arXiv subjects

Jongwon Park

Publications and source records attributed to Jongwon Park.

At least 19 recordsLinked to original sources

arXiv2026

Predicting Space Groups of Double Perovskites by LLM with Dynamic Few-Shot Learning

Double perovskites (DPs) offer broad compositional tunability, but predicting the space groups (SGs) of stable structures remains difficult because available datasets are often strongly imbalanced toward dominant SG classes. We refer to dominant SG classes as major SGs and underrepresented classes as minor SGs. We introduce Dynamic and Diversity-enhanced Few-shot Retrieval and Rule-Guided Inference for Space-Group Prediction (DyRIS), an LLM-agent-based framework that predicts ranked SG candidates from a given DP composition. DyRIS uses diversity-enhanced dynamic few-shot prompting to retrieve relevant in-context examples while limiting the dominance of frequently represented SGs. It further incorporates rule-guided inference based on B/B' cation ordering, quantitative indicators, and major-SG bias control to refine and rank the final Top-3 SG candidates. We evaluate DyRIS on 3,528 thermodynamically filtered DP entries and compare it with composition-based and descriptor-based baselines. At a training-data ratio of 0.5, DyRIS achieves competitive overall accuracy while obtaining the best Overall Top-1 macro-F1 score and the best performance across all Minor-SG metrics. DyRIS improves Minor-SG Top-1 accuracy by 3.26 percentage points relative to CrabNet and achieves higher Minor-SG Top-3 accuracy than the strongest PyCaret-based baseline. Ablation studies show that diversity-enhanced retrieval, quantitative indicators, major-SG bias control, and B/B' ordering information each contribute to prediction performance. Additional experiments show that the final rule-guided inference step is not easily replaced by conventional classifier- or ranker-based models. These findings demonstrate the potential of combining retrieval-based LLM reasoning with crystallographic domain knowledge for SG prediction in imbalanced materials datasets.

cs.AI↗

arXiv2026

Population III star formation in an X-ray background: IV. On-the-fly calculation of radiation backgrounds and their impact on the intergalactic medium

In this paper, part of a series on the effects of X-ray sources in promoting Population III (Pop III) star formation, we investigate the ionisation and heating of the intergalactic medium (IGM) and the consequent enhancement of molecular hydrogen (H$_{2}$) and Pop III formation using cosmological zoom-in simulations. We adopt a minimal X-ray feedback model in which X-rays originate solely from Pop III supernovae, and compute the global X-ray and Lyman-Werner (LW) radiation backgrounds on-the-fly during the simulation of a mean-density region of the Universe. This approach self-consistently captures the feedback loop between Pop III stars and the radiation backgrounds they produce. Pop III supernovae generate a weak X-ray background (J$_{\mathrm{X,21}} \sim 10^{-5}$) and a moderate LW background (J$_{\mathrm{LW,21}} \sim 10^{-1}$); the latter intensifies below $z \approx 12$ (J$_{\mathrm{LW,21}} \sim 10^{1}-10^{2}$) with the onset of Pop II star formation. Applying these backgrounds to regions of varying mean density produces a net positive X-ray feedback that increases the Pop III number density, with stronger enhancement in underdense regions. The positive feedback is more pronounced when the X-ray background is computed on-the-fly rather than by post-processing, demonstrating the importance of the feedback loop. The X-ray background also raises the Thomson scattering optical depth at high redshift, while the total optical depth remains consistent with Planck 2018 constraints. Because our model includes only Pop III supernovae as X-ray sources, it represents the most conservative scenario; stronger X-ray feedback is expected when additional sources are included, as will be explored in future work.

astro-ph.GA↗

arXiv2026

Population III star formation in an X-ray background: V. Environmental dependence and halo occupation probability

