Search for: All records

Coreset selection, a technique for compressing large datasets while preserving performance, is crucial for modern machine learning. This paper presents a novel method for generating high-quality Wasserstein coresets using the Sinkhorn loss, a powerful tool with computational advantages. However, existing approaches suffer from numerical instability in Sinkhorn’s algorithm. We address this by proposing stable algorithms for the computation and differentiation of the Sinkhorn optimization problem, including an analytical formula for the derivative of the Sinkhorn loss and a rigorous stability analysis of our method. Extensive experiments demonstrate that our approach significantly outperforms existing methods in terms of sample selection quality, computational efficiency, and achieving a smaller Wasserstein distance. 

Training modern LLMs is extremely resource intensive, and customizing them for various deployment scenarios characterized by limited compute and memory resources through repeated training is impractical. In this paper, we introduce Flextron, a network architecture and post-training model optimization framework supporting flexible model deployment. The Flextron architecture utilizes a nested elastic structure to rapidly adapt to specific user-defined latency and accuracy targets during inference with no additional fine-tuning required. It is also input-adaptive, and can automatically route tokens through its sub-networks for improved performance and efficiency. We present a sample-efficient training method and associated routing algorithms for systematically transforming an existing trained LLM into a Flextron model. We evaluate Flextron on the GPT-3 and LLama-2 family of LLMs, and demonstrate superior performance over multiple end-to-end trained variants and other state-of-the-art elastic networks, all with a single pretraining run that consumes a mere 7.63% tokens compared to original pretraining. 

We demonstrate a methodology for diagnosing the multiscale dynamics and energy transfer in complex HED flows with realistic driving and boundary conditions. The approach separates incompressible, compressible, and baropycnal contributions to energy scale-transfer and quantifies the direction of these transfers in (generalized) wavenumber space. We use this to compare the kinetic energy (KE) transfer across scales in simulations of 2D axisymmetric vs fully 3D laser-driven plasma jets. Using the FLASH code, we model a turbulent jet ablated from an aluminum cone target in the configuration outlined by Liao et al. [Phys. Plasmas, 26 032306 (2019)]. We show that, in addition to its well known bias for underestimating hydrodynamic instability growth, 2D modeling suffers from significant spurious energization of the bulk flow by a turbulent upscale cascade. In 2D, this arises as vorticity and strain from instabilities near the jet's leading edge transfer KE upscale, sustaining a coherent circulation that helps propel the axisymmetric jet farther (≈25% by 3.5 ns) and helps keep it collimated. In 3D, the coherent circulation and upscale KE transfer are absent. The methodology presented here may also help with inter-model comparison and validation, including future modeling efforts to alleviate some of the 2D hydrodynamic artifacts highlighted in this study. 

This paper investigates the mechanical behaviour of a bi-layered panel containing many particles in one layer and demonstrates the size effect of particles on the deflection. The inclusion-based boundary element method (iBEM) considers a fully bounded bi-material system. The fundamental solution for two-jointed half spaces has been used to acquire elastic fields resulting from source fields over inclusions and boundary-avoiding multi-domain integral along the interface. Eshelby’s equivalent inclusion method is used to simulate the material mismatch with a continuously distributed eigenstrain field over the equivalent inclusion. The eigenstrain is expanded at the centre of the inclusion, which provides tailorable accuracy based on the order of the polynomial of the eigenstrain. As a single-domain approach, the iBEM algorithm is particularly suitable for conducting virtual experiments of bi-layered composites with many defects or reinforcements for both local analysis and homogenization purposes. The maximum deflection of solar panel coupons is studied under uniform vertical loading merged with inhomogeneities of different material properties, dimensions and volume fractions. The size of defects or reinforcements plays a significant role in the deflection of the panel, even with the same volume fraction, as the substrate is relatively thin. 

