Invited Talk at CCP2026, the IUPAP Conference on Computational Physics

Application scientist Yuta Aoki, Ph.D., will be giving an invited talk at the international conference "XXXVII IUPAP Conference on Computational Physics (CCP2026)," which will be held in Seoul, South Korea from August 9th (Sun) to 14th (Fri), 2026.
The Computational Physics Conference (CCP) is an international conference in the field of computational physics held under the auspices of IUPAP (International Union of Pure and Applied Physics). This 37th conference is hosted by the Korean Physical Society and brings together researchers from a wide range of fields in computational physics, including materials science, quantum many-body physics, lattice field theory, and soft matter/biophysics.
This presentation will be given in session "T12. Artificial Intelligence and Data-Driven Computational Physics," which focuses on AI and data-driven computational physics. We will present our efforts to accelerate and scale up atomic-level simulations while maintaining first-principles calculation level accuracy using the universal machine learning interatomid potential "Matlantis PFP."

Presentation

Title: Accelerating and Scaling Up Ab-Initio-Level Atomistic Materials Simulations via Universal Machine Learning Interatomic Potentials

Abstract: 
Density functional theory (DFT) has long been the foundation of computational materials science, but its cost rises steeply with system size, placing large models and long-timescale dynamics out of reach. Universal machine learning interatomic potentials (MLIPs) are changing that. This talk reports recent progress in carrying first-principles-level accuracy into large-scale, long-timescale atomistic simulation with Matlantis PFP, a universal MLIP trained on tens of millions of first-principles calculations and built on a graph neural network that embeds physical requirements such as three-dimensional symmetry and consistency between energy and forces.

Applications range from high-throughput screening to dynamics in complex materials systems. Crystal structure exploration shows how reliable accuracy at scale makes it possible to survey candidate structures ahead of experiment, while ionic diffusion in solid electrolytes, molecular adsorption and surface reactions on inorganic materials, interfaces in multi-component alloys and complex amorphous structures cover systems for which conventional first-principles calculation is impractical and classical force fields lack transferability. Knowledge distillation from PFP into lighter models extends the reach of simulation further in both space and time, and PFP's internal representations serve as descriptors for materials informatics, predicting even properties that conventional MLIPs cannot evaluate directly, such as the dielectric constant.

Together these results point toward a simulation-first approach to industrial R&D, grounded in high-throughput screening with reliable accuracy. The talk closes with directions for AI-agent-driven automation of atomistic modeling and large-scale simulation workflows.

Presenter

Event Overview

Conference Name: XXXVII IUPAP Conference on Computational Physics (CCP2026)
Date: August 9th (Sunday) – August 14th (Friday), 2026
Venue: Sejong University Convention Center, Seoul, South Korea
Session: T12. Artificial Intelligence and Data-Driven Computational Physics
Talk Time: Thursday, August 13, 2026, 16:00-16:30 (local time)

Event Details
For more details about the event, please check the link below.
https://ccp2026.kr/