arXiv subjects

Pranav Kumar

Publications and source records attributed to Pranav Kumar.

At least 19 recordsLinked to original sources

arXiv2026

Contextual Tamil Spelling and Grammar Correction Using Progressively Fine-Tuned Sequence-to-Sequence Transformers

Tamil spell and grammar correction is challenging because Tamil is an agglutinative low-resource language with rich verbal morphology, complex sandhi (phonetic transformation) rules at word boundaries, and a script of 247 distinct letters. Prior work targets word-level surface errors with rule-based methods, statistical n-gram models, Minimum Edit Distance, or hybrid pipelines with a transformer re-ranker; such methods cannot reliably handle contextual errors - subject-verb agreement, tense consistency, or cross-word sandhi - which require sentence-level understanding. We propose an end-to-end sequence-to-sequence formulation and fine-tune mT5-small and mBART-50 on a synthetic corpus of up to 657,720 noisy-clean Tamil sentence pairs spanning ten error categories. Both backbones follow the same four-stage progressive schedule, each stage targeting one weakness: surface noise (v2), contextual grammar (v3), single-site sandhi (v4), and multi-site cross-word sandhi (v5). On a 1,000-sentence balanced diagnostic set verified disjoint from all training data, our best model, mBART-50 v5, reaches 69.3% top-1 exact-match accuracy, with 87.5% on sandhi and 43.5% on subject-verb agreement. The schedule is what produces these gains: subject-verb accuracy rises from 1.0% to 52.5% once contextual pairs are introduced, and sandhi from 0% to 87.5% once multi-site sandhi pairs are. We additionally quantify a precision-recall trade-off this literature has not reported: sandhi recall is paid for monotonically in identity accuracy. Finally, Tamil-LLaMA-7B-Instruct reaches 19.0% zero-shot and 24.7% with three demonstrations against a 20.0% copy baseline, showing that a Tamil-adapted instruction model does not transfer to specialised sentence-level correction without task-specific supervision.

cs.CL↗

arXiv2026

Grain boundary evolution in nanoparticles

Grain boundary evolution is a key dynamical process that enables structural rearrangements in nanoparticles and drives them towards low energy configurations. Grain boundaries can also enhance catalytic properties, making it important to understand the elementary processes underlying their evolution for grain boundary engineering in nanoparticles. Compared with bulk materials, nanoparticles have additional rotational and translational degrees of freedom, can accommodate structural changes through shape relaxation, and grain boundaries terminate at free surfaces. However, the atomic-scale mechanisms of GB evolution under these less constrained conditions remain comparatively less understood than in bulk materials. Here, using atomistic simulations, we focus on low energy $\Sigma3$ (coherent twin boundary) and $\Sigma11$ grain boundaries, which are among the persistent compact GB structures that emerge during nanoparticle structural evolution. We identify two fundamental atomic displacements, column shift (C) and screw shift (S), that recur during the evolution of these grain boundaries and their junctions. These displacements occur through different atomic pathways and combine in different ways to generate grain boundary migration, structural transformations, and junction evolution. In particular, the same initial and final grain boundary configurations can be connected through different atomic pathways, and a column shift can occur either as a full shift or through disconnection kinks. C and S remain identifiable even when the grain boundary character changes, for example during a $\Sigma11$ to $\Sigma3$ transformation. An understanding of these elementary GB processes can help identify strategies to control grain boundary evolution and thereby engineer GB structures in nanoparticles.

cond-mat.mtrl-sci↗

arXiv2026

An Ontology for Machine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs) approximate quantum-mechanical energies and forces---conventionally computed by density functional theory (DFT) or wave-function methods---at a fraction of the cost. The field encompasses a growing ecosystem of algorithms, training datasets, hyperparameters, and target materials, yet the metadata needed to systematically compare, reproduce, and build upon MLIP studies remains scattered across papers, scripts, and ad-hoc file formats. We present the MLIPs ontology, an OWL 2 DL ontology that captures the concepts needed to describe MLIP methods, their hyperparameters, training datasets with DFT provenance, and published benchmarks. The ontology is organized into three modules---Method, Training Data, and Benchmark---and connects existing ontologies in materials science (MDO, CMSO/ASMO) and machine learning (ML-Schema), complementing dataset-side schemas such as Croissant. It declares 27 formal axioms enforcing data completeness and consistency, including property chains that link trained models to their methods and training data. We demonstrate the ontology through a running example based on Moment Tensor Potentials and evaluate it through competency-question execution on a 20-paper seeded knowledge graph, OWL reasoning, and comparison with existing ontologies.