An X-ray background in the early Universe enhances molecular hydrogen formation, the main coolant of primordial gas, thereby lowering the threshold for Pop III star formation. Continuing our series on X-ray impacts on Pop III star formation, we investigate how a soft X-ray background promotes Pop III star formation using cosmological zoom-in simulations of ten cosmic volumes spanning a range of halo number densities. Each volume is irradiated by the Lyman-Warner (LW) H$_{2}$ dissociating background and a weak (J$_{21} \sim 10^{-5}$), soft ($E \sim 0.2-2.0$ keV) X-ray background produced by pair-instability SNe (PISNe) from Pop III stars and calculated self-consistently as described in a companion paper. We also compare the same simulations with and without X-rays to isolate the X-ray effect. The background promotes Pop III star formation in two ways: (1) by reducing the mean host halo mass by a factor of $\sim 2-3$, and (2) by enabling Pop III star formation in haloes that would otherwise remain sterile, thereby increasing the halo occupation fraction. The resulting gain in Pop III number density is largest in underdense regions (a factor of $\approx 3$ on average, reaching up to 7). In the most extreme case, Pop II stars form only in the presence of X-rays and the gas-phase metallicity rises by an order of magnitude, suggesting that dwarf galaxies in underdense regions may be significantly influenced by an early X-ray background. We also provide fitting functions for the halo occupation probability of Pop III stars as a function of redshift for both X-ray and LW-only simulations, which can serve as inputs for semi-analytic models.

astro-ph.GA↗

arXiv2026

HetCCL: Accelerating LLM Training with Heterogeneous GPUs

The rapid growth of large language models is driving organizations to expand their GPU clusters, often with GPUs from multiple vendors. However, current deep learning frameworks lack support for collective communication across heterogeneous GPUs, leading to inefficiency and higher costs. We present HetCCL, a collective communication library that unifies vendor-specific backends and enables RDMA-based communication across GPUs without requiring driver modifications. HetCCL introduces two novel mechanisms that enable cross-vendor communication while leveraging optimized vendor libraries, NVIDIA NCCL and AMD RCCL. Evaluations on a multi-vendor GPU cluster show that HetCCL matches NCCL and RCCL performance in homogeneous setups while uniquely scaling in heterogeneous environments, enabling practical, high-performance training with both NVIDIA and AMD GPUs without changes to existing deep learning applications.

cs.DC↗

arXiv2025

Enhanced Conditional Generation of Double Perovskite by Knowledge-Guided Language Model Feedback

Double perovskites (DPs) are promising candidates for sustainable energy technologies due to their compositional tunability and compatibility with low-energy fabrication, yet their vast design space poses a major challenge for conditional materials discovery. This work introduces a multi-agent, text gradient-driven framework that performs DP composition generation under natural-language conditions by integrating three complementary feedback sources: LLM-based self-evaluation, DP-specific domain knowledge-informed feedback, and ML surrogate-based feedback. Analogous to how knowledge-informed machine learning improves the reliability of conventional data-driven models, our framework incorporates domain-informed text gradients to guide the generative process toward physically meaningful regions of the DP composition space. Systematic comparison of three incremental configurations, (i) pure LLM generation, (ii) LLM generation with LLM reasoning-based feedback, and (iii) LLM generation with domain knowledge-guided feedback, shows that iterative guidance from knowledge-informed gradients improves stability-condition satisfaction without additional training data, achieving over 98% compositional validity and up to 54% stable or metastable candidates, surpassing both the LLM-only baseline (43%) and prior GAN-based results (27%). Analyses of ML-based gradients further reveal that they enhance performance in in-distribution (ID) regions but become unreliable in out-of-distribution (OOD) regimes. Overall, this work provides the first systematic analysis of multi-agent, knowledge-guided text gradients for DP discovery and establishes a generalizable blueprint for MAS-driven generative materials design aimed at advancing sustainable technologies.

cs.AI↗

arXiv2025

Federated Recommender System with Data Valuation for E-commerce Platform

Federated Learning (FL) is gaining prominence in machine learning as privacy concerns grow. This paradigm allows each client (e.g., an individual online store) to train a recommendation model locally while sharing only model updates, without exposing the raw interaction logs to a central server, thereby preserving privacy in a decentralized environment. Nonetheless, most existing FL-based recommender systems still rely solely on each client's private data, despite the abundance of publicly available datasets that could be leveraged to enrich local training; this potential remains largely underexplored. To this end, we consider a realistic scenario wherein a large shopping platform collaborates with multiple small online stores to build a global recommender system. The platform possesses global data, such as shareable user and item lists, while each store holds a portion of interaction data privately (or locally). Although integrating global data can help mitigate the limitations of sparse and biased clients' local data, it also introduces additional challenges: simply combining all global interactions can amplify noise and irrelevant patterns, worsening personalization and increasing computational costs. To address these challenges, we propose FedGDVE, which selectively augments each client's local graph with semantically aligned samples from the global dataset. FedGDVE employs: (i) a pre-trained graph encoder to extract global structural features, (ii) a local valid predictor to assess client-specific relevance, (iii) a reinforcement-learning-based probability estimator to filter and sample only the most pertinent global interactions. FedGDVE improves performance by up to 34.86% on recognized benchmarks in FL environments.

