2026-10-14 | | Total: 46
Continuum gyrokinetic codes use structured, field-aligned grids on nested flux surfaces; particle-in-cell codes support unstructured meshes and open field lines but carry a statistical noise floor. We present a continuum drift-kinetic solver that occupies neither position: a semi-Lagrangian discontinuous Galerkin discretization on unstructured triangular poloidal planes extruded toroidally, whose parallel advance is a map along the exact field line in the flux-coordinate independent manner, so the mesh is never aligned with the field, with adiabatic electrons and a long-wavelength polarization field solve. Every operator is verified against an independently computed reference, from Landau damping to the geodesic-acoustic root and a linear ion-temperature-gradient eigenvalue problem, every measured rate within 4% except the geodesic-acoustic damping, within 12%. A linear toroidal ion-temperature-gradient demonstration on a canonical-Maxwellian equilibrium recovers a radially coherent mode whose growth rate, 0.065 c_s/a, is stationary in time. The collisionless zonal-flow residual is the one measurement still moving: it rises with poloidal refinement from 0.72 to 0.87 of the model's own orbit-quadrature prediction for the measured observable and has not converged. The solver is fully device resident, and every result is reproduced between CPU and GPU builds and across MPI process counts. The code, the data behind every table and an artifact description are in a public repository.
General-purpose molecular dynamics software supports diverse interatomic potentials and simulation requirements. Within a simulation stage, the loaded potential model and the required calculations typically remain fixed as the atomic configuration evolves. We present the Wisevolve Potential Engine (WPE), which uses this information from GPUMD to select and reuse a GPU execution policy tailored to the model and simulation stage. GPUMD retains simulation control and ownership of the simulation data. On a single NVIDIA RTX 5090, WPE-enabled GPUMD (GPUMD-WPE) achieves a median throughput of 109 million atom-step/s across 106 neuroevolution potential (NEP) workloads. For a 15.1-million-atom diamond Si system, it sustains 142 million atom-step/s. Relative to GPUMD, geometric-mean speedups are 3.49x across these NEP workloads, 3.61x across 16 workloads using NEP with dynamic charges (qNEP), 1.24x across 28 classical-potential workloads, and 10.5x across 56 path-integral molecular dynamics (PIMD) configurations. End-to-end calculations of water structure, lattice thermal conductivity, and PIMD thermal expansion show close agreement with corresponding GPUMD results and published simulation results. A separate 12.5-million-atom Cu shock-compression simulation demonstrates its applicability to large-scale nonequilibrium dynamics. These results support model- and stage-specific specialization as a practical way to accelerate GPU potential evaluation in general-purpose molecular dynamics workflows.
The use of artificial intelligence (AI) in physics research is increasingly common, yet few guidelines exist for its effective use. Starting with a principles-based approach, we outline a set of ten requirements for collaborative AI agents in experimental physics, spanning scientific norms, effective agent use, and domain expertise. We take a human-centered approach in which agents act as collaborators that amplify rather than replace the expertise of a human researcher (user) who directs the work and is responsible for its results. We then present a framework designed to enforce these requirements, detailing methods for improving off-the-shelf behavior of large language models (LLMs) with custom scaffolding and describing general-purpose scientific workflows and skills. In particular, we highlight the power of closed-loop campaigns where agents work iteratively toward a well-defined goal. We demonstrate the potential of this mode of research through six case studies in experimental physics, covering device design, hardware control, neural network optimization, physical model identification, and data analysis. All six case studies result in useful work products that add utility to our research, including a neural network architecture and solid-state quantum defect photodynamic model that substantially outperform our published work; a microwave waveguide design that extends a published design concept to a fabrication-ready device meeting our experimental constraints; a quantum control pulse sequence candidate that improves simulated performance by more than an order of magnitude; commissioning of an NMR spectrometer to $\sim$1~Hz frequency precision on live hardware; and finally a data-fitting pipeline that achieves, to our knowledge, the first reconstruction of the Ising microstate of an artificial spin-ice array from diffraction-limited magnetic imaging.
