Scientific Programme

I will introduce the general problem of first principles force fields: creating surrogate models for quantum mechanics that yield the energy of a configuration of atoms in 3D space, as we would find them in materials or molecules. Over the last decade significant advances were made in the attainable accuracy, and today we can model materials and molecules with a per-atom energy accuracy of up to 1 part in 10,000 with a speedup of over a million or more compared to the explicit quantum mechanical calculation, enabling molecular dynamics on large length and time scales. The most surprising aspect of the best model is its extreme generalisation: fitted only on small periodic crystals, it shows stable trajectories on arbitrary chemical systems, from water to nanoparticles and proteins. I will show some of the technical details behind the success of our models: equivariant many-body graph polynomials with very few and weak nonlinearities. The relationship between the architectural elements and the extreme generalisation is still largely a mystery. The locality of the graph structure is key to its success, as well as high body order and message passing. The force fields get significantly better with more data, yet model size and complexity can remain largely the same. Integrating explicit long range electrostatics with such general "foundational" force fields is a challenge, as well as combining large datasets for materials and organic molecules, due to the incompatibility of leading DFT approximations.

In the physical sciences, ML has found great synergy with physics-based simulations. On the one hand, simulations can provide abundant, reproducible and scalable training data. Machine learning models can act as surrogates of expensive simulators, allowing improvements in speed and accessible length- and time-scales. In this talk, we describe recent work in the use of differentiable programming and deep learning models to fuse optimization and learning in molecular simulations, including the use of differentiable uncertainty to power active learning, alchemical ML potentials that capture atomic disorder in solids, learning collective variables or transport operators for accelerated simulations, and using discrete generative models and reinforcement learning for surface reconstruction. Lastly, we encourage a discussion around the need to relate (ML-accelerated) simulations to tangible impact in chemicals and materials.

Discovery and understanding of next-generation materials requires a challenging combination of the high accuracy of first-principles calculations with the ability to reach large size and time scales. We pursue a multi-tier development strategy in which machine learning algorithms are combined with exact physical symmetries and constraints to significantly accelerate computations of electronic structure and atomistic dynamics. First, current DFT approximations fall short of the required accuracy and efficiency for predictive calculations of defect properties, band gaps, stability and electrochemical potentials of materials for energy storage and conversion. To advance the capability of DFT we introduce non-local charge density descriptors that satisfy exact constraints and learn exchange-correlation functionals called CIDER. These models are orders of magnitude faster in self-consistent calculations for solids than hybrid functionals but similar in accuracy. On a different level, we introduced equivariant neural network interatomic potentials (examples include NequIP, Allegro, SevenNet, GNoME, MACE) that are transforming how MD simulations are used to describe and design complex and reactive systems. We developed machine learning models for generalized potential and coarse-grained free energy functions with arbitrary dependence on external fields and temperature. We apply and demonstrate these methods via first principles ML MD simulations of dynamics of phase transformations, heterogeneous reactions, ferroelectric transitions, nuclear quantum effects, and soft materials.

I will discuss the progress we have made towards building a new generation of AI-powered infrastructure for scientific research.

Certain novel schemes for computing that use photons (instead of electrons).

N/A

Computational materials science has seen tremendous progress since the early days of Density Functional Theories. Stable algorithms enabled high-throughput computing which in turn enabled machine-learned potentials (MLP). Though far from perfect at this point, MLPs hold tremendous promise for accelerating materials simulation and discovery. Such progress is not parallelled on the experimental side, making it the gating factor in materials development. In response we built A-lab, an autonomous facility for the closed-loop synthesis of inorganic materials from powder precursors. All synthesis and characterization actions in A-lab, including powder mixing and grinding, firing, characterization by XRD and SEM, and all sample transfers between them are fully automated, leading to a lab that can synthesize and structurally characterize compounds within 10–20 hours of initiation. The A-lab leverages ab-initio computations through an API with the Materials Project, historical data sets that are text-mined from the literature, machine learning for optimization of synthesis routes and interpretation of characterization data, and active learning to plan and interpret the outcomes of experiments performed using robotics. The automation of synthesis and analysis can be further integrated into scientific workflows similar to computational workflows.

