arXiv subjects

Himanshu Mishra

Publications and source records attributed to Himanshu Mishra.

At least 19 recordsLinked to original sources

arXiv2026

F-WANDA: Fisher-Reweighted Post-Training Pruning for Sustainable Deployment of Large Language Models

One-shot post-training pruning is the most energy-frugal compression strategy for largelanguage models (LLMs), yet existing approaches trade either quality (WANDA) or compute cost (SPARSEGPT). We introduce F-WANDA, a drop-in modification of WANDA that reallocates the per-row keep budget across output neurons in proportion to the empirical Fisher information of the pre-activation. The Fisher signal is collected in a single additional backward pass over the same calibration corpus WANDA already uses; no weights are updated. On LLAMA-2-7B at 50 % unstructured sparsity, F-WANDA attains WikiText-2 perplexity of 6.85, matches WANDA fluency, and improves 5-shot MMLU by +1.6 pp over WANDA and +1.1 pp over SPARSEGPT, while incurring only one-third of SPARSEGPT pruning wall-clock and energy. The headline trade-off is achieved without extra calibration data or fine-tuning, placing F-WANDA on the Pareto frontier of quality versus pruning cost for sustainable LLM compression.

cs.AI↗

arXiv2026

Resolving the Bubble Puzzle: Hydrogen Peroxide Formation Precedes Hydroxyl Radicals in Microbubbles and is Governed by Solid-Water Interfaces

An alternative explanation is presented for recent reports that attribute sustained chemiluminescence (CL) and electrochemiluminescence (ECL) from electrogenerated microbubbles on steel or copper electrodes in aqueous luminol solutions (over 2-30 V range) to the spontaneous formation of hydroxyl radicals at the gas-water interface. Our experiments with a broad set of electrodes, viz., steel, copper, aluminium, and platinum, reveal that while microbubbles can be electrogenerated on all electrodes, CL is exhibited by steel and Cu only and not by Al and Pt. These observations establish that the gas-water interface of microbubbles is not the site for hydroxyl radical generation (else CL would be recorded in all cases). Complementary quantification of H2O2 in these experiments reveals its electrode dependence as follows: Al > Cu > Steel > Pt. This establishes that depending on the electrode, H2O2 forms first, and in some cases, hydroxyl radicals are observed (i.e., where CL/ECL is seen). Experiments with NMR and EPR spectroscopy revealed that: (i) H2O2 formation occurs only when O2 is present in water; and (ii) while steel and copper generate hydroxyl radicals through 1-electron reduction of H2O2, Al does not promote one-electron reduction of H2O2 to generate hydroxyl radicals, and Pt preferentially promotes disproportionation of H2O2 to H2O and O2. In fact, we demonstrate that H2O2 and hydroxyl radicals can be observed at specific metal-water interfaces even without microbubbles, confirming that the solid surface is the reactive site. Therefore, this work affords electrode-based predictions of whether or not electrogenerated microbubbles would yield CL in luminol solutions and calls into question the notion of spontaneous formation of hydroxyl radicals at gas-water interfaces.

physics.chem-ph↗

arXiv2026

Orientation dynamics of a settling spheroid in simple shear flow: bifurcations and stochastic alignment

We investigate the orientation dynamics of a settling spheroid in simple shear flow, combining a deterministic dynamical-systems analysis with a stochastic Fokker-Planck treatment. The dynamics is governed by the competition between the Jeffery torque from the background shear and the inertial torque from settling. For configurations in which gravity lies in the shear plane, the azimuthal dynamics reduces to overdamped motion in a tilted periodic potential controlled by a single effective parameter $\mathcal{R}$ that combines the particle shape anisotropy and the settling strength. A saddle-node bifurcation on an invariant circle (SNIC) at $\mathcal{R}=1$ governs the transition from sustained rotational motion to steady equilibrium, with the rotation period diverging as $(1-\mathcal{R})^{-1/2}$. When gravity is parallel to the vorticity axis, the attractor is a periodic orbit for all settling strengths. The stochastic analysis reveals that noise plays a fundamentally different role depending on whether settling-induced potential barriers are present: in the classical Jeffery problem it diffuses over the orbit constant, whereas with settling it drives Kramers-type phase slips whose rate is exponentially sensitive to the Péclet number, defined as the ratio of diffusive to convective time scales. Langevin simulations confirm the predicted intermittent dynamics, with phase slips becoming progressively rarer as the barrier height or Péclet number increases. Asymptotic results in both the small- and large-$\mathrm{Pe}$ limits, together with numerical solutions of the Fokker-Planck equation at arbitrary $\mathrm{Pe}$, quantify the orientation moments across all regimes.