cs.AI↗

arXiv2026

Machine learning via artificial neural networks coupled with density functional theory and experiments for thermodynamic optimization of high-entropy alloys for hydrogen storage at room temperature

High-entropy alloys (HEAs) have received considerable attention for hydrogen storage because of their compositional flexibility; however, designing HEAs with optimal thermodynamics is critical. This study employs machine learning via artificial neural networks (ANN) and density functional theory (DFT) to design a novel AB-type TixNb2-xVCrMnFe (x = 0.5-2.0) high-entropy system for hydrogen storage at ambient temperature (A: Ti, V and Nb, and B: Cr, Mn and Fe). Both ANN and DFT predict that the hydride formation enthalpy decreases to negative values with increasing the titanium content. Two alloys with x > 1.5 are predicted to achieve enthalpies within the -25 to -39 kJ/mol range, making them appropriate for room-temperature hydrogen storage. Experiments demonstrate good agreement with the enthalpy predictions, with the Ti-rich alloys showing reversible hydrogen storage with fast kinetics at room temperature. These results provide a framework for reliable use of data analysis and ab initio calculations to explore high-entropy hydrides as hydrogen storage materials.

cond-mat.mtrl-sci↗

arXiv2026

Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies

Vision-Language-Action (VLA) policies translate language and visual inputs into robot actions, where their hidden representations directly shape closed-loop behavior. However, mechanistic interpretability tools from language and vision-language models do not transfer cleanly to VLAs: outputs are robot actions rather than human-readable tokens, and interventions can only be tested via expensive closed-loop rollouts. We propose an event-grounded interpretability pipeline that anchors SAE feature analysis to behavioral events rather than text contexts. End-effector keyframes are clustered within each task using visual, state, and temporal cues, linking SAE features to behaviorally salient events and, via optional VLM annotations, to semantic context. To our knowledge, our pipeline is among the first to ground SAE-based VLA analysis in closed-loop behavioral events. Across two simulation architectures and a real-robot study, event-grounded ranking yields the strongest causal effects on OpenVLA and transfers to the continuous action chunks of $\pi_{0.5}$. SAE is a sparse but imperfect intervention basis: usability varies with architecture and intervention site, and aggressive intervention reveals safety and interpretability limits. Overall, event-grounded SAE analysis emerges as a practical starting point for behavior-anchored VLA interpretability, motivating future work on SAE features beyond action-aligned coordinates, finer-grained closed-loop evaluation, and safe interventions for high-stakes VLA deployments. Code is available at \url{https://github.com/xc-j/Event-SAE}.

cs.RO↗

arXiv2026

Adding Compilation Metadata To Binaries To Make Disassembly Decidable

The binary executable format is the standard method for distributing and executing software. Yet, it is also as opaque a representation of software as can be. If the binary format were augmented with metadata that provides security-relevant information, such as which data is intended by the compiler to be executable instructions, or how memory regions are expected to be bounded, that would dramatically improve the safety and maintainability of software. In this paper, we propose a binary format that is a middle ground between a stripped black-box binary and open source. We provide a tool that generates metadata capturing the compiler's intent and inserts it into the binary. This metadata enables lifting to a correct and recompilable higher-level representation and makes analysis and instrumentation more reliable. Our evaluation shows that adding metadata does not affect runtime behavior or performance. Compared to DWARF, our metadata is roughly 17% of its size. We validate correctness by compiling a comprehensive set of real-world C and C++ binaries and demonstrating that they can be lifted, instrumented, and recompiled without altering their behavior.