Fueled by increased availability of materials data, machine learning is poised to revolutionize materials science by enabling accelerated discovery, design, and optimization of materials. As one of the first and most visible of materials data providers, the Materials Project (www.materialsproject.org) uses supercomputing and an industry-standard software infrastructure together with state-of-the-art quantum mechanical theory to compute the properties of all known inorganic materials and beyond. The data, currently covering over 160,000 materials and millions of properties, is offered for free to the community together with online analysis and design algorithms. Serving a rapidly expanding community of more than 600,000 registered users, the Materials Project delivers millions of data records daily through its API, fostering data-rich research across materials science. This wealth of data is inspiring the development of machine learning algorithms aimed at predicting material properties, characteristics, and synthesizability. However, we note that truly accelerating materials innovation also requires rapid synthesis, testing and feedback, seamlessly coupled to existing data-driven predictions and computations. The ability to devise data-driven methodologies to guide synthesis efforts is needed as well as rapid interrogation and recording of results – including ‘non-successful’ ones. This talk will outline the rise of data-driven materials design, predictive synthesis and showcase successes as well as comment on current pitfalls and future directions.

Materials discovery is fundamental to advance next-generation technologies as well as for sustainable and circular economy. Beyond computational screening, generative models are efficient at finding materials with desired properties. Here, we introduce a Materials Generative Design and Testing (Mat-GDT) framework. Mat-GDT integrates property-directed generative design with accelerated validation through high-throughput simulations and experiments. The central idea of the framework is constructed around a foundation AI model for inorganic materials, which enables Mat-GDT to solve key challenges in sustainable design of functional materials. We demonstrate that domain-specific implementations of Mat-GDT will solve the outstanding challenge of data driven inverse design of inorganic functional materials.

Our online lives are shaped by machine learning predictions, guiding purchase decisions and social connections. These models rely on "like buttons" that capture more intent and emotion than mere observation alone. In the physical world, urban analytics instead focuses primarily on measuring human experience. This presentation details an experiment in Singapore with over 100 participants using the open-source Project Cozie platform to collect human experiences with noise, heat, and activity. The resultant 10,000+ microsurveys demonstrate the value of a "like button" in the spatiotemporal city context, offering new insights for urban planners and nudging humans to exhibit positive behavior. These data were then used as the foundation for the Cool, Quiet City Competition on the Kaggle platform, which crowdsourced predictive models to develop the most effective AI-driven solutions.

The global incidence of dementia is on the rise, with more than 9.9 million new cases annually, equivalent to a new diagnosis every 3.2 seconds. This trend is set to escalate due to the aging population worldwide. In this talk, I will discuss groundbreaking research in the application of generative artificial intelligence (AI) for the early detection of dementia through conversations with the elderly. Our work focuses on identifying language markers that serve as early indicators of dementia and developing predictive models based on these markers, providing a cost-effective and widely accessible solution for mass screening. To enhance the precision of language markers, we have pioneered a cross-modality learning framework that aligns language markers with brain imaging through a contrastive loss technique and improves language-based predictions with auxiliary imaging variables generated from language. Finally, I will present our latest advancements in developing a large-language model-based conversational chatbot. This chatbot is designed for cognitively demanding interactions that are not only user-friendly but also hold the potential to act as a therapeutic tool for mitigating cognitive decline among the general population.

