Month: August 2025

For decades, researchers have relied on carbonaceous materials like graphite and carbon black as workhorse components in lithium-ion batteries (LIBs). These carbons usually function as hosts for lithium intercalation, where Li ions slip between graphene layers. But there’s always been a catch: lithium diffusion through bulk carbon is relatively slow, limiting performance and requiring the help of liquid or solid electrolytes. In this work, we collaborated with the group of Prof Ping Liu and researchers at the University of Maryland, and the University of Houston to identify a completely new superionic lithium transport pathway along the surface of carbon materials. We found that in carbon blacks with high surface area but limited lithium storage capacity, such as Ketjen black (KB), lithium ions move astonishingly fast across the surface once lithiated. At room temperature, lithiated KB exhibited an ionic conductivity of 18.1 mS/cm, outperforming many leading solid electrolytes. Our group’s contributions (led by Ji Qi) is in the form of DFT calculations to show that lithium at the surface encounters much lower migration barriers (as low as 0.15 eV) than lithium inside bulk graphite. In effect, the carbon surface becomes a highway for ultrafast lithium transport. This surface-mediated conduction has several […]

Graph deep learning is transforming materials science by enabling accurate, scalable, and efficient predictions of material properties and potential energy surfaces (PES). Our article on “Materials Graph Library (MatGL) — an open-source, modular framework purpose-built for materials science and chemistry” has been published in npj Computational Materials. MatGL started as a collaboration between Intel Labs and the Materials Virtual Lab to provide a “batteries-included” environment for: Implementing both invariant and equivariant graph neural networks (GNNs), including M3GNet, MEGNet, CHGNet, TensorNet, and SO3Net. Leveraging pretrained foundation potentials (FPs) covering the full periodic table for out-of-the-box property predictions and atomistic simulations. Seamless integration with ASE and LAMMPS for high-fidelity molecular dynamics, geometry optimization, and property calculations. Training custom models with PyTorch Lightning for efficient parallelization on CPUs, GPUs, and TPUs. We show that MatGL’s models achieving state-of-the-art accuracy on widely used datasets (QM9, Matbench, ANI-1x, MPF, MatPES) while maintaining competitive computational efficiency. The library’s design also supports fine-tuning, enabling rapid adaptation to new materials systems. Read the full paper here and explore MatGL on GitHub.