physics.flu-dyn↗

arXiv2026

Superhydrophobic Sand Mulch Shifts Soil Evaporation from Temperature-Controlled to Diffusion-Limited Regimes

In hot arid and semi-arid regions, substantial irrigation water is lost through surface evaporation under intense solar irradiation and high temperatures, limiting freshwater sustainability and crop productivity. Superhydrophobic Sand (SHS) mulch, a plastic-free, bio-inspired technology, has been proposed as a dry diffusion barrier to suppress evaporative losses. Here, we combine controlled column experiments with heat and mass transfer modeling to quantify how SHS thickness and soil properties govern evaporation under fixed irradiation. Relative to unmulched controls, a 5 mm SHS layer reduced evaporative flux by 65$\%$ in fine sand and 63$\%$ in coarse sand, while a 10 mm layer reduced flux by 83$\%$ and 70$\%$, respectively. Notably, soil-type trends reversed after mulching: although unmulched fine sand exhibited 37.5$\%$ higher evaporation than coarse sand, application of a 10 mm SHS layer reduced fine-sand evaporation to 40$\%$ below that of coarse sand. To explain this counterintuitive behavior, we developed a coupled heat and vapor transport model incorporating soil thermophysical properties and diffusion through the porous mulch layer. The model accurately predicted steady-state temperature profiles and evaporation rates for both mulched and unmulched systems. Our results show that SHS mulch shifts evaporation from a surface-temperature-controlled regime to a diffusion-limited regime governed by mulch thickness and soil thermal conductivity. This mechanistic understanding clarifies the performance of SHS and supports its potential to enhance irrigation efficiency in arid agricultural and landscaping applications.

physics.geo-ph↗

arXiv2026

RDumb++: Drift-Aware Continual Test-Time Adaptation

Continual Test-Time Adaptation (CTTA) seeks to update a pretrained model during deployment using only the incoming, unlabeled data stream. Although prior approaches such as Tent, EATA etc. provide meaningful improvements under short evolving shifts, they struggle when the test distribution changes rapidly or over extremely long horizons. This challenge is exemplified by the CCC benchmark, where models operate over streams of 7.5M samples with continually changing corruption types and severities. We propose RDumb++, a principled extension of RDumb that introduces two drift-detection mechanisms i.e entropy-based drift scoring and KL-divergence drift scoring, together with adaptive reset strategies. These mechanisms allow the model to detect when accumulated adaptation becomes harmful and to recover before prediction collapse occurs. Across CCC-medium with three speeds and three seeds (nine runs, each containing one million samples), RDumb++ consistently surpasses RDumb, yielding approx 3% absolute accuracy gains while maintaining stable adaptation throughout the entire stream. Ablation experiments on drift thresholds and reset strengths further show that drift-aware resetting is essential for preventing collapse and achieving reliable long-horizon CTTA.

cs.LG↗

arXiv2026

QUAIL: Quantization Aware Unlearning for Mitigating Misinformation in LLMs

Machine unlearning aims to remove specific knowledge (e.g., copyrighted or private data) from a trained model without full retraining. In practice, models are often quantized (e.g., 4-bit) for deployment, but we find that quantization can catastrophically restore forgotten information [1]. In this paper, we (1) analyze why low-bit quantization undermines unlearning, and (2) propose a quantization-aware unlearning method to mitigate this. We first compute weight-change statistics and bucket overlaps in quantization to show that typical unlearning updates are too small to cross quantization thresholds. Building on this insight, we introduce a logits space hinge loss: for each forget example, we force the output logits of the unlearned model to differ from the original model by at least a margin (half the quantization step). This ensures forgotten examples remain distinguishable even after quantization. We evaluate on language and classification tasks (including a Twitter misinformation dataset) and show our method preserves forgetting under 4-bit quantization, whereas existing methods almost entirely recover the forgotten knowledge.

cs.LG↗

arXiv2025

Transport and orientation of anisotropic particles settling in surface gravity waves

We study the translation and orientation dynamics of an anisotropic particle settling in monochromatic linear surface gravity waves. Recent work has shown that a neutrally buoyant spheroid attains a preferred mean orientation in such wave fields, independent of its initial state and determined solely by its aspect ratio. Comparing the settling parameter $\mathrm{Sv}$, the ratio of settling speed to wave speed, with the asymptotically small wave steepness $ε$, we investigate the long time dynamics of a negatively buoyant particle. We examine the transition from aspect ratio-dependent equilibrium orientation in the weak settling regime ($\mathrm{Sv} \ll ε^2$) to initial-condition-dependent alignment in the strong settling limit ($\mathrm{Sv} \gg 1$). Since translation and orientation are coupled for anisotropic particles, we use orientation dynamics to predict net horizontal transport. Fluid inertia induces an inertial torque that breaks the Stokesian degeneracy and drives broadside alignment. We analyze the influence of this torque on drift and alignment rate as functions of settling and wave parameters. Finally, we evaluate finite-size effects through the parameter $σ$, showing that a neutrally buoyant finite-size spheroid exhibits $σ$-dependent drift, validating the finite-size approximation when the spheroid size approaches the wavelength.

