordsLinked to original sources

arXiv2026

Reference-free certification of machine-learning interatomic potentials

Universal machine-learning interatomic potentials now reach held-out energy and force errors so small that they no longer predict how a model behaves in simulation. Here we show that a potential can be graded without any reference calculation, against properties the exact Born-Oppenheimer surface satisfies by mathematical or physical necessity. We organise these properties into four families, symmetry and invariance, self-consistency and integrability, statistical-mechanical equilibrium and regularity, and turn them into fourteen inexpensive probes, each certifying against a target that is exact and independent of chemistry and reference method. Every probe therefore returns an absolute, architecture-comparable score from the energies, forces and stresses a potential already exposes, and reveals failures a fixed test set cannot. Applied to 64 pretrained potentials, the suite resolves two orders of magnitude of certified quality at comparable reported accuracy, and flags a fine-tuned potential whose reference errors improve while its surface degrades until molecular dynamics fails. An interactive leaderboard of all 64 models is available at https://jhaens.github.io/pescert-bench.

cond-mat.mtrl-sci↗

arXiv2026

Enabling Domain-Specific Atomistic Models: A Machine Learning Potential for the Solid Acid Family

Machine-learned interatomic potentials trained across the periodic table have made atomistic simulation broadly accessible, and specializing them to a single compound class is widely expected to improve accuracy. Yet examples remain scarce, and fewer still surpass universal models in speed or reach a higher level of electronic-structure theory. Here we present a potential that is universal within the class of water-free solid-state hydrogen-bond network mediated proton conductors rather than across chemistry and a database of 4.4 million first-principles configurations spanning 55 materials. It surpasses leading general-purpose potentials across this domain, recovering measured activation energies and the ordering of anion rotational dynamics that those models miss; agreement on static structure does not imply agreement on transport. Accuracy falls for compositions far from the training set but stays competitive for close structural relatives, and a higher level of electronic-structure theory is reached with a few hundred additional configurations per material. We further introduce a compact variant carrying a fifth of the parameters, faster still and yet more accurate than every general-purpose model tested. It sustains more than a quarter of a million atoms on a single graphics processor, placing grain boundaries and the transition into the highly conducting phase within reach, and making a quantum treatment of the protons affordable. Reference accuracy and accessible system size thus become largely independent, offering a template for other compound classes.

cond-mat.mtrl-sci↗

arXiv2026

Data-Efficient Construction of Material-Specific Machine-Learning Interatomic Potentials from Ab Initio Molecular Dynamics Trajectories

Pretrained machine-learning interatomic potentials, so-called universal or foundation models offer an appealing starting point for atomistic simulations, but their accuracy for material-specific observables often remains limited without additional reference data (fine-tuning). Here, we systematically quantify how much first-principles data are required to convert universal models into ab initio-accurate material-specific potentials, and ask whether fine-tuning is necessarily preferable to training from scratch. We compare five universal MLIP frameworks, MACE-MP-0, SevenNet-0, GRACE-1L-OAM, MatterSim-v1-5M and ORB-v2, across seven chemically diverse systems incorporating rare and reactive events. Fine-tuning on only 10 AIMD-derived configurations is insufficient for the investigated systems; 200 configurations succeed in favorable cases, but the outcome remains strongly system-dependent. By contrast, 2000 AIMD configurations constitute a robust default, yielding low force and energy errors and reproducing the target material-specific observables. Moderately dense sub-sampling of the AIMD trajectory reduces the required trajectory length tenfold with little loss in model quality. Training from scratch on the same datasets is competitive with, and often slightly more accurate than, naive fine-tuning for MACE and SevenNet, whereas GRACE requires more data. The energy profile for a sulfur-vacancy jump in MoS$_2$ reveals that low trajectory-level errors do not guarantee a correct reaction profile, highlighting the need for observable-level validation. Finally, we show that averaging independently trained models improves predictions in scarce-data regimes at no additional first-principles cost. Together, these results provide practical guidelines for converting limited AIMD reference data into reliable material-specific MLIPs for nanosecond-timescale simulations at near-DFT accuracy.

cond-mat.mtrl-sci↗

arXiv2026

Quantum nuclear and band-dispersion effects recover near-UV absorption in short-hydrogen-bonded organic crystals