Projection-based model-order reduction (MOR) offers a promising approach for accelerating expensive full-order model (FOM) by constructing reduced-order models (ROMs) that retain the majority of FOM accuracy while significantly reducing the computational cost. However, conventional static ROMs, often following offline/online training strategy, can exhibit limited predictive capability for convection-dominated problems featuring strongly evolving physics such as shocks. Adaptive ROMs emerge as a promising alternative and address such limitation by dynamically updating the low-dimensional subspace during online simulation to track locally evolving system dynamics, thereby enabling accurate predictions even for shock-dominated flows. However, such online adaptation introduces additional computational overhead, and frequent basis and sampling-point updates can substantially reduce their computational advantage relative to their static counterpart. To reconcile this tradeoff, the current work introduces a self-adaptive MOR framework that dynamically constructs and deploys the most appropriate ROM formulation (either adaptive or static) on the fly according to the evolving characteristics of the solution. to achieve optimal computational efficiency while maintaining high modeling accuracy of the local physics during the online model execution. In this regard, the proposed framework seeks to exploit the advantages of both adaptive- and static-ROM approaches to achieve an optimal computational efficiency and predictive accuracy while reducing reliance on extensive offline training. The framework is developed and evaluated for modeling rotating detonation engines (RDEs), which provide a challenging test environment involving strongly convection-dominated, transient, and multiscale reacting-flow dynamics.
We present a mass-weighted meshless formulation for the shifted incompressible lattice Boltzmann method (LBM) whose streaming step is exactly unitary, combining with the Carleman embedding. In traditional meshless LBM, on a scattered node set the semi-Lagrangian streaming step is an interpolation, which is not unitary; a generic block encoding of this matrix therefore requires a normalization proportional to its sparsity and logarithmically many ancillas. To overcome the non-unitarity, we instead express the interpolation as the exponential exp(a Dx + b Dy) of mimetic derivative operators. Endowing each node with its Voronoi mass and applying a mass-weighted similarity transform renders these derivative operators skew-Hermitian by enforcing moment constraints together with low-wave number trigonometric differentiation conditions, so that the interpolation becomes exactly unitary. Such unitarization of interpolation leave the local collision step unchanged up to a rescaling of its quadratic term, where the same weights define the amplitude encoding of the distribution functions. For quasi-uniform node sets, the mass-weighting diagonal remains within a unit-order normalization that degrades only mildly as the Carleman truncation order grows, thereby mitigating the cost of block encoding. Numerical experiments on the Taylor Green vortex confirm that the method reproduces the full meshless LBM, with the velocity error decreasing as the Carleman order increases.
The sudden growth in AI data centers is raising numerous environmental concerns. Low-frequency noise pollution is a concern for local residents, that may not just be an annoyance but also carry significant health impacts. An FDTD model of sound propagation over the topography near an AI data center is presented. The topography plays a large role in the level of noise pollution, with areas of sudden elevation on either side of the river near the AI data center more susceptible to high noise levels, while areas further afield largely protected from noise pollution by the topography. Understanding the role that topography plays in noise pollution, and capturing the noise pollution in specific topographies, can help identify areas where noise mitigation strategies might be best employed.
acousticHelmholtzFoam is an OpenFOAM application that computes time-harmonic acoustic pressure fields in heterogeneous gas-liquid media on unstructured polyhedral meshes. The complex pressure is stored as real and imaginary volume fields, the material coefficients are constructed from a volume-of-fluid phase fraction with geometric face-fraction averaging, and open boundaries are truncated by Cartesian perfectly matched layers (PMLs) whose damping tensor is generated by the companion utility setPMLFields. The two coupled real-valued equations are discretized with the cell-centred finite-volume operators of OpenFOAM and assembled explicitly, including processor couplings and boundary contributions, into a distributed PETSc matrix that is solved by default with the MUMPS sparse direct solver. The solver writes pressure amplitude, complex velocity, acoustic radiation pressure and time-averaged momentum flux for acoustic-force evaluation. This paper documents the implemented equations, the case interface, the assembly and solution algorithm, the build procedure and the test infrastructure. A layered gas-liquid-PML case with an analytical solution serves as an end-to-end regression: on 8000 cells the whole-domain relative complex-pressure error is $1.197\times10^{-3}$, and serial and MPI solutions on up to eight ranks agree to within $10^{-12}$ in relative $L_2$ norm. A single-axis acoustic-levitator height sweep with constant-volume spheroidal drops demonstrates the two-phase material input and the force output, and strong- and problem-size-scaling measurements characterize the direct-solver workflow. The numerical formulation and its verification are presented in a companion paper (arXiv:2610.01593).