physics.flu-dyn↗

arXiv2025

How Surface Make-Up and Receding Electrokinetics Determine the Sign and Magnitude of Electrification at Water-Hydrophobe Interfaces?

It has been widely reported that as water contacts hydrophobic materials such as air or hydrocarbons (liquid or solid), the interfaces acquire a negative charge. It is not entirely clear whether this occurs due to the nature of water, or the hydrophobe, or purely the interface. Here, we probe the effects of surface chemistry and the speed of liquid-solid contact formation and separation on electrification. Glass capillaries grafted with mixed self-assembled monolayers of octadecyltrichlorosilane (ODTS) and (3-aminopropy)triethoxysilane (APTES) were exploited. Water was drawn inside these capillaries from an electroneutral reservoir, and the excess charge carried by the pendant droplets, if any, was quantified using an electrometer with a 100 fC resolution. Depending on the APTES content, the surface charge density at the water-hydrophobe interface ranged from negative (for ODTS) to near-neutral (for APTES 2-sec-exposure followed by ODTS) to positive APTES(5s)-ODTS. Next, we probed the charge (Q) contributions of the following steps on the electrification: (i) Contact (Q_1^n): as a dry capillary enters the water reservoir; (ii) Liquid uptake (Q_2^n): as water is uptaken; (iii) Capillary lift (Q_3^n): the filled capillary is removed from the reservoir; and (iv) Liquid release (Q_4^n): releasing the liquid back into the reservoir. This revealed that the electrification during water uptake (Q_2^n) varied with the rate of release during the previous cycle (Q_4^(n-1)), and it did not depend on the uptake rate. We explain these findings based on the electrical double layer theory, electrokinetics, and charge conservation, advancing the current understanding of electrification.

cond-mat.soft↗

arXiv2025

Mic-hackathon 2024: Hackathon on Machine Learning for Electron and Scanning Probe Microscopy

Microscopy is a primary source of information on materials structure and functionality at nanometer and atomic scales. The data generated is often well-structured, enriched with metadata and sample histories, though not always consistent in detail or format. The adoption of Data Management Plans (DMPs) by major funding agencies promotes preservation and access. However, deriving insights remains difficult due to the lack of standardized code ecosystems, benchmarks, and integration strategies. As a result, data usage is inefficient and analysis time is extensive. In addition to post-acquisition analysis, new APIs from major microscope manufacturers enable real-time, ML-based analytics for automated decision-making and ML-agent-controlled microscope operation. Yet, a gap remains between the ML and microscopy communities, limiting the impact of these methods on physics, materials discovery, and optimization. Hackathons help bridge this divide by fostering collaboration between ML researchers and microscopy experts. They encourage the development of novel solutions that apply ML to microscopy, while preparing a future workforce for instrumentation, materials science, and applied ML. This hackathon produced benchmark datasets and digital twins of microscopes to support community growth and standardized workflows. All related code is available at GitHub: https://github.com/KalininGroup/Mic-hackathon-2024-codes-publication/tree/1.0.0.1

cond-mat.mtrl-sci↗

arXiv2025

Disentangling the Roles of Dissolved Oxygen, Common Salts, and pH on the Spontaneous Hydrogen Peroxide Production in Water: No O2, No H2O2

Despite the mounting evidence proving that the air-water interface or the microdroplet geometry has nothing to do with the spontaneous formation of hydrogen peroxide (H2O2), the myth persists. Three recent studies by George and co-workers give credence to the myth by showing connections between the spontaneous formation of hydroxyl (HO) radicals and hydrogen peroxide (H2O2) in sprayed microdroplets with the solution pH, dissolved salts, nebulizing gas, and the gaseous environment. They report that among halides (chloride, bromide, and iodide), bromide dominates the H2O2 formation because of its ability to donate electrons. Also, they conclude that the H2O2 production at the air-water interface scales with waters alkalinity. In response, we apply a broad set of techniques, spanning NMR, potentiodynamic polarization, electron microscopy, and hydrogen peroxide assay kit (HPAK) fluorometry, to reexamine these claims. Our experiments reveal that regardless of the halide present in water, the air-water interface or the microdroplet geometry does not drive the H2O2 formation. It is the reduction of O2 at the solid-water interface that produces H2O2, i.e., in the absence of O2, no H2O2 is formed regardless of the halide ions. We explain the relative dependence of H2O2 concentrations on the halides based on their propensity to drive pitting corrosion (Chloride > Bromide > Iodide). As the pits appear in the passivating layer, exposing the metal, H2O2 is consumed in further oxidation. Next, we disprove the claim of alkalinity-driven H2O2 formation by demonstrating that aluminum and titanium surfaces produce more H2O2 in acidic and alkaline conditions, respectively. Taken together, these findings refute the conclusions of George and co-workers and others regarding spontaneous H2O2 generation at the air-water interface. The following mnemonic captures our conclusion: no O2, no H2O2.