The design of novel materials has been a cornerstone of technological progress, driving transformative innovations such as the adoption of electric vehicles, the development of highly efficient solar cells, and the widespread use of superconductors in magnetic resonance imaging (MRI) systems. At Microsoft Research, we aim to advance foundational artificial intelligence (AI) capabilities to accelerate the materials discovery process and deliver real-world impact across diverse domains. We have created two state-of-the-art AI models to accelerate materials design. MatterGen is a generative AI model that improves the efficiency of exploring the vast chemical space. It directly generates novel material structures given a “prompt” specifying material properties for the target application. MatterSim is an AI emulator that predicts which of those proposed materials are viable and if they really have the properties we requested. It efficiently simulates a broad range of properties of many different materials, including metals, oxides, sulfides, halides, and their various states such as crystals, amorphous solids, and liquids. Working together, MatterGen and MatterSim function as a flywheel to accelerate the design of novel materials by thousands of times and deliver real-world impact in broad domains.

Quantum AI advantage, also known as quantum AI supremacy, represents an ambitious goal of showcasing the ability of a programmable quantum device to solve AI problems that exceed the capabilities of current classical computers. This pursuit encompasses substantial challenges for both experimentalists and theorists, shaping a vibrant research frontier that has garnered increasing attention across diverse communities. In this talk, I will provide a concise introduction to the field of quantum AI advantage, highlighting its significance and potential implications. I will delve into recent progress made in this burgeoning area, shedding light on the strides taken towards achieving quantum AI advantage, especially quantum learning advantage. Drawing from specific concrete examples, I will explore promising instances where quantum systems display the potential to surpass classical limitations in certain AI tasks. Furthermore, I will outline the notable challenges inherent in demonstrating quantum AI advantage with existing proposals and devices.

AI is often seen as a data analysis tool, but its real strength may lie in reading, synthesizing knowledge, and suggesting connections—like an average but somewhat reflective scientist with exceptional recall. This talk explores how AI aids discovery not by generating insights but by managing literature reviews, background research, and pattern recognition, allowing scientists to focus on critical thinking. Through some examples from my experience, I will show how AI helps navigate scientific information, synthesize ideas, and uncover overlooked connections.

Data scarcity is a major impediment for the development of effective AI models. The lack of unbiased ‘real world’ data is especially acute in Chemistry. There has been a significant effort in recent years to address this issue. In this presentation, I will share my personal journey of setting up the Centre for Rapid Online Analysis of Reactions to support AI-enabled research, and highlight some challenges in generating meaningful data to improve AI predictions.

The discovery of novel chemicals and materials will revolutionize various industries such as energy storage, environmental remediation, and manufacturing. Traditional discovery methods have been time-consuming and costly, often taking years or even decades to move from hypothesis to production. However, with the advent of AI, this paradigm is rapidly changing. AI technologies have the potential to revolutionize the way chemicals and materials are discovered and developed, making the process faster, more efficient, and more cost-effective. NVIDIA ALCHEMI (AI Lab for Chemistry and Materials Innovation) is at the forefront of this revolution, enabling AI to accelerate chemical and materials discovery by accelerating the components needed to deploy these AI methods in real-world workloads with maximum efficiency and usability.

The prevailing paradigms of artificial intelligence (AI), built on the restrictive assumption of independent and identically distributed (i.i.d.) data, are fundamentally limited in their capacity for genuine creativity and adaptation in a non-stationary world1. While powerful within their training distribution, these systems fail when faced with true novelty, a stark contrast to biological intelligence, which exhibits continuous, lifelong learning and an open-ended growth in complexity2. This paper presents a comprehensive framework for a new computational paradigm designed to bridge this gap. We propose a move away from executing pre-programmed logic on a passive substrate and towards harnessing the intrinsic, emergent evolution of active matter3. Our central thesis is that realizing open-ended computation requires a new foundation based on physical hypercomputation: the deliberate engineering of non-halting, unpredictably complex dynamics—akin to those defined as "uncomputable" by classical theory—into physical materials4. This work outlines the theoretical foundations, candidate materials, experimental validation program, and architectural principles for this transformative approach.