Extending the Fokker--Planck equation to a coupled system of equations through non-Abelian symmetry groups offers a powerful theoretical framework for coupled multi-state stochastic systems. However, translating this abstract gauge field formulation into viable numerical predictions remains an unaddressed challenge. Motivated by our previous work on ice crystal habit evolution, in this work we bridge this gap by presenting the first numerical solver for an SU(2) gauge-coupled Fokker--Planck system in (1+1) dimensions. Using an operator-splitting scheme combined with a non-oscillatory advection solver, our algorithm strictly ensures probability conservation, non-negativity, and numerical stability. We demonstrate that non-Abelian gauge coupling acts as a cross-reactive feedback loop that fundamentally alters system dynamics, driving continuous probability exchange between states, suppressing localised density peaks, and forcing long-time spatial delocalisation. Validated across linear and quadratic coupling regimes up to strong interactions (g <= 1.0), this work establishes the foundational computational infrastructure needed to integrate complex multi-state physics. Overall, while our approach provides a key step toward improving ice cloud representation in atmospheric and climate models, the underlying physical framework and computational engine remain fundamentally general and multidisciplinary, offering a versatile paradigm for any domain requiring coupled multi-state stochastic equations.
SIREN (Ship Inviscid Resistance and Elevation eNgine) is a Rankine-source panel solver for the steady wave resistance of a ship, in the tradition of Dawson's method. The free-surface condition is linearised about a base flow, and the linearisation is iterated. In the present formulation the base flow is the solution of a Neumann problem on the actual boundary: sources on the entire hull, wet and dry, and on the free surface panels, with impermeability imposed on both, so that the free surface acts as a rigid lid at the position it currently occupies. This replaces the classical double-body flow, whose image about z=0 is meaningful only for a flat surface and a hull cut at z=0, and it removes the need for desingularised free-surface panels. In the non-linear iteration the free-surface panels are moved rigidly in the vertical to the relaxed wave elevation at their centroid, the hull panels are marked wet or dry against the moved surface, and the base flow is recomputed on the moved geometry; impermeability applies to the whole hull regardless of this marking, which is used only to select which panels enter the force integral, and under-relaxation is confined to the passes where the non-linear correction accumulates without being reset. The radiation condition is imposed by a one-row upstream shift of the collocation points, and the wave resistance is obtained by pressure integration and, independently, by transverse wave-cut analysis.
Cu-C materials and interfaces are central to diamond thermal management, metallization, graphene growth and carbon-coating technologies, but atomistic simulations across these applications require a potential that describes metallic Cu, diverse carbon bonding, heterogeneous Cu-C interactions and close atomic encounters. Here we develop NEP-CuC, a transferable neuroevolution potential trained on a dataset of 2,727 uniformly relabelled configurations selected from a heterogeneous Cu-C candidate pool. Disagreement within an ensemble of five independently trained models provides an internal indicator of model sensitivity. A Ziegler-Biersack-Littmark (ZBL) correction describes short-range nuclear repulsion, while a runtime D3(BJ) correction supplements dispersion-sensitive calculations. Tests on source-derived held-out and independent external datasets, together with structural, elastic, defect, interfacial, vibrational and finite-temperature transport benchmarks, establish broad but property-dependent physical fidelity across elemental Cu, diverse carbon environments and heterogeneous Cu-C configurations. Large-scale simulations of amorphous-carbon formation and orientation-dependent shock compression of Cu further test transferability under strongly non-equilibrium conditions. For Cu/diamond interfaces, the same potential describes deposition, annealing and tensile response: preparation history leaves persistent differences in interphase structure, while the mean peak nominal tensile stress at 500 K decreases from $12.35\pm0.71$ to $6.65\pm0.98$ GPa as the incident Cu energy during the initial deposition stage increases from 1 to 50 eV. NEP-CuC provides an atomistic framework for linking processing, interfacial structure and physical response across Cu-C systems, with direct relevance to Cu/diamond thermal-management interfaces, diamond metallization and non-equilibrium Cu-C processing.