cond-mat.soft↗

arXiv2024

Hydrogen peroxide forms spontaneously in water (bulk, film, or microdroplet) via reduction of dissolved oxygen at solid-water interface

Zare and co-workers have recently claimed that hydrogen peroxide is spontaneously generated on the air-water interface of sprayed microdroplets, i.e., that H2O2 forms without an external energy source or co-reactant or catalyst. Specifically, they find that the H2O2(aq) concentration in sprayed microdroplets increases by a factor of 3.5 (or 2.5) as the spray chamber's relative humidity (RH) is changed from 15% to 50% (or from 15% to 95%). Building on these results, they imply causation for the seasonality of viral infections arising from the RH-dependent H2O2 generation in environmental microdroplets. Here, we present an alternative explanation for their observations.

cond-mat.soft↗

arXiv2024

Why Some Metal Ions Spontaneously Form Nanoparticles in Water Microdroplets? Disentangling the Contributions of Air-Water Interface and Bulk Redox Chemistry

Water microdroplets containing 100 micromolar HAuCl4 have been shown to reduce gold ions into gold nanoparticles spontaneously. It has been suggested that this chemical transformation is driven by ultrahigh electric fields at the air-water interface, albeit without mechanistic insight. We investigated the fate of several metallic salts in water, methanol, ethanol, and acetonitrile in bulk and microdroplets. This revealed that when HAuCl4 (or PtCl4) is added to bulk water (or methanol or ethanol), metal NPs appear spontaneously. Over time, the nanoparticles grow in bulk, as evidenced by the solution's changing colors. If the same bulk solution is sprayed pneumatically and collected, the NP size has no significant enhancement. Interestingly, the reduction of metal ions is accompanied by the oxidation of water (or alcohols); however, these redox reactions are minimal in acetonitrile. We establish that the spontaneous reduction of metal ions is (i) not limited to water or gold ions, (ii) not driven by the air-water interface of microdroplets, and (iii) appears to be a general phenomenon for solvents containing hydroxyl groups. These results advance our understanding of liquids in general and should be relevant in soil chemistry, biogeochemistry, electrochemistry, and green chemistry.

cond-mat.soft↗

arXiv2024

Directional Wetting of Submerged Gas-entrapping Microtextures

Numerous natural and industrial processes entail the spontaneous entrapment of gas/air as rough/patterned surfaces are submerged under water. As the wetting transitions ensue, the gas diffuses into the water leading to the fully-water-filled state. However, the standard models for wetting do not account for the microtexture's topography. In other words, it is not clear whether the lifetime of n cavities arranged in a I-D line or a II-D (circular or square) lattice would be the same or not as a single 0-D cavity. In response, we tracked the time-dependent fates of gas pockets trapped in I-D and II-D lattices and compared them with wetting transitions in commensurate 0-D cavities. Interestingly, the wetting transitions in the I-D and the II-D arrays had a directionality such that the gas from the outermost cavities was lost the first, while the innermost got filled by water the last. In essence, microtexture's spatial organization afforded shielding to the loss of the gas from the innermost cavities, which we probed as a function of the microtexture's pitch, surface density, dimensionality, and hydrostatic pressure. These findings advance our knowledge of wetting transitions in microtextures, and inspiring surface textures to protect electronic devices against liquid ingression.

cond-mat.soft↗

arXiv2023

Coating-free Underwater Breathing via Biomimicry

Numerous natural and engineering scenarios necessitate entrapment of air pockets or bubbles on submerged surfaces, e.g., aquatic insects, smartphones, and membranes for separation and purification. Current technologies for bubble entrapment rely heavily on perfluorocarbon coatings, which limits their sustainability and applications. Here, we investigate doubly reentrant cavities, a biomimetic microtexture capable of entrapping air under wetting liquids, under static and dynamic pressure cycling. The effects of positive, negative, and positive-negative cycles are studied across a range of pressure amplitudes, ramp rates, intercycle intervals, and water column heights. Remarkably, the fate of the trapped air under pressure cycling falls into the following three distinct regimes: the bubble (i) monotonically depletes, (ii) remains indefinitely stable, or (iii) starts growing. This hitherto unrealized richness of underwater bubble dynamics will guide the development of coating-free underwater technologies and provide clues into the curious lives of air-breather aquatic/marine insects.

