Token Cartography: principled applications of machine learning techniques to retain scientific interpretability

Unsupervised machine learning, such as clustering and latent reduction, is commonly applied to large-scale scientific data. These methods serve as the bedrock for more complex scientific foundation models. Lacking introspective approaches to access the trained latent spaces, scientific insights that might have been learned cannot be revealed. Furthermore, outlier and extrapolation detection is often not baked into the model, leaving users with outputs that further require confidence assessment. Together, these shortcomings render many existing foundation models unexplainable, significantly limiting scientific elucidation. To address this drawback, we propose a strategy to construct interpretable latent spaces by inducing semantic directions and mapping the space with known chosen scientific descriptors, alongside data regularisation via vector quantisation and feature engineering. This talk will elucidate the strategy, termed Token Cartography, and show an early use case on a massive complex petabyte-scale diffraction imaging dataset, achieved with HPC resources. We believe that this strategy can serve as a conceptual guideline for procedures to maximise extractable information from experimental data but also lay the groundwork for developing explainable scientific foundation models.

About the Speaker

Ervin Chia is a Doctoral candidate of Physics at the National University of Singapore and is part of an international collaboration aimed at understanding anomalous low-temperature properties of the most vital molecule for life on Earth: water. He studies petabyte-scale data collected from fourth-generation light sources, using machine learning methods enabled by HPC resources, to gain insights on the microstructures that form in the brief moments as water crystallises into ice.

Navigating Chemical Space: Techniques for mapping chemistry for molecular identification and reaction path discovery.

Most modern materials discovery AI techniques typically target the prediction of a final product with desirable properties. There is a growing understanding that more information about the chemical processes that result in materials is required in order to ensure these materials are synthesisable, among other desirable information, such as prediction of observables so we can ensure the correct materials are being produced, and behaving as expected.

This presentation will present two bodies of work.

1. Mass Spectroscopy is a standard technique for the identification of molecules in a laboratory setting, particularly for molecular species previously known, and catalogued into libraries. Identification of unknown molecules requires more extensive work, either through use of further molecular analysis techniques (such as Nuclear Magnetic Resonance or Infrared spectroscopy, among others), or computationally predictive techniques that can generate a synthetic library data entry to match a molecule to before the molecule is ever synthesised. This latter approach is particularly attractive when molecules are expected to be dangerous. Our lab had developed a molecular fragmentation and characterisation workflow for the prediction of Mass Spectrum using Density Functional Theory methods, which will be discussed.

2. Reaction Route Maps are a technique of searching chemical space through potential energy surface (PES) minima, and transition states between these minima, these can be labelled as zeroth, and first order saddle points respectively. We have developed the Molecular high-Index Saddle Dynamics (MHiSD) method which seeks to map chemical space by searching up the PES landscape the 2nd, 3rd and higher order saddle-points, and then back down in a branching tree approach. This provides alternative routes across the PES, and additional information on the structure of the PES, ideal for a broader, and more detailed search of chemical space. This technique is general for any molecular modelling method that can generate gradients and hessian information, and early results will be discussed.

About the Speaker

Stephen Dale is an Assistant Professor in the Department of Materials Science and Engineering, and the Institute of Functional Intelligent Materials at the National University of Singapore. His research focuses on the development of density-functional theory (DFT) and its applications in the discovery and characterisation of new materials. Stephen obtained his PhD from the University of California, Merced in 2017 and BSc from the University of Western Australia in 2012. Before joining NUS, he was a Postdoctoral Research Fellow at Dalhousie University, the Australian National University, the University of Sydney, and Griffith University.

Digital Twin Development for Taiwan's tokamak (FIRST) research: Multi-Scale Plasma Simulation, AI Equilibrium Reconstruction, and 3D Visualization

This talk presents recent progress by our team in developing a digital twin for Formosa Integrated Research Spherical Tokamak (FIRST), Taiwan’s premier nuclear fusion tokamak project. Currently in Phase 1 of development, our research integrates simulation, state reconstruction, and high-fidelity visualisation. We have established a temporal baseline for tokamak plasma evolution and implemented PLaNet [1] to accelerate magnetic equilibrium reconstruction. To capture multi-scale phenomena, we investigate critical plasma instabilities using targeted numerical codes: JOREK [2] for magnetohydrodynamic (MHD) dynamics such as vertical displacement events (VDEs), and GTC [3] for micro-instabilities like ion temperature gradient (ITG) modes. Furthermore, to bridge theoretical simulation with immersive visual analytics, we have rendered 3D JOREK simulation outputs within an NVIDIA Omniverse environment. These milestones establish the computational infrastructure needed for Phase 2 development as we prepare for FIRST’s initial plasma discharge.