Magnetic racetrack memory requires localized carriers that combine rapid motion, resistance to perturbations and a detectable signal. Here we show how a 360° domain wall in a hollow ferromagnetic tube combines these properties within a single continuous magnetic shell. Two successive magnetization reversals form a bound object whose surrounding domains point in the same direction. Analytical energetics and micromagnetic simulations establish its finite separation, recovery from tested perturbations and current driven transport without microscopic Dzyaloshinskii-Moriya interaction or a maintaining field. The texture's winding resists unwinding while radial rotation carries an energy cost, providing a specific control over stability. Relative rotation of the constituent spins accounts quantitatively for additional displacement during current onset. The wall supports motion at hundreds of metres per second, with transient velocities approaching one kilometre per second at stronger drive. A prepared seven-bit pattern retains its carriers and local readout contrast after pulsed transport and recovery. Curved-track and differential-current tests identify radial stiffness and drive uniformity as controls for intact motion. These results establish a physical basis for fast, locally distinguishable tubular magnetic bits and identify a continuous switching pathway that can be targeted by tailored excitation.
Kohn-Sham density-functional theory is expensive because it needs orbitals that extend over the whole system. For insulators this can be avoided by using only the diagonal of the one-electron resolvent, $g(\mathbf{r},z)=\langle\mathbf{r}|(z-\hat h)^{-1}|\mathbf{r}\rangle$, at a few complex energies $z$. We call the scheme local-resolvent density-functional theory (LRes-DFT). The density and the band energy are linear in $g$, so neither orbitals nor a kinetic-energy functional are needed. The complex energies are $M$ equally spaced points on a circle that encloses the occupied levels, offset by half a spacing so that no point lies on the real axis. Summing the resolvent over them is equivalent to occupying each level with $1/(1+u^M)$, where $u$ is the level's energy measured from the circle's center in units of its radius. This is a smooth step that is near 1 inside the circle and near 0 outside, and it stays between 0 and 1 for every $M$. If the circle moves with the chemical potential, forces at fixed electron number are exact. Gauss-Legendre and Zolotarev rules need fewer points when the gap is small. In three dimensions we use eigenpairs of small windows around each molecule instead of resolvents. This is divide-and-conquer (DC) with a different occupation function, and it is as accurate as DC with Fermi-Dirac occupations. For H$_2$ molecular crystals of up to 1024 electrons it agrees with Kohn-Sham to $2.5\times10^{-7}$ Ha per electron. On one core it is slower than dense diagonalization below about 500 electrons and 3.5 times faster at 1024 electrons. It loses accuracy for disordered or compressed systems and for diffuse basis functions.
Although machine learning potentials (MLPs) generalize across configurational and chemical compound spaces, their accuracy is commonly believed to be limited by their training methods, with improvements beyond this limit achieved so far only by refitting to experimental observables. In this Perspective, we argue that MLPs can also be made more accurate than their training methods non-empirically by generalizing across a third space, that of quantum-chemical methods. We propose routes to method-transferable MLPs that predict at unseen methods, with working implementations in both density-functional and wavefunction-based theories showing that they can be more accurate than any of their training methods. Aiming beyond the accuracy of the training methods could also make universal MLPs more accurate than single-method models, even when the method-transferable MLPs remain less accurate than those QM methods. Method-transferable MLPs will shape the construction of training datasets and the evaluation of models, pushing the accuracy limits of chemical simulations.
We introduce STRUCTURES-P, a machine-learned all-electron orbital-free density functional for periodic CHNO systems. Electrostatics are evaluated analytically in an atom-centered Gaussian basis, while an equivariant network learns the kinetic and exchange-correlation energy from energies and gradients of non-ground-state densities. Minimizing the functional from a superposition of atomic densities recovers the Kohn-Sham ground state: on held-out crystals, the density error falls from 6.7% to 0.4% of the electron count and the energy error reaches 0.25 mHa per atom, half that of a same-architecture interatomic potential. A single functional can be trained for both isolated molecules and periodic systems, and the model generalizes to larger unit cells. It largely preserves relative-energy rankings and reproduces convex-hull vertices on the tested H-N composition slice. Together, these results demonstrate the viability of machine-learned all-electron orbital-free DFT for variational density optimization and energy evaluation in fixed-geometry solids.