physics.flu-dyn↗

arXiv2023

Busting the Myth of Spontaneous Formation of H2O2 at the Air-Water Interface: Contributions of the Liquid-Solid Interface and Dissolved Oxygen Exposed

Recent reports on the spontaneous formation of H2O2(aq) at the air-water interface and the solid-water interface have been sensational. The speculated mechanism at the air-water interface is based on instantaneous ultrahigh electric fields and the micro-scale of droplets, whereas the solid-water interface is speculated to be the site for oxidation of water (or hydroxide ions) and reduction of the solid surface. We utilized 1H-NMR spectroscopy to investigate the effects of the nebulizing gas, the dissolved oxygen content, and solid substrates on the H2O2(aq) formation (detection limit 50 nM). Experiments revealed that contrary to the sensational claims, the air-water interface is not the site for H2O2(aq) formation; instead, it is the solid-water interface where H2O2(aq) is formed during the reduction of dissolved oxygen and oxidation of the solid surface. Curiously, the tendencies of solid substrates towards forming H2O2(aq) follow the classic Galvanic series. This report advances the current understanding of aquatic chemistry and should be relevant to corrosion science, surface science, and electrochemistry.

cond-mat.soft↗

arXiv2023

General Predictive Framework for Droplet Detachment Force

Liquid droplets hanging from solid surfaces are commonplace, but their physics is complex. Examples include dew or raindrops hanging onto wires or droplets accumulating onto a cover placed over warm food or windshields. In these scenarios, determining the force of detachment is crucial to rationally design technologies. Despite much research, a quantitative theoretical framework for detachment force remains elusive. In response, we interrogated the elemental droplet surface system via comprehensive laboratory and computational experiments. The results reveal that the Young Laplace equation can be utilized to accurately predict the droplet detachment force. When challenged against experiments with liquids of varying properties and droplet sizes, detaching from smooth and microtextured surfaces of wetting and non wetting chemical makeups, the predictions were in an excellent quantitative agreement. This study advances the current understanding of droplet physics and will contribute to the rational development of technologies.

cond-mat.soft↗

arXiv2022

Direct imaging of polymer filaments pulled from rebounding drops

Polymer filaments form the foundation of biology from cell scaffolding to DNA. Their study and fabrication play an important role in a wide range of processes from tissue engineering to molecular machines. We present a simple method to deposit stretched polymer fibers between micro-pillars. This occurs when a polymeric drop impacts on and rebounds from an inclined superhydrophobic substrate. It wets the top of the pillars and pulls out liquid filaments which are stretched and can attach to adjacent pillars leaving minuscule threads, with the solvent evaporating to leave the exposed polymers. We use high-speed video at the microscale to characterize the most robust filament-forming configurations, by varying the impact velocity, substrate structure and inclination angle, as well as the PEO-polymer concentration. Impacts onto plant leaves or randomized nano-structured surface leads to the formation of a branched structure, through filament mergers at the free surface of the drop. SEM shows the deposition of filament bundles which are thinner than those formed by evaporation or rolling drops. Raman spectroscopy identifies mode B stretched DNA filaments from aqueous-solution droplets.

physics.flu-dyn↗

arXiv2021

Zwitterions Layer at but Do Not Screen Electrified Interfaces

The role of ionic electrostatics in colloidal processes is well-understood in natural and applied contexts; however, the electrostatic contribution of zwitterions, known to be present in copious amounts in extremophiles, has not been extensively explored. In response, we studied the effects of glycine as a surrogate zwitterion, ion, and osmolyte on the electrostatic forces between negatively charged mica-mica and silica-silica interfaces. Our results reveal that while zwitterions layer at electrified interfaces and contribute to solutions' osmolality, they do not affect at all the surface potentials, the electrostatic surface forces (magnitude and range), and solutions' conductivity across 0.3-30 mM glycine concentration. We infer that the zwitterionic structure imposes an inseparability among the positive and negative charges and that this inseparability prevents the buildup of a counter charge at interfaces. These elemental experimental results pinpoint how zwitterions enable extremophiles to cope with the osmotic stress without affecting finely tuned electrostatic force balance.

cond-mat.soft↗