Near-UV optical absorption is increasingly reported in hydrogen-bonded organic and biomolecular materials lacking aromatic or extended pi-conjugated chromophores, yet its microscopic origin remains unresolved and electronic-structure calculations often overestimate experimental absorption onsets. Here, we combine machine-learned interatomic potentials for large-scale classical and quantum nuclear sampling with periodic excited-state calculations to address this discrepancy in L-pyroglutamine ammonium, an experimentally established glutamine-derived crystal containing a well-resolved short hydrogen bond and exhibiting non-aromatic near-UV optical response. Using controlled in silico ion substitutions that vary the surrounding hydrogen-bond environment while preserving this scaffold, we compute optical spectra from configurations sampled along classical and quantum nuclear trajectories using hybrid-functional time-dependent density functional theory. We show that nuclear quantum effects stabilise proton-sharing configurations that are strongly suppressed classically, redshifting the lowest bright excitations by 0.5-0.8 eV and raising the fraction of configurations with bright excitations below 6 eV from approximately 3% to approximately 30%. Explicit Brillouin-zone sampling provides a further, mechanistically distinct redshift of 0.5-1.1 eV, reflecting modest but significant indirect electronic character. Only when both effects are incorporated does the calculated onset recover the experimental 3.8-4.5 eV range. These results establish quantum proton fluctuations and reciprocal-space convergence as cooperative but physically distinct ingredients required for predictive optical spectroscopy of strongly hydrogen-bonded molecular materials.

physics.chem-ph↗

arXiv2026

Universal Interatomic Potentials as Configuration-Space Generators for One-Shot and Iterative Fine-Tuning of Ab Initio-Accurate Material-Specific Models

Universal machine-learning interatomic potentials (MLIPs) are rapidly becoming general-purpose tools for atomistic simulation, but their role in quantitative materials modeling when reactive events are involved remains unsettled. We compare five universal MLIPs across seven chemically diverse systems and find that strong performance on standard benchmarks does not guarantee accurate predictions of target observables. In particular, zero-shot models do not reliably reproduce reactive, transport, or high-barrier processes, exemplified here in particular by the sulfur-vacancy jump in MoS$_2$. We therefore propose a practical alternative: universal MLIPs are used to generate long molecular dynamics trajectories, the resulting configurations are sub-sampled and relabeled with DFT, and material-specific MLIPs are subsequently trained or fine-tuned on the resulting first-principles datasets. This workflow converts universal models into efficient configuration-space generators while retaining ab initio reference labels for training. Across the tested systems, $2{,}000$ DFT-recalculated structures are often sufficient to obtain accurate fine-tuned or trained-from-scratch models. For the most challenging case, iterative self-training progressively refines the sampled configuration space and recovers the DFT MoS$_2$ potential energy profile with only $600$ first-principles calculations in total. The resulting workflow enables the generation of $1$ ns ab initio-quality trajectories - including training data generation and model creation - within three days.

cond-mat.mtrl-sci↗

arXiv2026

Revealing Hydroxide Ion Transport Mechanisms in Commercial Anion-Exchange Membranes at Nano-Scale from Machine-learned Interatomic Potential Simulations

Hydroxide ion transport in anion-exchange membranes fundamentally limits the efficiency of alkaline water electrolysis for green hydrogen production, yet the atomic-scale transport mechanisms remain poorly understood due to the computational challenges associated with modeling ion dynamics. Given that anion-exchange membranes enable alkaline electrolysis with abundant catalysts while avoiding perfluoroalkyl and polyfluoroalkyl materials, a deeper mechanistic understanding of hydroxide transport in these systems is essential for advancing sustainable hydrogen production. Here, we show that large-scale molecular dynamics simulations with fine-tuned machine-learned interatomic potentials provide atomistic insight into hydroxide mobility in a commercial membrane over tens of nanoseconds and over ten nanometer. We find that increasing water content transforms isolated water clusters into a connected hydrogen-bond network that enables long-range proton transfer. Under dry conditions hydroxide ions are trapped near positively charged groups and transport is strongly hindered, whereas well-hydrated membranes exhibit extended proton migration and diffusion coefficients approaching those of dilute aqueous solutions. The simulations reproduce experimental trends in diffusion and activation energies. Our results establish a direct link between nano-scale structure and macroscopic transport. Beyond mechanistic insight, the presented simulation framework enables predictive, simulation-guided optimization of membrane chemistry and architecture, opening a pathway toward the rational design of more efficient anion-exchange membranes for green hydrogen technologies.