cs.CR↗

arXiv2026

Classical Unattainability of Extremality in non-BPS D-brane Systems

For the worldline theory of an extremal black hole, extremality amounts to vanishing ground state energy. In light of recent gravity results one would expect much like the ground state degeneracy this fine tuned condition too will not be met. It is unclear though whether this should be a quantum artefact or classical. In this paper we consider a non-BPS extremal four-charge Reissner Nordstrom black hole in N = 8 String theory. It is shown that the microscopic D-brane description fails to admit any extremal state, even classically. This positive energy is expected to destabilize the near horizon AdS2. The positive minimum energy is a direct consequence of the pattern of supersymmetry breaking by the D-branes. The black hole entropy is found to be the logarithm of the number of isolated minima and hence is related to the configurational entropy of the microscopic potential. We also find multiple continua of local minima corresponding to marginally bound states of the constituent D-branes.

hep-th↗

arXiv2025

Hydrogen diffusion in TiCr$_2$H$_x$ Laves phases: A combined ab initio and machine-learning-potential study

The kinetics of hydrogen diffusion in C15 cubic and C14 hexagonal TiCr$_2$H$_x$ (0 < $x$ <= 4) Laves-phase hydrogen storage alloys is investigated with density functional theory (DFT) and machine learning interatomic potentials (MLIPs). Generalized solid-state nudged elastic band calculations are conducted based on DFT for all symmetrically inequivalent paths between the first-nearest-neighbor face-sharing interstitial sites. The hydrogen migration barriers are substantially higher for the paths that require breaking a Ti-H bond than for those that require breaking a Cr-H bond. Molecular dynamics (MD) simulations with the MLIPs also demonstrate that hydrogen migration occurs more frequently within the hexagonal rings made of the A$_2$B$_2$ interstitial paths, each requiring the breaking of Cr-H bonds, than along the inter-ring paths. The diffusion coefficients of hydrogen obtained from the MD simulations reveal a non-monotonic dependence on hydrogen concentration, which is more pronounced at lower temperatures. Time-averaged radial distribution functions of hydrogen further show that hydrogen avoids face-sharing positions during diffusion and that the hydrogen occupancy at the second-nearest-neighbor edge-sharing positions increases with increasing hydrogen concentration. The diffusion coefficients of hydrogen within 400-1000 K follow an Arrhenius relationship, with activation barriers consistent with most experimental values. One-order of magnitude overestimation of diffusion coefficients compared with some experiments suggests a substantial impact of hydrogen trapping by defects such as Cr vacancies and Ti anti-sites in non-stoichiometric TiCr$_2$ in experiments.

cond-mat.mtrl-sci↗

arXiv2025

Bounds and asymptotic expansions for the radii of convexity and uniform convexity of normalized Bessel functions

This paper explores the asymptotic behaviour of the radii of convexity and uniform convexity for normalized Bessel functions with respect to large order. We provide detailed asymptotic expansions for these radii and establish recurrence relations for the associated coefficients. Additionally, we derive generalized bounds for the radii of convexity and uniform convexity by applying the Euler-Rayleigh inequality and potential polynomials. The asymptotic inversion method and Rayleigh sums are the main tools used in the proofs.

math.CV↗

arXiv2025

The radius of starlikeness of regular Coulomb wave functions

Motivated by the pioneering work of M.S. Robertson [Ro54] and R.K. Brown [Br60], [Br62], who examined the geometric properties of some normalised solutions of second-order homogeneous differential equations, in this paper we investigate the radii of univalence and starlikeness for two kind of normalised regular Coulomb wave functions. Moreover, a generalized normalised Bessel function is introduced, and its radius of starlikeness is studied by using two different approaches. In addition, the asymptotic behaviour, with respect to the large order, of the radius of starlikeness of one type of normalised Coulomb wave functions is considered, which is in fact the first zero of the derivative of the regular Coulomb wave function. We derive a complete asymptotic expansion for this radius of starlikeness and provide a recurrence relation for the coefficients of this expansion. The proof is based on Rayleigh sums of the zeros of Coulomb wave functions, asymptotic inversion and some basic results on regular Coulomb wave functions developed by \v{S}tampach and \v{S}\'{t}ov\'{\i}\v{c}ek [SS14].