[1] Bonotto, Matteo, Domenico Abate, and Leonardo Pigatto. Fusion Engineering and Design 200 (2024): 114193.

[2] M. Hoelzl et al, 2021 Nucl. Fusion 61 065001

[3] Z. Lin, T. S. Hahm, W. W. Lee, W. M. Tang, and R. B. White, Science 281 (5384), 1835-1837 (1998).

About the Speaker

Dr Yi-Cheng Chuang is a researcher at NCHC specialising in magnetic control fusion. He holds a Ph.D. from William & Mary, where he studied neutral particle transport in tokamaks. Today, his work centres on advancing Phase 2 of digital twin development. When he’s not modelling plasma environments, he enjoys hiking local trails and playing board games.

From Physics to AI and back

In this talk, we explore how we can use pre-existing knowledge (e.g., physics) to improve AI systems, and how we can possibly extract some knowledge from AI systems. On the first topic, we present a novel physics-enhanced deep learning hybrid method, namely CondensNet, for resolving cloud physics in climate models. On the second topic, we present some results on the use of explainable AI for climate applications. We conclude with an overall perspective bridging the two topics.

About the Speaker

Dr Gianmarco Mengaldo is an Assistant Professor in the Department of Mechanical Engineering and the Department of Mathematics - by courtesy - at National University of Singapore. He is also a member of the Joint Advisory Group for the World Meteorological Organization (WMO), a United Nations (UN) agency. He received his BSc and MSc in Aerospace Engineering from Politecnico di Milano (Italy), and his PhD in Aeronautical Engineering from Imperial College London (United Kingdom). After his PhD he undertook various roles both in industry and academia, including at the European Centre for Medium-Range Weather Forecasts (ECMWF), and at the California Institute of Technology (Caltech). Dr Mengaldo adopts an interdisciplinary approach at the intersection of mathematical engineering, computational physics and AI to study complex systems that arise in various branches of applied science. His current research interests involve (i) integrating domain knowledge (e.g., physics) and AI; (ii) explainable AI, both theoretical and applied, among others. Dr Mengaldo’s main application areas include weather and climate, robotics, and finance.

Addressing Fusion Energy Challenges through Multiscale Simulation and AI

The development of practical fusion energy presents significant scientific and engineering challenges, from understanding complex plasma behaviour to managing interactions between high-temperature plasmas and material surfaces. Addressing these challenges requires modelling physical processes across vastly different spatial and temporal scales, while also extracting useful insights from data generated by fusion experiments.

This presentation highlights research at A*STAR’s Future Energy Acceleration and Translation (FEAT) Centre, where high-performance computing and artificial intelligence are being applied to advance fusion science and technology. Drawing on research in computational plasma physics and plasma–material interactions, we will illustrate how numerical simulations are being used to understand plasma behaviour and predict material response under fusion conditions. In parallel, machine learning is being applied to extract meaningful information from diagnostic data collected by fusion experiments worldwide. Together, these efforts demonstrate how HPC-enabled simulation and AI can help researchers tackle key scientific and engineering challenges on the path towards practical fusion energy.

About the Speaker

Dr Eddie Chua is Co-lead of Plasma Physics & Diagnostics and a Senior Scientist at A*STAR’s Future Energy Acceleration and Translation (FEAT) Centre. He obtained his Ph.D. in Astrophysics from Harvard University, where his research centred on large-scale cosmological simulations of galaxy formation. His current research focuses on computational plasma physics, including edge and sheath dynamics, plasma–material interactions, and impurity transport. He also leads research in artificial intelligence for fusion diagnostics, integrating physics-based modelling and machine learning to advance predictive capabilities for future fusion energy devices.

Revealing Earth’s Interior with TomoATT: HPC-Enabled Seismic Imaging and Discovery

Understanding subsurface structures is essential for exploring energy resources, assessing geohazards, and supporting underground development. These structures must be inferred from indirect seismic observations by solving computationally demanding inverse problems. This talk introduces TomoATT, an open-source software package for adjoint-state traveltime tomography, and explains how algorithmic advances and high-performance computing enable detailed three-dimensional seismic imaging. I will highlight our ongoing studies across the Japanese archipelago, California, and Singapore, illustrating how TomoATT supports investigations of Earth’s structure at regional and local scales.

About the Speaker

Tong Ping is an Associate Professor at Nanyang Technological University, Singapore. His research focuses on applied & computational mathematics, computational geophysics, and seismic imaging, with applications to energy exploration, geohazards, and subsurface characterisation. His team develops advanced mathematical methods and open-source software, including TomoATT and SurfATT, for imaging Earth’s interior using seismic observations. By combining adjoint-state methods, high-performance computing, and geophysical data, his work connects computational innovation with practical challenges in geothermal exploration and geotechnical engineering.

