Bridging Accuracy and Scale: AI-Accelerated Atomistic Simulation with Matlantis™

Taku Watanabe

Principal researcher

Matlantis

Accelerating materials discovery requires bridging the gap between accurate simulations and industrially relevant length and time scales. Machine-learning interatomic potentials (MLIPs) are emerging as a powerful tool to achieve this, as they promise to deliver the accuracy of first-principles calculations at a fraction of the computational cost. However, their practical application is often hindered by the need for costly, system-specific fine-tuning and limited predictive accuracy across diverse chemical spaces.

In this talk, we introduce the Matlantis platform, which operationalizes AI-accelerated atomistic simulation in the cloud. Our core MLIP, PreFerred Potential (PFP), is trained across 96 elements and a unique, diverse dataset of materials, enabling out-of-the-box predictive power across a vast chemical and structural space. This universal potential supports large-scale molecular dynamics simulations and property estimation across diverse systems, including catalysis, polymers, semiconductors, and energy materials.

We highlight its predictive power through case studies in catalyst optimization and reaction engineering, ion diffusion in solid electrolytes, and surface reactions for semiconductor processing. We then show integration patterns with existing workflows and validation against reference DFT calculations and targeted experiments. Finally, we introduce LightPFP, which, for fixed domains, distills PFP to lower inference cost and extend tractable system sizes into the 10^5-atom class, broadening design-space exploration.

By overcoming key bottlenecks in accuracy and scale, Matlantis makes high-fidelity simulation a practical and routine tool for accelerating industrial materials R&D.