cs.LG↗

arXiv2025

Thunder-LLM: Efficiently Adapting LLMs to Korean with Minimal Resources

Since state-of-the-art LLMs often underperform in languages other than English or Chinese, improving the capability of LLMs in new languages has become an essential task. Moreover, LLMs' entire end-to-end training process remains largely unknown to the public due to proprietary reasons, technical complexity, inconsistent documentation, and ethical considerations. The complete picture remains a closely guarded secret within the industry. This paper presents methods to adapt an existing English-based LLM to Korean in a low-budget scenario. We describe the entire end-to-end process: collecting Korean datasets, preprocessing the data, training the model, creating downstream benchmarks, and conducting evaluations. The evaluation results indicate that our method can effectively and cost-efficiently add new language capabilities to existing LLMs. Our new bilingual models, Thunder-LLM and Thunder-LLM-Ins, achieve superior Korean performance compared to state-of-the-art models while utilizing minimal data and computational resources. We share our comprehensive experience and make the code publicly available.

cs.CL↗

arXiv2025

CIMAGE: Exploiting the Conditional Independence in Masked Graph Auto-encoders

Recent Self-Supervised Learning (SSL) methods encapsulating relational information via masking in Graph Neural Networks (GNNs) have shown promising performance. However, most existing approaches rely on random masking strategies in either feature or graph space, which may fail to capture task-relevant information fully. We posit that this limitation stems from an inability to achieve minimum redundancy between masked and unmasked components while ensuring maximum relevance of both to potential downstream tasks. Conditional Independence (CI) inherently satisfies the minimum redundancy and maximum relevance criteria, but its application typically requires access to downstream labels. To address this challenge, we introduce CIMAGE, a novel approach that leverages Conditional Independence to guide an effective masking strategy within the latent space. CIMAGE utilizes CI-aware latent factor decomposition to generate two distinct contexts, leveraging high-confidence pseudo-labels derived from unsupervised graph clustering. In this framework, the pretext task involves reconstructing the masked second context solely from the information provided by the first context. Our theoretical analysis further supports the superiority of CIMAGE's novel CI-aware masking method by demonstrating that the learned embedding exhibits approximate linear separability, which enables accurate predictions for the downstream task. Comprehensive evaluations across diverse graph benchmarks illustrate the advantage of CIMAGE, with notably higher average rankings on node classification and link prediction tasks. Notably, our proposed model highlights the under-explored potential of CI in enhancing graph SSL methodologies and offers enriched insights for effective graph representation learning.

cs.LG↗

arXiv2024

Finetuning Pre-trained Model with Limited Data for LiDAR-based 3D Object Detection by Bridging Domain Gaps

LiDAR-based 3D object detectors have been largely utilized in various applications, including autonomous vehicles or mobile robots. However, LiDAR-based detectors often fail to adapt well to target domains with different sensor configurations (e.g., types of sensors, spatial resolution, or FOVs) and location shifts. Collecting and annotating datasets in a new setup is commonly required to reduce such gaps, but it is often expensive and time-consuming. Recent studies suggest that pre-trained backbones can be learned in a self-supervised manner with large-scale unlabeled LiDAR frames. However, despite their expressive representations, they remain challenging to generalize well without substantial amounts of data from the target domain. Thus, we propose a novel method, called Domain Adaptive Distill-Tuning (DADT), to adapt a pre-trained model with limited target data (approximately 100 LiDAR frames), retaining its representation power and preventing it from overfitting. Specifically, we use regularizers to align object-level and context-level representations between the pre-trained and finetuned models in a teacher-student architecture. Our experiments with driving benchmarks, i.e., Waymo Open dataset and KITTI, confirm that our method effectively finetunes a pre-trained model, achieving significant gains in accuracy.