Immersed-boundary methods avoid body-fitted meshes, but the energy or entropy stability of the underlying flow discretisation does not automatically extend to the immersed treatment. This manuscript analyses the volume-penalised compressible Navier--Stokes equations for both energy and entropy stability, on the continuous and semi-discrete levels, and integrates the isolated source exactly in time. The source is formulated as a one-parameter family indexed by $θ$, whose momentum component relaxes the fluid velocity towards the prescribed solid velocity; $θ$ sets how much of the resulting kinetic-energy loss leaves the domain rather than being converted to internal energy. The two stability requirements impose opposing bounds on $θ$, with dissipation of the relative total energy requiring $θ\geq0$ and stability of the mathematical entropy requiring $θ\leq1$. Where penalisation acts, both conditions are therefore satisfied if and only if $0\leqθ\leq1$, and the symmetric value $θ=1/2$ divides the kinetic-energy loss equally between the domain and the internal energy. We conclude that any entropy-stable split-form discontinuous Galerkin spectral element method remains semi-discretely entropy stable once the source is added at the collocation nodes, for stationary and prescribed moving bodies alike.
We present a Python-based program that combines maximally localized Wannier functions-based first-principles electronic structure with the nonequilibrium Green's function formalism to enable efficient simulations of scanning tunneling microscopy (STM). The workflow builds upon standard density functional theory packages to construct compact transport Hamiltonians that preserve the properties of realistic materials. Superconductivity is incorporated via an effective pairing potential within the Nambu-spinor formalism, while coupling to the STM tip and substrate electrodes is treated using the wide-band limit approximation. Within this formulation, transport quantities such as current-voltage characteristics and differential conductance are evaluated in a closed analytical form. This avoids explicit numerical energy integrations and enables rapid spectroscopic calculations over broad parameter spaces. The framework provides a unified and computationally efficient approach for modeling superconducting hybrid nanostructures with spin-orbit coupling, magnetic textures, and spin polarization. Representative applications in spin chains, two-dimensional and molecular magnets on a superconductor demonstrate its robust capabilities for both simple model systems and realistic ab initio molecular structures relevant to modern STM experiments.
Microscopic simulations resolve spatial patterns but are feasible only over small domains and short timescales, whereas bulk descriptions erase those patterns through spatial averaging. Mesoscopic fields preserve this spatial organisation, but their effective equations are known only for selected microscopic systems and scaling limits. We introduce Deep Mesoscopic Dynamics (DeepMesoD), which uses neural networks to learn an effective mesoscopic field equation from microscopic trajectories at a prescribed coarse-graining scale. For the nonconserved dissipative dynamics studied here, DeepMesoD learns the free-energy functional with a spatial interaction kernel and a state-dependent dissipation operator constrained by the fluctuation-dissipation relation. We first validate the construction in the Ising-Kac model, where the learned mesoscopic equation recovers the known hydrodynamic limit and rollouts remain stable far beyond the training horizon. We then treat two systems without closed-form mesoscopic descriptions. For a two-dimensional Heisenberg ferromagnet governed by atomistic spin dynamics, DeepMesoD learns a mesoscopic model whose predictions match atomistic references up to 1 ns, and a rollout with an unseen initial condition remains bounded up to 25 ns. For layered CrBr$_3$, the learned kernel captures the suppression and sign reversal of interlayer coupling at a stacking fault whose location and strength are not supplied as inputs. Controlled changes to the learned interaction show how fault number and placement shape multistep magnetic responses, complementing our magneto-photoluminescence measurements. Across the two atomistic systems, DeepMesoD achieves a three-to-four-order-of-magnitude reduction in wall-clock time compared with atomistic spin dynamics.
Noncollinear ferrimagnets combine a net magnetic moment with internal spin degrees of freedom, creating opportunities to couple magnetic, mechanical, thermal, and electronic responses. Using first-principles calculations, symmetry analysis, atomistic spin simulations, and transport modeling, we identify stable NbMnX$_2$ ($X$ = S, Se, Te) monolayers with noncollinear ferrimagnetic ground states and Curie temperatures up to 313 K. Competing exchange, Dzyaloshinskii--Moriya interactions, and magnetic anisotropy produce field-responsive canted states that couple strongly to the lattice and charge carriers. Magnetization rotation changes the elastic modulus of NbMnTe$_2$ by up to 36\%, while lattice entropy substantially enhances the near-room-temperature magnetocaloric response of NbMnS$_2$ and NbMnSe$_2$. In NbMnSe$_2$, rotation restores a mirror-protected crossing, continuously closes an approximately 0.13 eV gap, and produces a conductivity-based anisotropic magnetoresistance ratio up to 10,000\% at 100 K. The combined results establish noncollinear NbMnX$_2$ monolayers as a multifunctional platform for two-dimensional spintronics, nanomechanics, and thermal control.
