SandboxAQ was named Software and Digital Innovation of the Year at The Battery Show for the second year running. This webinar explores the technology behind that recognition.
Machine-learned interatomic potentials (MLIPs) promise near-quantum accuracy at a fraction of the cost. However, general-purpose models are limited by their training data, which mostly comes from standard DFT and near-equilibrium structures. That makes them unreliable for the reactive, high-temperature conditions that decide battery safety and performance.
In this webinar, we will demonstrate how SandboxAQ's Large Quantitative Models close this gap by generating physics-based data where public data runs out. Attendees will learn how strategic data generation, foundation potentials and accelerated computing together can shorten battery materials development from years to weeks.
Smarter Data Generation - We developed AQVolt26, a dataset of 323,000 off-equilibrium lithium-halide structures labeled with r2SCAN DFT. Strategic sampling lets us capture a landscape of more than 200 million structures by sampling only about 2% of it.
Higher Accuracy & Stability - Models trained on AQVolt26 are more accurate and more stable in molecular dynamics. As we extend into lithium and sodium sulfides and hydrides, they cut energy and force errors by 54% and 37% compared with open-source models.
Accelerated Discovery with NVIDIA - Together with NVIDIA, we show how batched GPU simulation with the NVIDIA ALCHEMI Toolkit speeds up inference by 100x. That speed enables active-learning and generative workflows that have identified 363 promising solid-electrolyte candidates and led to a new halide electrolyte patent.
