Search for: All records

Throughout its lifecycle, an LLM incurs significantly higher carbon emissions during inference than training. Inference requests vary in batch size, prompt length, and token generation, while cloud providers deploy heterogeneous GPU configurations to meet diverse service-level objectives. Unlike training, inference exhibits lower and highly variable hardware utilization, making equation-based carbon models unreliable. Existing network-based estimators lack accuracy, as they fail to account for the distinct prefill and decode phases, hardware-specific features, and realistic request distributions. We propose LLMCO₂, a graph neural network (GNN)-based model, to improve the accuracy of LLM inference carbon footprint estimation by ~ 67% over prior approaches. Source code is available at https://github.com/fuzhenxiao/LLMCO₂. 

Voids—the nothingness—broadly exist within nanomaterials and impact properties ranging from catalysis to mechanical response. However, understanding nanovoids is challenging due to lack of imaging methods with the needed penetration depth and spatial resolution. Here, we integrate electron tomography, morphometry, graph theory and coarse-grained molecular dynamics simulation to study the formation of interconnected nanovoids in polymer films and their impacts on permeance and nanomechanical behaviour. Using polyamide membranes for molecular separation as a representative system, three-dimensional electron tomography at nanometre resolution reveals nanovoid formation from coalescence of oligomers, supported by coarse-grained molecular dynamics simulations. Void analysis provides otherwise inaccessible inputs for accurate fittings of methanol permeance for polyamide membranes. Three-dimensional structural graphs accounting for the tortuous nanovoids within, measure higher apparent moduli with polyamide membranes of higher graph rigidity. Our study elucidates the significance of nanovoids beyond the nothingness, impacting the synthesis‒morphology‒function relationships of complex nanomaterials. 

Cation‐disordered rock salts (DRXs) are well known for their potential to realize the goal of achieving scalable Ni‐ and Co‐free high‐energy‐density Li‐ion batteries. Unlike in most cathode materials, the disordered cation distribution may lead to more factors that control the electrochemistry of DRXs. An important variable that is not emphasized by research community is regarding whether a DRX exists in a more thermodynamically stable form or a more metastable form. Moreover, within the scope of metastable DRXs, over‐stoichiometric DRXs, which allow relaxation of the site balance constraint of a rock salt structure, are particularly underexplored. In this work, these findings are reported in locating a generally applicable approach to “metastabilize” thermodynamically stable Mn‐based DRXs to metastable ones by introducing Li over‐stoichiometry. The over‐stoichiometric metastabilization greatly stimulates more redox activities, enables better reversibility of Li deintercalation/intercalation, and changes the energy storage mechanism. The metastabilized DRXs can be transformed back to the thermodynamically stable form, which also reverts the electrochemical properties, further contrasting the two categories of DRXs. This work enriches the structural and compositional space of DRX families and adds new pathways for rationally tuning the properties of DRX cathodes.