In recent years, machine learning (ML) and artificial intelligence (AI) methods have begun to play a more and more enabling role in the sciences and in industry. In particular, the advent of large and/or complex data corpora has given rise to new technological challenges and possibilities. In his talk, Müller will touch upon the topic of ML applications in the sciences, in particular in medicine, physics, chemistry. He will focus on techniques from explainable AI and their use for extracting information from machine learning models in order to further our understanding by explaining nonlinear ML models. Finally, Müller will briefly discuss perspectives and limitations.

NVIDIA, as a full-stack computing platform company, is at the forefront of accelerating scientific discovery by providing comprehensive solutions that span hardware, software, and cloud services. The increasing diversity and scale of scientific applications, and the introduction of AI into scientific applications and workflows, introduce significant complexity to scientific computing infrastructure. NVIDIA addresses these challenges through the development of microservices, reference architectures & workflows, and AI development frameworks. By abstracting away complexity, NVIDIA enables scientists to focus on research rather than computing infrastructure management. Streamlined deployment and optimized performance shorten the time from hypothesis to discovery. NVIDIA’s ongoing scientific software development exemplifies its commitment to accelerating scientific discovery. These solutions empower researchers to harness the full potential of AI and high-performance computing, driving faster and more impactful scientific breakthroughs.

Following the sequence and structure revolutions, predicting the dynamical mechanisms of proteins that implement biological function remains an outstanding scientific challenge. Several experimental techniques and molecular dynamics (MD) simulations can, in principle, determine conformational states, binding configurations and their probabilities, but suffer from low throughput. Here we develop a Biomolecular Emulator (BioEmu), a generative deep learning system that can generate thousands of statistically independent samples from the protein structure ensemble per hour on a single graphical processing unit. By leveraging novel training methods and vast data of protein structures, over 200 milliseconds of MD simulation, and experimental protein stabilities, BioEmu’s protein ensembles represent equilibrium in a range of challenging and practically relevant metrics. Qualitatively, BioEmu samples many functionally relevant conformational changes, ranging from formation of cryptic pockets, over unfolding of specific protein regions, to large-scale domain rearrangements. Quantitatively, BioEmu samples protein conformations with relative free energy errors around 1 kcal/mol, as validated against millisecond-timescale MD simulation and experimentally-measured protein stabilities. By simultaneously emulating structural ensembles and thermodynamic properties, BioEmu reveals mechanistic insights, such as the causes for fold destabilization of mutants, and can efficiently provide experimentally-testable hypotheses.

In this talk we will cover the application of AI to disease modeling, target discovery, indication prioritization, indication expansion, and small molecule drug design. We will explore key case studies and cover the current state of the industry, highlighting its limitations, bottlenecks, and opportunities for advancing drug discovery. We will also discuss the applications of generative AI to development of foundational models for chemistry and multi-species multi-omics life models for aging and fundamental biological research.

Modern computing needs are increasingly limited by the latency and energy costs of memory access. Emerging memory devices such as resistive random-access memory (RRAM) have shown potential to enable efficient computing architectures, as the data can be mapped as the conductance values of RRAM devices and computation can be directly performed in-memory. Specifically, by converting input activations into voltage pulses, vector-matrix multiplications (VMM) can be performed in analog domain, in place and in parallel, thus achieving high energy efficiency during operation. In this presentation, I will discuss how practical neural network models can be mapped onto realistic RRAM arrays in a modular design. System performance metrics including throughput and energy efficiency will be discussed. Challenges such as quantization effects, finite array size, and device non-idealities will be analyzed, and techniques such as fine-grained structured-pruning and tensor-train factoring are explored to address the memory capacity concerns. At the architecture level, effective compiler needs to be developed to map the network graph on to the tiled weight-stationary architecture, and examples of different generations of networks will be presented.

While quantum devices promise extraordinary capabilities, discerning genuine advantages from mere illusions remains a formidable challenge. In this endeavor, quantum theorists are like prophets, striving to foretell a future where quantum technologies will transform our world. Most people understand quantum advantage primarily as offering faster computation. This talk explores the vast world of quantum advantages that extends far beyond computation to include fundamental advantages in learning, sensing, cryptography, and memory storage. I will also demonstrate how some quantum advantages are inherently unpredictable using classical resources alone, suggesting a landscape far richer than currently envisioned. As quantum technologies proliferate, these unpredictable advantages may enable transformative applications beyond the reach of our classical imagination.