cs.CV↗

arXiv2024

Violent starbursts and quiescence induced by FUV radiation feedback in metal-poor galaxies at high-redshift

JWST observations of galaxies at $z\gtrsim 8$ suggest that they are more luminous and clumpier than predicted by most models, prompting several proposals on the physics of star formation and feedback in the first galaxies. In this paper, we focus on the role of ultraviolet (UV) radiation in regulating star formation by performing a set of cosmological radiation hydrodynamics simulations of one galaxy at sub-pc resolution with different radiative feedback models. We find that the suppression of cooling by far UV (FUV) radiation (i.e., $\mathrm{H_2}$ dissociating radiation) from Pop II stars is the main physical process triggering the formation of compact and massive star clusters and is responsible for the bursty star formation observed in metal-poor galaxies at $z\gtrsim 10$. Indeed, artificially suppressing FUV radiation leads to a less intense continuous mode of star formation distributed into numerous, but low-mass open star clusters. Due to the intense FUV field, low-metallicity clouds remain warm ($\sim 10^4\,\mathrm{K}$) until they reach a relatively high density ($\gtrsim 10^3\,\mathrm{cm^{-3}}$), before becoming self-shielded and transitioning to a colder ($\sim 100\,\mathrm{K}$), partially molecular phase. As a result, star formation is delayed until the clouds accumulate enough mass to become gravitationally unstable. At this point, the clouds undergo rapid star formation converting gas into stars with high efficiency. We, therefore, observe exceptionally bright galaxies (ten times brighter than for continuous star formation) and subsequent quenched "dead" galaxies that did not form stars for tens of Myrs.

astro-ph.GA↗

arXiv2024

On the origin of outward migration of Population III stars

Outward migration of massive binary stars or black holes in their circumbinary disc is often observed in simulations and it is key to the formation of wide black hole binaries. Using numerical simulations of Population III (Pop III) star formation, we study the angular momentum of Pop III binaries and the torques between stars and gas discs to understand the origin of outward migration and high ellipticity. The outward migration of protostars is produced by gravitational torques exerted on them by their circumstellar minidiscs. The minidiscs, on the other hand, migrate outward mainly by gaining angular momentum by accreting gas from the circumbinary disc. The angular momentum transfer is most efficient for rapidly accreting equal-mass binaries, and weaker when the secondary mass is small or the massive companion evaporates the gas disc via radiative feedback. We conclude that outward migration and the formation of wide equal-mass massive binaries is common in metal-free/metal-poor star formation, mainly driven by their large accretion rates. We expect that the lower gas temperature and accretion rates in metal-enriched circumstellar discs would lead more often to inward migration and closer binary separations. We also observe inward migration for smaller mass Pop III protostars/fragments, leading to the rapid merging of sink particles and likely the formation of close binary black holes that, however, reach separations below the resolution of our simulations. We discuss the implications that Pop III separations and ellipticity may have on the interpretation that gravitational wave signals from merging intermediate-mass black holes come from Pop III remnants.

astro-ph.GA↗

arXiv2022

Star Cluster Formation and Survival in the First Galaxies

Using radiation-hydrodynamic cosmological simulations, we present a detailed ($0.1$ pc resolution), physically motivated portrait of a typical-mass dwarf galaxy before the epoch of reionization, resolving the formation and evolution of star clusters into individual $10\:\mathrm{M_{\odot}}$ star particles. In the rest-frame UV, the galaxy has an irregular morphology with no bulge or galactic disk, dominated by light emitted from numerous, compact, and gravitationally-bound star clusters. This is especially interesting in light of recent HST and JWST observations that -- aided by the magnifying power of gravitational lenses -- have imaged, at parsec-scale resolution, individual young star clusters in the process of forming in similar galaxies at $z>6$. Because of their low metallicities and high temperatures, star-forming gas clouds in this galaxy have densities $\sim 100$ times higher than typical giant molecular clouds; hence, their star formation efficiencies are high enough ($f_*\sim10-70$ per cent) to produce a sizeable population of potential globular cluster progenitors but typically smaller (between a few $100\:-\: 2\times10^4\:\mathrm{M_{\odot}}$, sizes of $0.1-3$ pc) and of lower metallicities ($10^{-3.5}-10^{-2.5}\:\mathrm{Z_{\odot}}$). The initial mass function of the star-forming clouds is log-normal while the bound star cluster mass function is a power-law with a slope that depends mainly on $f_*$ but also on the temporal proximity to a major starburst. We find slopes between $-0.5$ and $-2.5$ depending on the assumed sub-grid $f_*$. Star formation is self-regulated on galactic scales; however, the multi-modal metallicity distribution of the star clusters and the fraction of stars locked into surviving bound star clusters depends on $f_*$.