Density-functional tight binding has yet to yield a transferable description of vanadium oxides; for the standard second-order construction and the tested cells, we demonstrate that two of its approximations deviate by electron volts and offset each other in the band-index centroid separation between the V 3d and O 2p bands: the neutral two-centre reference omits electron-volt environmental terms, whereas the Mulliken charge model assigns the compact V 3d shell less than half of the self-consistent potential shift observed in PBE. Instead, we build a tight-binding model where the reference Hamiltonian derives from superposed neutral atoms, the full PBE exchange-correlation potential and frozen-core orthogonality, without adjustable terms, and where the orbital-resolved charge response employs three screening factors obtained from PBE linear response and an unscreened Coulomb range. Against the primary PAW reference, the reference energy at fixed geometry and the stated numerical settings matches the plane-wave Harris-Foulkes energy differences of VO2, exhibiting signed deviations of +2.6 meV/atom for M1 relative to rutile (an energy difference of 92 meV/atom) and +6.2 and -5.2 meV/atom under 3% volume compression and expansion; meanwhile, the charge response captures 21 of 22 fitted and all 8 held-out matched PBE responses within tolerance, providing the V 3d shell with nearly the same self-consistent shift as PBE (-3.89 versus -3.92 eV in rutile). The convexity of the spinless model at fixed geometry prevents it from spontaneously differentiating symmetry-equivalent sites under the symmetries respected by its implemented operators and sampling, so within this model such differentiation in the Magnéli phases must arise from electron-lattice coupling or an explicit spin or correlation term.
The clean-energy transition is driving unprecedented demand for metals, whose extraction is among the largest industrial sources of carbon emissions. Molten-salt electrolysis offers a path toward carbon-free production of aluminum, iron, titanium, and critical metals. However, corrosive operating conditions and complex, multicomponent feedstocks make experimental development costly and challenging. Here, we introduce a first-principles framework for constructing electrochemical phase diagrams of multicomponent systems to accelerate electrode and electrolyte design and guide the selection of compatible materials and operating conditions for electrochemical cells. We derive a grand potential that generates these diagrams at scale, and supply free energies using first-principles data obtained through a fine-tuned foundation machine-learned interatomic potential. For silicon in molten CaCl$_2$, the construction not only recovers the experimental diagram when using the first-principles framework on a limited set of structures, but also predicts further domains when extended to a larger set of crystal structures. Finally, we demonstrate the predictive power of the approach by screening ten alkali and alkaline-earth halide electrolytes for aluminum electrowinning. Three thermodynamic criteria identify NaF, where cryolite emerges as the stable aluminum-bearing phase, recovering the chemistry that Hall and Héroult established empirically in 1886. The framework provides a first-principles basis for selecting compatible electrodes, electrolytes, and operating conditions toward low-carbon metal production from complex source materials.
A key challenge in the design of next-generation colliders is predicting the effects of beam-beam interactions and using these predictions to optimize machine performance. Beam-beam interactions are characterized by strong focusing forces, radiative effects, and the generation of background particles. While analytical formulas provide insight into the dynamics of beam-beam interactions, these nonlinear, coupled effects are most accurately described with numerical methods. In this paper, we introduce the open-source particle-in-cell code WarpX as a modern tool for modeling the interaction point in future lepton colliders. WarpX provides high-performance execution on distributed GPU architectures within an open-source and actively maintained code base. We present convergence scans and benchmarks against the established beam-beam codes GUINEA-PIG (v1.2.2) and CAIN (v2.42) for luminosity spectra, photon emission, and secondary-particle production under a wide range of beam parameters corresponding to different collider concepts. Our results show that WarpX (v26.09) reproduces the benchmark beam-beam observables while enabling GPU-portable, extensible simulations. These results establish WarpX as a scalable platform for high-fidelity beam-beam modeling in future collider studies.