math.CA↗

arXiv2025

Asymptotic behavior of zeros of Bessel function derivatives

We derive two distinct asymptotic expansions for the zeros $j_{\nu,k}^{(n)}$ of the $n$-th derivative of Bessel function $J_\nu^{(n)}(x)$. The first is a McMahon-type expansion for the case when $k \to \infty$ with fixed $\nu$, for which we also establish an explicit error bound. The second addresses the case when $\nu \to \infty$ with fixed $k$ and it involves the zeros of Airy functions and their derivatives. These results extend and refine the classical work of Wong, Lang, and Olver on the zeros of Bessel functions. In the course of obtaining our main results, we also generalize several auxiliary results, which in turn provide a broader framework for the study of zeros of special functions.

math.CA↗

arXiv2025

Machine-learning interatomic potentials achieving CCSD(T) accuracy for systems with extended covalent networks and van der Waals interactions

Machine-learning interatomic potentials (MLIPs) enable large-scale atomistic simulations at moderate computational cost while retaining ab initio accuracy. MLIPs trained on coupled-cluster data, particularly CCSD(T), have emerged as a promising route to achieve chemical accuracy beyond the limits of density functional theory (DFT) and to incorporate non-empirical van der Waals (vdW) interactions. Most existing approaches are, however, still not straightforwardly applicable for systems with extended covalent networks such as covalent organic frameworks (COFs) due to the limited availability of CCSD(T) for periodic systems. Here we present a methodology to train MLIPs with CCSD(T) accuracy for these systems. The approach uses the {\Delta}-learning method with a dispersion-corrected tight-binding baseline. This strategy enables training on compact molecular fragments while preserving transferability toward the periodic systems. Dispersion interactions are accounted for by adding vdW-bound multimers in the training set, and the combination with a vdW-aware tight-binding baseline allows the formally local MLIP to attain CCSD(T)-level accuracy even for systems dominated by long-range vdW forces. The resulting potential yields root-mean-square energy errors below 0.4 meV/atom on training and test sets and reproduces electronic total atomization energies, bond lengths, harmonic vibrational frequencies, and inter-molecular interaction energies for benchmark molecular systems. We apply the method to a prototypical quasi-two-dimensional COF composed of carbon and hydrogen. The COF structure, inter-layer binding energies, and hydrogen absorption are analyzed at CCSD(T) accuracy. The developed methodology opens a practical route to large-scale atomistic simulations for systems with extended covalent networks and vdW interactions with chemical accuracy.

cond-mat.mtrl-sci↗

arXiv2025

Direct and inverse problem for bi-wave equation with time-dependent coefficients from partial data

In this article, we study a direct and an inverse problem for the bi-wave operator $(\Box^2)$ along with second and lower order time-dependent perturbations. In the direct problem, we prove that the operator is well-posed, given initial and boundary data in suitable function spaces. In the inverse problem, we prove uniqueness of the lower order time-dependent perturbations from the partial input-output operator. The restriction in the measurements are considered by restricting some of the Neumann data over a portion of the lateral boundary.