cond-mat.mtrl-sci↗

arXiv2026

High-throughput screening and mechanistic insights into solid acid proton conductors

Proton-conducting solid acids could enable water-free operation of high-temperature fuel cells. However, systematic materials screening has, hitherto, been computationally prohibitive. Here, we introduce a two-stage high-throughput screening strategy that directly computes proton diffusion coefficients, enabled by machine-learned interatomic potentials fine-tuned to ab initio data. Starting from more than six million materials, our screening -- based on structural motifs rather than empirical descriptors -- identifies $27$ high-performing proton conductors, including over ten previously unexplored compounds. These include sustainable and commercially available materials, candidates that have not yet been synthesized, organic systems that fall outside conventional design rules, and known proton conductors that validate our approach. Importantly, our findings reveal a universal oxygen--oxygen distance of approximately $2.5$~Å at the moment of proton transfer across diverse chemistries, providing mechanistic insight and showing that macroscopic proton conductivity emerges from the interplay between anion rotational dynamics, hydrogen-bond network connectivity, and proton-transfer probability.

physics.chem-ph↗

arXiv2025

Bridging Atomistic and Mesoscale Lithium Transport via Machine-Learned Force Fields and Markov State Models

Lithium diffusion in solid-state battery anodes occurs through thermally activated hops between metastable sites often separated by large energy barriers, making such events rare on ab initio molecular dynamics (AIMD) timescales. Here, we present a bottom-up multiscale workflow that integrates AIMD, machine-learned force fields (MLFFs), and Markov state models (MSMs) to establish a quantitatively consistent link between atomistic hopping mechanisms and mesoscale transport. MLFFs fine-tuned on AIMD reference data retain near-DFT accuracy while enabling large-scale molecular dynamics simulations extending to tens of nanoseconds. These extended trajectories remove the strong finite-size bias present in AIMD and yield diffusion coefficients in excellent agreement with experiment. Furthermore, from these long MLFF trajectories, we obtain statistically converged lithium jump networks and construct MSMs that remain Markovian across more than two orders of magnitude in the lag times used for their construction. The resulting MSMs faithfully reproduce mean-square displacements and recover rare diffusion processes that do not occur on AIMD timescales. In addition to propagating lithium distributions, the MSM transition matrices provide mechanistic insight: their eigenvalues and eigenvectors encode characteristic relaxation timescales and dominant transport pathways. Although demonstrated for defect-free crystalline Li$_x$Si$_y$ phases, the AIMD$\rightarrow$MLFF$\rightarrow$MSM framework is general and provides a transferable approach for describing lithium transport in amorphous materials, defect-mediated diffusion, and next-generation solid-state anodes.

physics.chem-ph↗

arXiv2025

Fine-Tuning Unifies Foundational Machine-learned Interatomic Potential Architectures at ab initio Accuracy

This work demonstrates that fine-tuning transforms foundational machine-learned interatomic potentials (MLIPs) to achieve consistent, near-ab initio accuracy across diverse architectures. Benchmarking five leading MLIP frameworks (MACE, GRACE, SevenNet, MatterSim, and ORB) across seven chemically diverse compounds reveals that fine-tuning universally enhances force predictions by factors of 5-15 and improves energy accuracy by 2-4 orders of magnitude. The investigated models span both equivariant and invariant, as well as conservative and non-conservative, architectures. While general-purpose foundation models are robust, they exhibit architecture-dependent deviations from ab initio reference data; fine-tuning eliminates these discrepancies, enabling quantitatively accurate predictions of atomistic and structural properties. Using datasets constructed from equidistantly sampled frames of short ab initio molecular dynamics trajectories, fine-tuning reduces force errors by an order of magnitude and harmonizes performance across all architectures. These findings establish fine-tuning as a universal route to achieving system-specific predictive accuracy while preserving the computational efficiency of MLIPs. To promote widespread adoption, we introduce the aMACEing Toolkit, which provides a unified and reproducible interface for fine-tuning workflows across multiple MLIP frameworks.

