arXiv subjects

Harsh Jain

Publications and source records attributed to Harsh Jain.

7 recordsLinked to original sources

arXiv2026

Agent Stack for ML Training Lifecycle: Generation, Validation, and Optimization

Production model development requires repeated architecture design, platform integration, debugging, training, and evaluation. We present an agent stack deployed at LinkedIn with two independently usable agent loops: AutoTorch for model generation and migration, and AutoResearch for model optimization. AutoTorch generates PyTorch models and training pipelines for TensorFlow migration, XGBoost replacement, and new-model development. Executable validation hooks check platform API compliance, trainability, and conversion fidelity on sampled data, guiding iterative repair before engineer review and full-scale distributed training. AutoResearch explores architecture and training-configuration changes through expert-steered or autonomous experimentation. An external orchestrator manages parallel trials, experiment history, candidate selection, and wall-clock budgets across agent sessions. A shared runtime preserves execution state independently of agent reasoning and allocates GPUs separately from agent execution. Across ranking, content understanding, and recommendation workloads, AutoTorch completed generation and full-scale training in 92 of 115 workflow runs (80%). Of these, 77 (83.7%) produced models meeting post-training quality requirements. In a separate cohort, AutoResearch improved the primary task metric on held-out data for 28 of 100 production pipelines. Engineering teams estimate that the stack reduced elapsed end-to-end migration time from approximately eight weeks to two weeks, a 4x speedup.

cs.SE↗

arXiv2026

Intrinsic Paramagnetism of Optical Phonons and Thermal Hall Effect in Quantum Paraelectrics

We propose a microscopic theory of the phonon-mediated thermal Hall effect in quantum paraelectrics, driven by the intrinsic paramagnetism due to quantum fluctuations of the (nearly) critical optical phonon modes. The effective magnetization due to temporal fluctuations of the assosciated dipoles couple to an external magnetic field resulting the magnetoelectric term $\mathbf{H} \cdot (\mathbf{P} \times \dot{\mathbf{P}})$ in the effective Lagrangian density of the polarisation field $\mathbf{P}$. This coupling splits the transverse optical phonon spectrum into right- and left-circularly polarized chiral branches. Using the chiral phonon Green's functions and the Kubo formalism, we derive the thermal Hall conductivity $κ_{xy}$ from the heat-current correlation function. We observe that $κ_{xy}$ is proportional to the difference in spectral weights of the two chiral modes, $κ_{xy} \propto A_+^2 - A_-^2$, establishing a direct microscopic link between phonon chirality and transverse heat transport. Our theory predicts three key observations: (i) a linear-in-magnetic-field response $κ_{xy} \propto H$ at low fields, evolving into nonlinear behavior at higher fields (ii) a non-monotonic temperature dependence of $κ_{xy}$ with a peak arising from the competition between thermal population and phonon scattering and (iii) strong enhancement of $κ_{xy}$ near the quantum critical point $ω_0 \to 0$, it exhibits power-law scaling $κ_{xy} \propto T^α$ with $α\approx 0.41$. Our findings provide an understanding of thermal Hall transport in quantum paraelectrics due to intrinsic paramagnetism and is aligned with the observed behaviour in SrTiO$_3$ and related materials.

cond-mat.str-el↗

arXiv2026

Compression-controlled dynamic buckling in thin soft sheets

We investigate experimentally the dynamic phase transition from compressed to buckled phases for thin sheets of rubber. We find that the rubber strips enter a highly compressed, metastable state when compression speed is high. During the compressed phase, higher modes grow, followed by mode coarsening. Mode growth is accompanied by an expansion of length while the system is still being compressed. We measure the forces and length of the sheet to confirm this, and we develop a mechanism for how modes grow and coarsen during dynamical buckling. The influence of crucial control parameters in the experiments, such as the material cross section and compression speed, on the buckling dynamics, are explained theoretically.

cond-mat.soft↗

arXiv2026

Multi-particle gates on driven one-dimensional paths: probing deep traps

We study single-file transport of driven overdamped colloidal particles on a periodic path with deep potential wells. In the small trap limit (i.e., trap size smaller than particle size), the particle current transitions from zero to finite as the number of particles on the path exceeds a critical number $n_c$. Beyond this threshold, $n_c$ particles cluster behind the trap, demonstrating collective correlated motion. The remaining `extra' particles circulate, giving a finite current. We study this phenomenon numerically using overdamped Brownian dynamics simulations, and present an experimental realization of this behaviour for micron-scale colloidal particles driven in an optical vortex. Using our experimental observations, we present results characterizing potential wells as deep as several hundred $k_BT$.

cond-mat.soft↗

arXiv2023

Emergent reprogrammable mechanical memory in soft rods network via friction tuning

We present emergent mechanical memory storage behavior in soft cellular materials. The cellular materials are a network of soft hyperelastic rods which store shape changes, specifically local indentation. This happens under an applied global compressive strain on the material. The material transits under strain from an elastic state (capable of `forgetting' any applied indentation after un-indentation) to plastic state (indefinitely storing the shape change due to indentation). The memory can be erased via removal of applied global strains and is therefore re-programmable. We characterise this behaviour experimentally and present a simple model that makes use of friction for understanding this behavior.

cond-mat.soft↗

arXiv2020

Recharging and rejuvenation of decontaminated N95 masks

N95 respirators comprise a critical part of the personal protective equipment used by frontline health-care workers, and are typically meant for one-time usage. However, the recent COVID-19 pandemic has resulted in a serious shortage of these masks leading to a worldwide effort to develop decontamination and re-use procedures. A major factor contributing to the filtration efficiency of N95 masks is the presence of an intermediate layer of charged polypropylene electret fibers that trap particles through electrostatic or electrophoretic effects. This charge can degrade when the mask is used. Moreover, simple decontamination procedures (e.g. use of alcohol) can degrade any remaining charge from the polypropylene, thus severely impacting the filtration efficiency post decontamination. In this report, we summarize our results on the development of a simple laboratory setup allowing measurement of charge and filtration efficiency in N95 masks. In particular, we propose and show that it is possible to recharge the masks post-decontamination and recover filtration efficiency.

physics.med-ph↗

arXiv2019

Molding 3D curved structures by selective heating

It is of interest to fabricate curved surfaces in three dimensions from easily available homogeneous material in the form of flat sheets. The aim is not just to obtain a surface $M$ in $\mathbb{R}^3$ which has a desired intrinsic Riemannian metric, but to get the desired embedding $M \subset \mathbb{R}^3$ up to translations and rotations (the Riemannian metric alone need not uniquely determine this). In this paper, we demonstrate three generic methods of molding a flat sheet of thermo-responsive plastic by selective contraction induced by targeted heating. These methods do not involve any cutting and gluing, which is a property they share with origami. The first method is inspired by tailoring, which is the usual method for making garments out of plain pieces of cloth. Unlike usual tailoring, this method produces the desired embedding in $\mathbb{R}^3$, and in particular, we get the desired intrinsic Riemannian metric. The second method just aims to bring about the desired new Riemannian metric via an appropriate pattern of local contractions, without directly controlling the embedding. The third method is based on triangulation, and seeks to induce the desired local distances. This results in getting the desired embedding in $\mathbb{R}^3$, in particular, it also gives us the target Riemannian metric. The second and the third methods, and also the first method for the special case of surfaces of revolution, are algorithmic in nature. We give a theoretical account of these methods, followed by illustrated examples of different shapes that were physically molded by these methods.

cond-mat.soft↗