Large-scale combustion simulations of hydrogen and ammonia fuels

About the Speaker

HPC for Quantum Utility: From Digital Twins to Autonomous AI Scientists

Quantum computing is progressing rapidly, yet realising practical quantum utility remains challenging due to limited quantum resources and the complexity of developing reliable quantum applications. In this talk, I will discuss how high-performance computing (HPC) can accelerate the path toward quantum utility through two complementary directions. First, HPC enables us to build digital twins of quantum systems, using large-scale simulation and AI to learn predictive models that reduce the need for costly quantum experiments. Second, HPC provides the computational foundation for Omni, our autonomous AI framework for quantum error correction, enabling AI agents to design, simulate, verify, and improve quantum workflows at scale. These results indicate a new avenue about how HPC, AI, and quantum computing can work synergistically to make useful quantum computing more accessible and accelerate scientific discovery.

About the Speaker

Yuxuan Du is an Assistant Professor at Nanyang Technological University (NTU), jointly appointed to the College of Computing and Data Science and the School of Physical and Mathematical Sciences. His research focuses on autonomous quantum science, quantum error correction, and fundamental algorithms for quantum computing. His work has been published in top-tier journals and conferences across physics and computer science as first or corresponding author, including Nature Reviews Physics, Nature Communications, Physical Review Letters, Physical Review X Quantum, npj Quantum Information, IEEE Transactions on Information Theory, and IEEE Transactions on Pattern Analysis and Machine Intelligence. He has been listed as a Top 2% Scientist Worldwide in the field of AI and Physics, based on citation indicators published by Stanford University and Elsevier. His research has been featured by major media outlets such as MIT Technology Review and Phys.org. He has served as an Area Chair for NeurIPS and ICLR. He serves as an Editor of Neural Networks and Quantum Machine Intelligence, and has regularly held Area Chair roles for ICML, NeurIPS, and ICLR.

About the Speaker

Energy-efficient building blocks for modern data centers

About the Speaker

Designing Hybrid Quantum–Classical Systems Before Deployment: Insights from HybridTwin

Hybrid quantum–classical systems bring together HPC resources and quantum processors, but their performance depends on more than the capabilities of either component alone. Communication latency, resource contention, and scheduling decisions can all change how a hybrid workload behaves. How can we evaluate these interactions before committing to a system design?

In this talk, I will introduce HybridTwin, a multi-fidelity digital twin framework for architecture-aware exploration of hybrid quantum–classical systems. It connects early-stage system simulation with execution-level analysis and real-device evaluation. Using HybridTwin, we examine how CPU–QPU communication latency changes the relative performance of quantum, classical, and quantum simulation. We then show how, as the number of concurrent jobs grows, waiting time and scheduling increasingly shape end-to-end performance. The talk concludes with a comparison of job-level and task-level scheduling and discusses what these findings mean for designing scalable hybrid computing systems.

About the Speaker

Dr Tao Luo is Senior Scientist and Head of the High Performance Computing (HPC) Chapter at the Institute of Advanced Intelligence and Computing (IAIC), A*STAR, Singapore. His research focuses on high-performance computing systems, AI infrastructure, scientific computing, computer architecture, and quantum computing, with an emphasis on enabling advanced computing capabilities for scientific discovery and industrial innovation. He leads initiatives spanning HPC platforms, AI systems, and large-scale computing infrastructure, and works closely with academia, industry, and national stakeholders to strengthen Singapore’s computational research ecosystem.

Dr Luo has led and contributed to multiple national and industry-funded programmes in AI and HPC. His work has been published in leading venues including Nature Communications, Communications of the ACM, IEEE TPDS, IEEE TCAD, SC, ICLR, NeurIPS, ECCV, AAAI, etc. He was an Associate Editor of IEEE TNNLS and currently serves as an Associate Editor of IEEE TCAD and has held numerous leadership roles in the international research community, including Track Chair of HPCAsia 2026, Program Chair of NVMSA 2024, and General Chair of NVMSA 2025.

Efficient data representation

About the Speaker

About the Speaker

About the Speaker

ROCm 10: A deep dive

As ROCm celebrates the 10 year anniversary with the release of ROCm10, a peek into the history and the latest and greatest features offered by ROCm for production grade AI deployments.

About the Speaker

Paul is a solution lead at AMD. He has over a decade of experience in HPC/AI infrastructure including system design and implementation. He plays a crucial role in building AMD's AI/HPC ecosystem across ASEAN region.