KronosAI has achieved state-of-the-art for silicon photonics. We are building foundation model-based physics solvers that are faster, more accurate and more generalizable than traditional numerical solvers. Yujie will discuss insights gained from KronosAI's training, and how our work is paving the way for sustainable, eco-friendly engineering by integrating inverse design and design space exploration into everyday practice.

Materials representation across multiple length scales is essential for enabling AI models to solve structure-to-property prediction problems in complex systems such as superalloys. Properties like Vickers hardness are primarily governed by microstructural features and formation energy correlates to atomic arrangements and, making it crucial to capture relevant information from different structural hierarchies. Accurate and interpretable representations across these scales allow machine learning approaches to accelerate materials discovery and design. At the microstructural level, three frameworks are used to represent image-based information: (1) statistical representations using 2-point spatial correlations to capture phase distribution patterns, (2) geometry-driven image processing techniques that extract morphological descriptors such as area, perimeter, and shape of precipitates, and (3) deep learning models like convolutional neural networks that automatically learn hierarchical features directly from raw SEM images. These image-derived features are combined with metadata such as composition and processing history to predict mechanical properties like Vickers hardness. At the atomic level, graph-based representations like the CLEAR (Chemistry and Local Environment Adaptive Representation) descriptor-based model represents crystal structures by combining elemental properties with interatomic distances through Voronoi-based neighbours. By applying pooling operations, these graph features are transformed into fixed-size vectors that enable predictive modelling of formation energy and phase stability.

N/A

N/A

Since the launch of ChatGPT in Nov 2022, AI has taken the world by storm. It is without doubt that AI has the potential to significantly improve how we teach and learn. However, it also said that managing faculty is akin to herding cats. As a university, how can we encourage faculty members to adopt AI effectively in their teaching? In this talk, we will describe our work at the AI Centre for Educational Technologies (AICET) and also our strategy for deploying AI for teaching at NUS. In particular, we are of the view that AI adoption is not a technology issue. Instead, it is a problem of change management. We will describe how we have since Jul 2024 set up an internal consulting team to help faculty develop new ways to teach with AI and technology. We will describe our Request for Proposals (RFP) process and some of the projects that the consultancy team has done and share some of the lessons that we have learnt.

"In October of 2024, Prof. Dean Ho and team launched DELTA, a first-in-kind human trial - with Prof. Dean Ho as the test subject. This N-of-1 protocol harnesses a combination of fasting, fitness, and food to optimise metabolic health, monitored using an array of digital health platforms. Built from an unprecedented dataset, this study will culminate in a digital twin of Dean to hyper-personalise his cardiometabolic health protocol. Outcomes from this trial will create data collection frameworks to power population-scale healthspan optimisation and design regimens that do not require sustained digital monitoring to impact even larger communities. Prof. Ho will share his trial experiences, actionable learnings, and broadly-applicable opportunities to re-imagine population health."

We have adopted methods to establish recommender systems useful for efficient discovery of currently unknown chemically relevant compositions (CRCs) of inorganic ionic compounds from vast candidates. Using information of compounds registered in inorganic crystal structure database (ICSD) and tensor-based recommender system, we systematically calculated recommendation scores for chemical compositions of up to 23 billion five-element ionic systems. The high success rate, despite not using any prior knowledge or first principles theoretical results, proves the powerfulness of current recommendation systems. Synthesis experiments were made at the chemical compositions with high recommendation scores. Novel oxides and nitrides were efficiently discovered accordingly. Recommender system was also used to predict successful processing conditions for new compounds based on our parallel experiment-dataset.