Machine learning surrogates of partial differential equations (PDEs) are commonly trained on data generated by numerical solvers. Therefore, the training data obey not the PDE but the modified equation of the numerical scheme, and the test loss, computed on data from the same solver, cannot distinguish a model that learned the physics from one that learned the discretization error. Here, we show that the difference is nevertheless visible in the trained model. We propose a suite of structure-preserving fidelity diagnostics, simple scores that measure whether a learned model respects one known property of the physics, such as a conserved quantity, a symmetry or a Hamiltonian structure, without access to the training data or the governing equation. Over five systems, the diagnostics track the error with respect to the true dynamics with a rank correlation of up to $0.94$, while the test loss is uncorrelated or anti-correlated with it. For the Burgers, Korteweg-de Vries and Navier-Stokes equations, the diagnostics separate numerical schemes that the test loss ranks as equivalent, including schemes of the same order. order.
We present a method for enforcing exactly Dirichlet, Neumann, and Robin conditions on general triangular domains with arbitrarily curved boundaries. The algorithmic constructions developed here complements those in a previous work for general quadrilateral domains. Our method combines (i) the exact mapping between a general curved triangular domain and the standard triangle, (ii) TFC constrained expressions, and (iii) Barnhill transfinite interpolations, to provide explicit closed-form trial ansatze that mathematically satisfy the imposed boundary conditions exactly. Corner compatibility induced by the boundary conditions is pivotal, and enforcing exactly the imposed boundary conditions hinges on the exact enforcement of the induced compatibility constraints. The corner compatibility associated with the acute vertices of the standard triangle and the coupling between gradient components when the hypotenuse corresponds to a Neumann/Robin boundary pose particular challenges. Two strategies are critical to resolving these issues, by reformulating conditional corner constraints into equivalent unconditional forms and by introducing a separate TFC-style auxiliary edge construction together with an associated linear solve on the hypotenuse. We analyze the corner compatibility constraints and present formulations for the general form of functions that exactly satisfy the imposed Dirichlet, Neumann, or Robin conditions on curved triangular domains. The proposed method has been implemented using extreme learning machine type randomized neural networks for solving PDE problems. We present ample numerical experiments with linear/nonlinear PDE problems on a variety of domains with complex boundary geometries. Numerical results show that the current method has enforced the Dirichlet, Neumann, and Robin conditions on curved domain boundaries to the machine precision.
Small polarons, formed when excess electrons become localized through lattice distortions, play a critical role in charge transport, catalysis, and degradation processes in transition-metal (TM) oxides. Whether universal machine-learning interatomic potentials (uMLIPs), which lack explicit electronic degrees of freedom, can describe them remains unclear. Here, we investigate H-induced electron localization across a range of TM oxides using a geometric descriptor based on TM-O bond-length expansion (Z-score) to identify polaronic distortions. We find that uMLIPs exhibit varying degrees of electron localization, from delocalized states to metastable small polarons, with their behavior strongly dependent on the training data and treatment of electronic correlations. In particular, the absence of Hubbard U corrections for specific TM cations in the training data leads to electron delocalization, in contrast to DFT+U predictions. Although some uMLIPs reproduce polaron hopping barriers during coupled proton-polaron transfer, all of them fail to stabilize spatially separated polaron-donor configurations, instead favoring localization near the H dopant.
Fast, high-fidelity molecular generation requires accurate chemical choices and efficient sampling. Chemical building blocks reduce graph size, but make each node a choice among tens of thousands of related identities. We introduce MIRRA (Molecular Identity Representation, Refresh and Assembly), a discrete flow-matching generator that learns from chemical knowledge and evaluates identity probabilities only when needed, while keeping every identity in its vocabulary available. Chemical features share information across related blocks, and learned identity corrections distinguish them in molecular context. By deciding which positions will refresh before computing their probabilities, the sampler avoids unused predictions while preserving each sampling transition. Against the strongest baselines in each task, MIRRA reduces chemical and graph distribution errors on ZINC250k by 29-31% and 53-60%, respectively, with generation up to 8.6 times faster. Property targeting yields an estimated 7.8-20.8 times more distinct novel hits per second. On a held-out polymer test set, MIRRA achieves 100% periodic validity with known chain topology, versus 57.0% for a polymer variational autoencoder. These results demonstrate fast, high-fidelity generation with the full chemical vocabulary available throughout sampling.