To provide energy-efficient space heating and cooling, a thermoelectric building envelope (TBE) embeds thermoelectric devices in building walls. The thermoelectric device in the building envelope can provide active heating and cooling without requiring refrigerant use and energy transport among subsystems. Thus, the TBE system is energy and environmentally friendly. A few studies experimentally investigated the TBE under limited operating conditions, and only simplified models for the commercial thermoelectric module (TEM) were developed to quantify its performance. A holistic approach to optimum system performance is needed for the optimal system design and operation. The study developed a holistic TBE-building system model in Modelica for system simulation and performance analysis. A theoretical model for a single TEM was first established based on energy conversion and thermoelectric principles. Subsequently, a TBE prototype model combining the TEM model was constructed. The prototype model employing a feedback controller was used in a whole building system simulation for a residential house. The system model computed the overall building energy efficiency and dynamic indoor conditions under varying operating conditions. Simulation results indicate the studied TBE system can meet a heating demand to maintain the desired room temperature at 20 °C when the lowest outdoor temperature is at -26.3 °C, with a seasonal heating COP near 1.1, demonstrating a better heating performance than electric heaters. It suggests a potential energy-efficient alternative to the traditional natural gas furnaces and electric heaters for space heating. 

Because of the thermoelectric (TE) effect (or Seebeck effect), a difference of potential is generated as a consequence of a temperature gradient across a sample. The TE effect has been mostly studied and engineered in semiconducting materials and it already finds several commercial applications. Only recently the TE effect in cement-based materials has been demonstrated and there is a growing interest in its potential. For instance, a temperature gradient across the external walls of a building can be used to generate electricity. By the inverse of the TE effect (or Peltier effect), one can also seek to control the indoor temperature of a building by biasing TE elements embedded in its external walls. In designing possible applications, the TE properties of cement-based materials must be determined as a function of their chemical composition. For instance, the TE properties of cement paste can be enhanced by the addition of metal oxide (e.g., Fe2O3) powder. In this paper, a single thermoelectric leg is studied using the finite element method. Metal oxide additives in the cement paste are modelled as spherical inhomogeneities. The thermoelectric properties of the single components are based on experimental data, while the overall thermoelectric properties of the composites are obtained from the numerical model. The results of this numerical study are interpreted according to the effective medium theory (EMT) and its generalisation (GEMT). 

Cable bacteria are multicellular filamentous bacteria that conduct electrons nonlocally between anoxic and oxic sediment regions, creating characteristic electrogenic pH fingerprints. These microbes aggregate in 3D patterns near biogenic structures, and filament fragments are also dispersed throughout deposits. Utilizing pH-sensitive planar optodes to investigate the dynamic response of electrogenic pH fingerprints to sediment reworking, we found that mobile bioturbators like nereid polychaetes (ragworms) can disturb the pH signatures. Sudden sediment disturbance associated with burrows at sub- to multi-centimeter scales eliminates detection of pH signatures. However, electrogenic pH fingerprints can recover in as little as 13 h near abandoned, closed burrows. Sequential collapse and regeneration of electrogenic pH fingerprints are associated with occupied and dynamic burrow structures, with the response time positively related to the scale of disturbance. In the case of relatively stable tube structures, built by benthos like spionid polychaetes and extending mm to cm into deposits, the electrogenic pH fingerprint is evident around the subsurface tubes. Cable filaments clearly associate with subsurface regions of enhanced solute exchange (oxidant supply) and relatively stable biogenic structures, including individual tubes and patches of tubes (e.g. made by Sabaco , a bamboo worm). Physically stable environments, favorable redox gradients, and enhanced organic/inorganic substrate availability promote the activity of cable bacteria in the vicinity of tubes and burrows. These findings suggest complex interactions between electrogenic activity fingerprints and species-specific patterns of bioturbation at multiple spatial and temporal scales, and a substantial impact of electrogenic metabolism on subsurface pH and early diagenetic reaction distributions in bioturbated deposits. 

Buildings use 40% of the global energy consumption and emit 30% of CO2 emissions. Of the total building energy, 30-40% is for building heating and cooling systems, which regulate the indoor thermal environment and provide thermal comfort to occupants. Most buildings use forced air technology in the United States to deliver heating/cooling to the targeted thermal zones. Researchers have suggested using radiant heating and cooling systems as a better alternative to all-air systems. Radiant systems supply heating or cooling directly to the building space using radiation released by the heated or cooled building enclosure via the embedded heating or cooling tubes. It is unsure whether the radiant heating and cooling system can provide better thermal comfort to occupants. Moreover, the evaluation method for thermal comfort in the current standard is only suitable for forced air systems. A new plan shall be developed to evaluate the radiation system’s thermal comfort. This paper reviews the experiment-based studies on the thermal comfort of radiant systems. According to the experimental studies regarding thermal comfort and radiant systems, the key findings are concluded to help guide the evaluation of thermal comfort for radiant systems.