astro-ph.GA↗

arXiv2022

Population III star formation in an X-ray background: III. Periodic radiative feedback and luminosity induced by elliptical orbits

We model Pop III star formation in different FUV and X-ray backgrounds, including radiation feedback from protostars. We confirm previous results that a moderate X-ray background increases the number of Pop III systems per unit cosmological volume, but masses and multiplicities of the system are reduced. The stellar mass function also agrees with previous results, and we confirm the outward migration of the stars within the protostellar discs. We find that nearly all Pop III star systems are hierarchical, i.e., binaries of binaries. Typically, two equal-mass stars form near the centre of the protostellar disc and migrate outward. Around these stars, mini-discs fragment forming binaries that also migrate outward. Stars may also form at Lagrange points L4/L5 of the system. Afterward, star formation becomes more stochastic due to the large multiplicity, and zero-metallicity low-mass stars can form when rapidly ejected from the disc. Stars in the disc often have eccentric orbits, leading to a periodic modulation of their accretion rates and luminosities. At the pericenter, due to strong accretion, the star can enter a red-supergiant phase reaching nearly Eddington luminosity in the optical bands ($m_{\rm AB} \sim 34$ for a $100~M_{odot}$ star at $z=6$). During this phase, the star, rather than its nebular lines, can be observed directly by JWST, if sufficiently magnified by a gravitational lens. The $\sim 10,000$ AU separations and high eccentricities of many Pop III star binaries in our simulations are favorable parameters for IMBH mergers - and gravitational waves emission - through orbital excitation by field stars.

astro-ph.GA↗

arXiv2022

Resolving Class Imbalance for LiDAR-based Object Detector by Dynamic Weight Average and Contextual Ground Truth Sampling

An autonomous driving system requires a 3D object detector, which must perceive all present road agents reliably to navigate an environment safely. However, real-world driving datasets often suffer from the problem of data imbalance, which causes difficulties in training a model that works well across all classes, resulting in an undesired imbalanced sub-optimal performance. In this work, we propose a method to address this data imbalance problem. Our method consists of two main components: (i) a LiDAR-based 3D object detector with per-class multiple detection heads where losses from each head are modified by dynamic weight average to be balanced. (ii) Contextual ground truth (GT) sampling, where we improve conventional GT sampling techniques by leveraging semantic information to augment point cloud with sampled ground truth GT objects. Our experiment with KITTI and nuScenes datasets confirms our proposed method's effectiveness in dealing with the data imbalance problem, producing better detection accuracy compared to existing approaches.

cs.CV↗

arXiv2021

Population III Star Formation in an X-ray background: I. Critical Halo Mass of Formation and Total Mass in Stars

The first luminous objects forming in the universe produce radiation backgrounds in the FUV and X-ray bands that affect the formation of Population III stars. Using a grid of cosmological hydrodynamics zoom-in simulations, we explore the impact of the Lyman-Warner (LW) and X-ray radiation backgrounds on the critical dark matter halo mass for Population III star formation and the total mass in stars per halo. We find that the LW radiation background lowers the H$_2$ fraction and delays the formation of the Population III stars. On the other hand, X-ray irradiation anticipates the redshift of collapse and reduces the critical halo mass, unless the X-ray background is too strong and gas heating shuts down gas collapse into the halos and prevents star formation. Therefore, an X-ray background can increase the number of dark matter halos forming Population III stars by about a factor of ten, but the total mass in stars forming in each halo is reduced. This is because X-ray radiation increases the molecular fraction and lowers the minimum temperature of the collapsing gas (or equivalently the mass of the quasi-hydrostatic core) and therefore slows down the accretion of the gas onto the central protostar.