physics.chem-ph↗

arXiv2025

Optimizing Machine Learning Potentials for Hydroxide Transport: Surprising Efficiency of Single-Concentration Training

We investigate the transferability of machine learning interatomic potentials across concentration variations in chemically similar systems, using aqueous potassium hydroxide solutions as a case study. Despite containing identical chemical species (K$^+$, OH$^-$, H$_2$O) across all concentrations, models fine-tuned on specific KOH concentrations exhibit surprisingly poor transferability to others, with force prediction errors increasing dramatically from ~30 meV/A (at training concentration) to ~90 meV/A (at very different concentrations). This reveals a critical limitation when applying such models beyond their training domain, even within chemically homogeneous systems. We demonstrate that strategic selection of training data can substantially overcome these limitations without requiring extensive computational resources. Models fine-tuned on intermediate concentrations (6.26 mol/L) exhibit remarkable transferability across the entire concentration spectrum (0.56-17.98 mol/L), often outperforming more computationally expensive models trained on multiple concentration datasets. This approach enables accurate simulation of hydroxide transport dynamics across varying electrolyte conditions while maintaining near-quantum accuracy. Our simulations further reveal the emergence of hydroxide-hydroxide hydrogen bonding at high concentrations - a phenomenon not explicitly represented in dilute training data but successfully captured by our intermediate-concentration model. This work establishes practical guidelines for developing broadly applicable machine learning force fields with optimal transferability, challenging the assumption that diverse training datasets are always necessary for robust performance in similar chemical environments.

physics.chem-ph↗

arXiv2025

Hydroxide Mobility in Aqueous Systems: Ab Initio Accuracy with Millisecond Timescales

We present a multiscale simulation approach for hydroxide transport in aqueous solutions of potassium hydroxide, combining ab initio molecular dynamics (AIMD) simulations with force field ensemble averaging and lattice Monte Carlo techniques. This method achieves near ab initio accuracy by capturing the femtosecond scale dielectric relaxation dynamics of the aqueous hydrogen bonding network, while extending the simulation capability to millisecond diffusion timescales. This extraordinary extension of the available length and time scales enables future studies of hydroxide mobility in functional materials such as nanostructured anion-exchange membranes, where hydroxide ions migrate through nanometer-sized channels. Remarkably, our approach demonstrates that a single AIMD trajectory is sufficient to predict hydroxide conductivity over a range of concentrations, underscoring its computational efficiency and relevance to the design of advanced energy materials.

physics.comp-ph↗

arXiv2025

Modelling complex proton transport phenomena -- Exploring the limits of fine-tuning and transferability of foundational machine-learned force fields

The solid acids CsH$_2$PO$_4$ and Cs$_7$(H$_4$PO$_4$)(H$_2$PO$_4$)$_8$ pose significant challenges for the simulation of proton transport phenomena. In this work, we use the recently developed machine-learned force field (MLFF) MACE to model the proton dynamics on nanosecond time scales for these systems and compare its performance with long-term ab initio molecular dynamics (AIMD) simulations. The MACE-MP-0 foundation model shows remarkable performance for all observables derived from molecular dynamics (MD) simulations, but minor quantitative discrepancies remain compared to the AIMD reference data. However, we show that minimal fine-tuning -- fitting to as little as 1 ps of AIMD data -- leads to full quantitative agreement between the radial distribution functions of MACE force field and AIMD simulations. In addition, we show that traditional long-term AIMD simulations fail to capture the correct qualitative trends in diffusion coefficients and activation energies for these solid acids due to the limited accessible time scale. In contrast, accurate and convergent diffusion coefficients can be reliably obtained through multi-nanosecond long MD simulations using machine-learned force fields. The obtained qualitative and quantitative behavior of the converged diffusion coefficients and activation energies now matches the experimental trends for both solid acids, in contrast to previous AIMD simulations that yielded a qualitatively wrong picture.

physics.comp-ph↗