Material Discovery in the age of AI

As semiconductor manufacturing advances to smaller process nodes, atomistic simulation has become essential for understanding and predicting material behavior at the scale where physics actually governs device performance. Surface deposition processes such as CVD and ALD depend on precise atomic-level understanding of how precursor molecules interact with substrate surfaces, reaction pathways, and film growth dynamics. Diffusion of dopants and defects through bulk materials and across interfaces determines electrical performance and reliability. Conductivity at heterogeneous interfaces between dielectrics, metals, and semiconductors drives critical design decisions that cannot be resolved with continuum models alone. At each of these scales and use cases, atomistic simulation is becoming a practical necessity, not a research luxury. 

However, conventional quantum chemistry and DFT methods remain highly compute-bound, requiring days of run time per molecule or configuration and creating bottlenecks that slow design cycles across process nodes and material systems. The pace of materials innovation in semiconductor manufacturing is increasingly constrained by the time and cost of traditional experimental and simulation workflows.

Machine learning interatomic potentials (MLIPs) offer a path forward: by learning from ab initio data, they can deliver near quantum chemistry accuracy at a fraction of the computational cost. Realizing that potential at scale, however, requires more than faster models. Surrounding operations such as neighbor list construction, dispersion corrections, and long-range electrostatics have historically remained CPU-bound and fragmented across tools, limiting end-to-end throughput regardless of model speed. GPU acceleration is changing the economics across the full simulation stack by delivering transformative speedups for high-fidelity DFT and quantum chemistry methods, and enabling MLIP-based workflows to run at speeds orders of magnitude beyond conventional ab initio approaches.

NVIDIA ALCHEMI is an open platform purpose-built for this challenge. The ALCHEMI Toolkit provides user-defined, composable simulation workflows powered by batched, GPU-accelerated operations, enabling teams to move from molecular inputs to actionable physical properties without sacrificing flexibility or reproducibility. Toolkit-Ops provides the modular, PyTorch-native building blocks for GPU-accelerated atomistic simulation. Partners including TSMC, Matlantis, Orbital Industries, and TorchSim are already building on this shared infrastructure to accelerate their simulation pipelines.

This session draws on real-world examples from the semiconductor industry to show how teams are compressing simulation cycles from months to days and translating those gains into faster, more robust process decisions. Attendees will leave with a practical framework for where AI-accelerated atomistic simulation fits in their R&D pipelines and how to get started.

Key Technologies Covered

• NVIDIA ALCHEMI
• Machine Learning Interatomic Potentials (MLIPs)
• GPU-accelerated molecular dynamics
• High-throughput screening and materials property prediction
• Multi-scale simulation: bridging DFT, MLIP-MD, and continuum methods