math.AP↗

arXiv2025

Machine Learning Potentials for Hydrogen Absorption in TiCr$_2$ Laves Phases

The energetics of hydrogen absorption in C15 cubic and C14 hexagonal TiCr$_2$H$_x$ Laves phases is investigated for $0 < x \le 6$ with density functional theory (DFT) and machine learning interatomic potentials (MLIPs). The MLIPs are trained with configurations generated through a series of active-learning schemes. Basin-hopping Monte Carlo (BHMC) simulations based on the MLIPs predict minimum-energy hydrogen configurations, along with enthalpies of formation and hydrogen orderings. The obtained phase transformations at 0 K agree well with the experiments at low temperatures. The hydrogen solubility limits in the low-concentration $\alpha$ phases at 0 K are predicted to be $x = 1.0$ and $x = 1.5$ for the C15 and the C14 phases, respectively. At these concentrations, C15 TiCr$_2$H shows the $Cc$ monoclinic symmetry, while C14 TiCr$_2$H$_{1.5}$ shows the $Ama2$ orthorhombic symmetry, both of which have not been reported for this system. The first and the second hydride phases, i.e., $\beta$ and $\beta'$, at 0 K are found around $x = 3$ and $x = 4$, respectively, for both the C15 and the C14 phases. In the second-hydride $\beta'$ phases, C15 TiCr$_2$H$_4$ shows the $I4_1/a$ tetragonal symmetry, while C14 TiCr$_2$H$_4$ shows the $R\bar3c$ rhombohedral symmetry. Hydrogen repulsion are found to extend to edge-sharing interstices, affecting the hydrogen ordering. Furthermore, the $6h_2$ A$_2$B$_2$ interstices are found to be energetically substantially more preferable for C14 TiCr$_2$H$_x$ than the other A$_2$B$_2$ interstices at low hydrogen concentrations, influencing the hydrogen-occupation trend.

cond-mat.mtrl-sci↗

arXiv2025

Density results of biharmonic functions on symmetric tensor fields and their applications to inverse problems

In this article we discuss density of products of biharmonic functions vanishing on an arbitrarily small part of the boundary. We prove that one can use three or more such biharmonic functions to construct a dense subset of smooth symmetric tensor fields up to order three, in a bounded domain. Furthermore, as an application of the density results, in dimension two or higher, we solve a partial data inverse problem for a biharmonic operator with nonlinear anisotropic third and lower order perturbations. For the inverse problem, we take the Dirichlet data to be supported in an arbitrarily small open set of the boundary and measure the Neumann data on the same set. Note that the analogous problem for linear perturbations are still unknown. So far, partial data problems recovering nonlinear perturbations were studied only up to vector fields. The full data analogues of the inverse problem has recently been studied for three or higher dimensions.

math.AP↗

arXiv2024

Inverse boundary value problem for the Convection-Diffusion equation with local data

We study a local data inverse problem for the time-dependent Convection-Diffusion Equation (CDE) in a bounded domain where a part of the boundary is treated to be inaccessible. Up on assuming the inaccessible part to be flat, we seek for the unique determination of the time-dependent convection and the density terms from the knowledge of the boundary data measured outside the inaccessible part. In the process, we show that there is a natural gauge in the perturbations, and we prove that this is the only obstruction in the uniqueness result.

math.AP↗

arXiv2024

Diffusion Meets DAgger: Supercharging Eye-in-hand Imitation Learning

A common failure mode for policies trained with imitation is compounding execution errors at test time. When the learned policy encounters states that are not present in the expert demonstrations, the policy fails, leading to degenerate behavior. The Dataset Aggregation, or DAgger approach to this problem simply collects more data to cover these failure states. However, in practice, this is often prohibitively expensive. In this work, we propose Diffusion Meets DAgger (DMD), a method to reap the benefits of DAgger without the cost for eye-in-hand imitation learning problems. Instead of collecting new samples to cover out-of-distribution states, DMD uses recent advances in diffusion models to synthesize these samples. This leads to robust performance from few demonstrations. We compare DMD against behavior cloning baseline across four tasks: pushing, stacking, pouring, and shirt hanging. In pushing, DMD achieves 80% success rate with as few as 8 expert demonstrations, where naive behavior cloning reaches only 20%. In stacking, DMD succeeds on average 92% of the time across 5 cups, versus 40% for BC. When pouring coffee beans, DMD transfers to another cup successfully 80% of the time. Finally, DMD attains 90% success rate for hanging shirt on a clothing rack.

cs.RO↗

arXiv2023

Non-trivial saddles in microscopic description of black holes

Non-trivial gravitational saddles have played a key role in the island proposal for the black hole information paradox. It is worth asking if non-trivial saddles exist in microscopic descriptions of black holes. We show this to be the case for 1/8 BPS black holes in N = 8 String Theory in a duality frame, where all charges are Ramond Ramond. The saddles are in the Coulomb branch, where they describe marginally stable bound states of the constituent branes, and correspond to vacua of the BFSS model. The non-perturbative suppression scale is determined by the binding energy.

hep-th↗