Advancing Science through AI

Chemistry underpins many fields critical to modern society, yet the rational design of chemical systems with targeted properties remains a formidable challenge due to the immense size of chemical space. Typical examples include the development of selective reagents for metal separations and robust chelators for Targeted Alpha Therapy (TAT)—a promising cancer treatment in which radioactive atoms destroy cancer cells upon contact. The clinical effectiveness of TAT relies on chelator molecules that can securely bind and transport these isotopes in vivo, as imprecise delivery can lead to damage to healthy tissue.

In this talk, we present an agentic AI–driven discovery loop that tightly integrates large language model (LLM)–guided hypothesis generation, generative molecular design, and first-principles quantum simulations to accelerate molecular discovery. We demonstrate how runtime integration enabled by automated workflow orchestration can harness large-scale, high-throughput physics-based simulations on leadership-class supercomputers to provide adaptive feedback to AI-guided exploration of chemical space. This approach offers an efficient, general, and scalable framework for transforming chemical discovery across high-impact applications.

Ping Yang graduated from Michigan Tech with a PhD in Chemistry in 2005 and is the Deputy Director of the G.T. Seaborg Institute for Transactinium Science and a Staff Scientist in the “Physics and Chemistry of Materials” group of the Theoretical Division at Los Alamos National Laboratory (LANL). She has extensive experience in computational approaches to modeling electronic structure and reactivity of actinides, surface chemistry, and nanomaterials in solution environments. She has published more than 150 papers and given over 100 invited presentations. Her research focuses on understanding fundamental electronic structures, optical and magnetic properties, reactivity, and dynamical behaviors of chemistry and materials that are important for energy security. Yang has broad interests in both applications of current high-performance computing frameworks and the development of new computational methods for long-time-scale simulations of complex actinide systems. She enjoys working with students and postdocs (>50) and was awarded a ‘Distinguished Postdoctoral Mentor Award’ by LANL.

When an ultrashort laser pulse (with a duration shorter than 10^-12 s) interacts with a material, the intense electric field strongly drives bonding electrons, leading to irreversible processes such as material removal or functionalization. These processes span a wide range of spatial and temporal scales, from femtosecond–nanometer-scale light–matter interactions to macroscopic material modification, and are highly sensitive to processing parameters. As a result, optimization based on trial-and-error experiments is costly in terms of time and resources.

To address this challenge, a digital twin for laser-based material processing is highly desirable. However, first-principles or physics-based numerical simulations of the underlying multi-scale, highly excited dynamics remain an open challenge. Recent advances in AI and data-driven modeling offer complementary approaches to bridge these gaps, but their effective deployment requires large-scale high-quality experimental data, which necessitate automated laboratory platforms as well as tight collaboration with HPC calculations. In this workshop, I will discuss how laser-based material processes can be modeled using AI technologies, and how such models can be combined with HPC calculations and experiments toward realizing a practical digital twin.

Dr. Shuntaro Tani is a RIKEN ECL Research Team Leader at the Center for Advanced Photonics, RIKEN. He received his Ph.D. in Physics from Kyoto University in 2014. His primary research spans ultrafast spectroscopy, lab automation, and laser-based precision material processing. His current work integrates cutting-edge laser technologies with deep learning techniques to enable high-precision processing through fully automated experimental systems.

AI for Science has emerged as a leading-edge research frontier, aiming to accelerate progress in meteorology, energy, materials, and life sciences through artificial intelligence. To advance the development and real-world adoption of AI for Science, NVIDIA has introduced the open-source framework PhysicsNeMo, which builds digital-twin models by integrating physics-governing equations (PDEs), simulation data, and observational data. Whether for domain experts with limited AI expertise or engineers pursuing advanced development, PhysicsNeMo provides strong support for the rapidly growing demand for AI across industries. It has been widely applied to scenarios such as CFD, heat transfer, solid mechanics, electromagnetics, seismic waves, weather forecasting, and super-resolution.

This talk will highlight representative PhysicsNeMo use cases, including meteorological downscaling, solar irradiance forecasting for photovoltaic power generation, and state-of-the-art AI models and high-fidelity datasets for vehicle external aerodynamics simulation.

Zhixiang Dai is a Developer Relations Manager at NVIDIA, responsible for the design and research of GPU computing solutions for higher education and research. His work focuses on the application of GPUs in high-performance computing, deep learning, data science, and AI for Science. He specializes in distributed and parallel GPU acceleration, as well as the migration and performance optimization of applications using CUDA and OpenACC.

Recent advances in artificial intelligence (AI), with scalable computational infrastructures, have significantly accelerated the application of AI and machine learning in the life sciences. In domains such as healthcare and drug discovery, it is essential to integrate heterogeneous data from multiple modalities, including genomics, molecular structures, and clinical information.

This talk highlights AI applications in the life sciences enabled by large-scale data and scalable computation, with a focus on graph neural networks for network-structured data and multimodal AI frameworks. We further discuss the development of molecular foundation models based on large-scale pretraining.

Ryosuke Kojima received his M.S. degree in 2014 and Ph.D. in Engineering in 2017 from Tokyo Institute of Technology. He joined the Graduate School of Medicine, Kyoto University, in 2017 as a Specially Appointed Assistant Professor and was promoted to Lecturer in 2021. Since 2024, he has been an Associate Professor at the Center for Digital Transformation of Healthcare and the Division of Biomedical Data Science, Graduate School of Medicine, Kyoto University. He is also a Team Director of the Laboratory for Multimodal AI Framework at the RIKEN BDR. His research focuses on artificial intelligence and data science for biomedical and life science applications.