Search for: All records

Self-supervised learning(SSL) is essential to obtain foundation models in NLP and CV domains via effectively leveraging knowledge in large-scale unlabeled data. The reason for its success is that a suitable SSL design can help the model to follow the neural scaling law, i.e., the performance consistently improves with increasing model and dataset sizes. However, it remains a mystery whether existing SSL in the graph domain can follow the scaling behavior toward building Graph Foundation Models~(GFMs) with large-scale pre-training. In this study, we examine whether existing graph SSL techniques can follow the neural scaling behavior with the potential to serve as the essential component for GFMs. Our benchmark includes comprehensive SSL technique implementations with analysis conducted on both the conventional SSL setting and many new settings adopted in other domains. Surprisingly, despite the SSL loss continuously decreasing, no existing graph SSL techniques follow the neural scaling behavior on the downstream performance. The model performance only merely fluctuates on different data scales and model scales. Instead of the scales, the key factors influencing the performance are the choices of model architecture and pretext task design. This paper examines existing SSL techniques for the feasibility of Graph SSL techniques in developing GFMs and opens a new direction for graph SSL design with the new evaluation prototype. Our code implementation is available online to ease reproducibility https://github.com/HaitaoMao/GraphSSLScaling. 

Abstract Mercury (Hg) is a global pollutant whose atmospheric deposition is a major input to the terrestrial and oceanic ecosystems. Gas‐particle partitioning (GPP) of gaseous oxidized mercury (GOM) redistributes speciated Hg between gas and particulate phase and can subsequently alter Hg deposition flux. Most 3‐dimensional chemical transport models either neglected the Hg GPP process or parameterized it with measurement data limited in time and space. In this study, CMAQ‐newHg‐Br (Ye et al., 2018,https://doi.org/10.1002/2017ms001161) was updated to CMAQ‐newHg‐Br v2 by implementing a new GPP scheme and the most up‐to‐date Hg redox chemistry and was run for the northeastern United States over January‐November 2010. CMAQ‐newHg‐Br v2 reproduced the measured spatiotemporal distributions of gaseous elemental mercury (GEM) and particulate bound mercury (PBM) concentrations and Hg wet deposition flux within reasonable ranges and simulated dry deposition flux in agreement with previous studies. The GPP scheme improved the simulation of PBM via increasing winter‐, spring‐ and fall‐time PBM concentrations by threefold. It also improved simulated Hg wet deposition flux with an increase of 2.1 ± 0.7 μgm²in the 11‐month accumulated amount, offsetting half of the decreasing effect of the updated chemistry (−4.2 ± 1.8 μgm²). Further, the GPP scheme captured the observedK_p‐T relationship as reported in previous studies without using measurement data and showed advantages at night and in rural/remote areas where existing empirical parameterizations failed. Our study demonstrated CMAQ‐newHg‐Br v2 a promising assessment tool to quantify impacts of climate change and emission reduction policy on Hg cycling. 

The summertime surface ozone (O3) concentrations over Southeast Michigan (SEMI) often exceed 70 ppbv. However, the associated O3 formation regime is still not well known. In this study, we examined the chemical drivers of O3 exceedances in SEMI, based on the Michigan-Ontario Ozone Source Experiment (MOOSE) field campaign during the period of May 20 – June 30, 2021. We employed a zero-dimensional (0-D) box model, which was constrained by measurements of meteorology and trace gas concentrations during MOOSE. Our model simulations demonstrated that the formaldehyde to nitrogen dioxide ratio (HCHO/NO2) for the transition between the VOC- and NOx-limited O3 production regimes was 3.0 ± 0.3 (mean ± 1σ) in SEMI. The midday (12:00-16:00) averaged HCHO/NO2 ratio during MOOSE was 1.62 ± 1.03, suggesting that O3 production in SEMI was likely limited by VOC emissions. Our study has significant implications for air quality policy and the design of effective O3 pollution control strategies through ground-based HCHO/NO2 measurements and model simulations. 

Our ability to predict the structure and evolution of stars is in part limited by complex, 3D hydrodynamic processes such as convective boundary mixing. Hydrodynamic simulations help us understand the dynamics of stellar convection and convective boundaries. However, the codes used to compute such simulations are usually tested on extremely simple problems and the reliability and reproducibility of their predictions for turbulent flows is unclear. We define a test problem involving turbulent convection in a plane-parallel box, which leads to mass entrainment from, and internal-wave generation in, a stably stratified layer. We compare the outputs from the codes FLASH , MUSIC , PPMSTAR , PROMPI , and SLH , which have been widely employed to study hydrodynamic problems in stellar interiors. The convection is dominated by the largest scales that fit into the simulation box. All time-averaged profiles of velocity components, fluctuation amplitudes, and fluxes of enthalpy and kinetic energy are within ≲3 σ of the mean of all simulations on a given grid (128 3 and 256 3 grid cells), where σ describes the statistical variation due to the flow’s time dependence. They also agree well with a 512 3 reference run. The 128 3 and 256 3 simulations agree within 9% and 4%, respectively, on the total mass entrained into the convective layer. The entrainment rate appears to be set by the amount of energy that can be converted to work in our setup and details of the small-scale flows in the boundary layer seem to be largely irrelevant. Our results lend credence to hydrodynamic simulations of flows in stellar interiors. We provide in electronic form all outputs of our simulations as well as all information needed to reproduce or extend our study.