First-principles methods based on density functional theory (DFT) have become indispensable tools in physics, chemistry, materials science, etc., but are bottlenecked by the efficiency-accuracy dilemma. The integration of first-principles methods with deep learning offers a transformative opportunity to overcome these limitations. In this talk, I will explore the emerging interdisciplinary field of deep-learning DFT, which employs advanced deep learning techniques to address key limitations in DFT computations. Specifically, I will present our recent work on developing a deep neural network framework, DeepH, that learns the relationship between the DFT Hamiltonian and atomic structures [1-3]. Trained on DFT data for small structures, these neural network models can generalize to predict properties of unseen large material structures without invoking time-consuming DFT self-consistent field iterations, making efficient and accurate study of large-scale materials feasible. Combined with recent methodological developments, these innovations pave the way for deep-learning electronic structure calculations [4-12]. This paradigm shift promises to transform the landscape of first-principles computations, significantly accelerating future materials discovery and design.

This talk investigates three important questions in agent research and development: how to build effective agents, how to endow them with advanced agentic capabilities—such as memory, knowledge digitization, reasoning, higher order thinking skills and general problem solving skills, and finally how to automate the construction and development of such agentic systems. To avoid the talk to be overly abstract, we use concrete applications in cybersecurity space to demonstrate how to automate cybersecurity expertise across the software development lifecycle, including vulnerability detection, diagnosis, proof-of-concept generation, and automated repair. The adoptability of this research could be applied to many other domains like vibe coding, medical analysis, material science and eventually auto-research. Finally, we discuss an interdisciplinary path toward Artificial General Intelligence (AGI), integrating insights from neuroscience, psychology, social sciences, and computer science to develop AI systems that are intelligent, agentic, and aligned with human values.

Traditional Machine Learning techniques, such as ANN, are based on classical (Bayesian) probabilities and Boolean algebra of events. At Qognitive, we have developed Quantum Cognition Machine Learning (QCML) by extending ML over quantum probabilities and non-Boolean (quantum) algebras of events. This allows QCML to beat the curse of dimensionality and achieve superiority over traditional ML when the number of features is large relative to the amount of data. QCML mimics human cognition and has been implemented on digital computers. We discuss applications of QCML in finance and medicine.

Recent breakthroughs in AI-driven methods such as AlphaFold have reshaped the landscape of protein structure prediction, raising questions about the continued relevance of traditional physics-based simulations. To bridge the gap between these two paradigms, we developed D-I-TASSER, a hybrid approach that integrates deep learning-derived inter-residue potentials with iterative Monte Carlo fragment assembly simulations, enhanced by a novel domain-splitting strategy tailored for large, multi-domain proteins. Benchmark tests and CASP assessments demonstrate that D-I-TASSER significantly surpasses both AlphaFold2 and AlphaFold3 in atomic-level accuracy, particularly for complex multi-domain targets. Large-scale folding experiments reveal that D-I-TASSER achieves correct folding in 81% of domains and 73% of full-length proteins across the human proteome. These results not only bring physics-based folding simulations back into the mainstream of protein structure prediction, but also underscore the compelling power of integrating cutting-edge deep learning techniques with physics-based modeling for high-accuracy, genome-scale prediction of protein structure and function.

Many of the most relevant observables of matter depend explicitly on atomistic and electronic structure, rendering physics-based approaches to chemistry and materials necessary. Unfortunately, due to the combinatorial scaling of the number of chemicals and potential reaction settings, gaining a holistic and rigorous understanding through exhaustive quantum and statistical mechanics-based sampling is prohibitive --- even when using high-performance computers. Accounting for explicit and implicit dependencies and correlations, however, will not only deepen our fundamental understanding but also benefit exploration campaigns (computational and experimental). I will discuss recently gained insights from my lab elucidating such relationships thanks to alchemical perturbation density functional theory and supervised machine learning.