astro-ph.GA↗

arXiv2021

Population III Star Formation in an X-ray background: II. Protostellar Discs, Multiplicity and Mass Function of the Stars

Disc fragmentation plays an important role in determining the number of primordial stars (Pop III stars), their masses, and hence the initial mass function. In this second paper of a series, we explore the effect of uniform FUV H$_2$-photodissociating and X-ray radiation backgrounds on the formation of Pop~III stars using a grid of high-resolution zoom-in simulations. We find that, in an X-ray background, protostellar discs have lower surface density and higher Toomre $Q$ parameter, so they are more stable. For this reason, X-ray irradiated discs undergo fewer fragmentations and typically produce either binary systems or low-multiplicity systems. In contrast, the cases with weak or no X-ray irradiation produce systems with a typical multiplicity of $6 \pm 3$. In addition, the most massive protostar in each system is smaller by roughly a factor of two when the disc is irradiated by X-rays, due to lower accretion rate. With these two effects combined, the initial mass function of fragments becomes more top-heavy in a strong X-ray background and is well described by a power-law with slope $1.53$ and high-mass cutoff of $61$ M$_\odot$. Without X-rays, we find a slope $0.49$ and cutoff mass of $229$ M$_\odot$. Finally, protostars migrate outward after their formation due to the accretion of high-angular momentum gas from outside and the migration is more frequent and significant in absence of X-ray irradiation.

astro-ph.GA↗

arXiv2017

Star Formation of Merging Disk Galaxies with AGN Feedback Effects

Using numerical hydrodynamics code, we perform various idealized galaxy merger simulations to study the star formation (SF) of two merging disk galaxies. Our simulations include gas accretion onto supermassive black holes and active galactic nucleus (AGN) feedback. By comparing AGN simulations with those without AGNs, we attempt to understand when the AGN feedback effect is significant. With ~70 simulations, we investigated the SF with the AGN effect in mergers with variety of mass ratios, inclinations, orbits, galaxy structures and morphologies. Using these merger simulations with AGN feedback, we measure merger-driven SF using the burst efficiency parameter introduced by Cox et al. We confirm the previous studies that, in galaxy mergers, AGN suppresses SF more efficiently than in isolated galaxies. However, we additionally find that the effect of AGNs on SF is larger in major mergers than in minor mergers. In minor merger simulations with different primary bulge-to-total ratios, the effect of bulge fraction on the merger-driven SF decreases due to AGN feedback. We create models of Sa, Sb and Sc type galaxies and compare their SF properties while undergoing mergers. With the current AGN prescriptions, the difference in merger-driven SF is not as pronounced as that in the recent observational study of Kaviraj. We discuss the implications of this discrepancy.

astro-ph.GA↗

arXiv2015

Demographics of Isolated Galaxies Along the Hubble Sequence

Isolated galaxies in low-density regions are significant in the sense that they are least affected by the hierarchical pattern of galaxy growth, and interactions with perturbers, at least for the last few Gyr. To form a comprehensive picture of the star formation history of isolated galaxies, we constructed a catalog of isolated galaxies and their comparison sample in relatively denser environments. The galaxies are drawn from the SDSS DR7 in the redshift range of $0.025<z<0.044$. We performed a visual inspection and classified their morphology following the Hubble classification scheme. For the spectroscopic study, we make use of the OSSY catalog. We confirm most of the earlier understanding on isolated galaxies. The most remarkable additional results are as follows. Isolated galaxies are dominantly late type with the morphology distribution (E: S0: S: Irr) = (9.9: 11.3: 77.6: 1.2)\%. The frequency of elliptical galaxies among isolated galaxies is only a third of that of the comparison sample. Most of the photometric and spectroscopic properties are surprisingly similar between isolated and comparison samples. However, early-type isolated galaxies are less massive by 50\% and younger (by H$β$) by 20\% than their counterparts in the comparison sample. This can be explained as a result of different merger and star formation histories for differing environments in the hierarchical merger paradigm. We provide an on-line catalog for the list and properties of our sample galaxies.

astro-ph.GA↗