Large Language Models (LLMs) have transformed Natural Language Processing by achieving remarkable performance across diverse tasks. Yet, alongside their widespread deployment, growing concerns have emerged about their susceptibility to adversarial attacks that can manipulate outputs, reveal sensitive information, or compromise reliability. These vulnerabilities raise important questions about the robustness and safety of LLMs in real-world applications. In this talk, I will present our recent efforts to systematically investigate the adversarial weaknesses of current LLMs and highlight the broader implications of these threats for trustworthy AI. I will also discuss emerging defense strategies and share insights into how the community can move toward building more robust and secure language models.

In recent years there have been concurrent rapid rises in high resolution remote sensing forest data and deep learning methods to exploit these for large-scale forest monitoring. Such methods have shown to be very powerful in rapidly retrieving properties of interest of large areas of forest with little effort, at least compared to standard field techniques. However, many deep learning approaches have been developed using data collected at small spatio-temporal scales, in a single or small number of forest types, and without ground validation data. Further, this approach often involves intensive and time-consuming data collection and processing, but generates results restricted to specific ecosystems and sensor types. There is a lack of widespread acknowledgement of how the types and structures of data used affects performance and accuracy of analysis algorithms, or the need for high quality ground truthing, leading to unrealiable validation data and inappropriate application of tools for monitoring.

Recent years have seen dramatic advances in both computational prediction and experimental determination of protein structures. These structures hold great promise for the discovery of highly effective drugs with minimal side effects, but using structures to design such drugs remains challenging. I will describe recent progress toward this goal, with a focus on machine learning approaches.

Large Language Models (LLMs) are increasingly deployed in interactions where they are prompted to adopt personas. This paper investigates whether such personaa conditioning affects model safety under bullying, an adversarial manipulation that applies psychological pressures in order to force the victim to comply to the attacker. We introduce a simulation framework in which an attacker LLM engages a victim LLM using psychologically grounded bullying tactics, while the victim adopts personas aligned with the Big Five personality traits. Experiments using multiple open-source LLMs and a wide range of adversarial goals reveal that certain persona configurations -- such as weakened agreeableness or conscientiousness -- significantly increase victim's susceptibility to unsafe outputs. Bullying tactics involving emotional or sarcastic manipulation, such as gaslighting and ridicule, are particularly effective. These findings suggest that persona-driven interaction introduces a novel vector for safety risks in LLMs and highlight the need for persona-aware safety evaluation and alignment strategies.

AI is becoming a critical component of Earth system science workflows that are experiencing rapid growth in data from model output of increasing resolution in weather and climate models, and Earth observation systems that produce orders of magnitude more data than their previous generations. Efforts are underway in the weather and climate modeling community towards refining the horizontal resolution of atmosphere GCMs towards km-scale to explicitly resolve certain small-scale convective cloud processes and provide more realistic local information on climate change. At the same time, exascale HPC systems have arrived and in most cases are designed with GPU accelerator technology that offers opportunities in reasonable simulation turn-around times balanced with efficiency in energy consumption. Ultimately, output from global storm-resolving models at km-scale will become the essential driver behind the deployment of Earth digital twins for programs like the EC Destination Earth and NVIDIA Earth-2. For model emulation of the Earth system, AI models become increasingly accurate as they train on more data, yet computational and storage requirements in data-distributed computing environments with energy-efficiency considerations are the current challenges for the HPC vendor community. This talk will describe advances in HPC for GPU-accelerated numerical models, AI software and system features for large-scale data handling, and ML model training and inference that when combined, provide the critical components towards the vision of Earth system digital twins.

The tropical urban nature of the weather and climate in the Singapore region places particular requirements on future observation, models and IT infrastructure. Significant progress has been made in recent years to meet these requirements through added value km-scale Numerical Weather Prediction (NWP) and regional climate projections based largely on physical climate modelling. However, recent advances in AI4Weather/Climate create exciting opportunities for further added value. This talk will provide an overview of CCRS’ plans to move from the current generation physical climate/weather modelling system (based on the Unified Model - UM) to next-generation weather and climate modelling approaches employing a combination of physical, hybrid AI and fully-data driven models suitable for applications in the Singapore region.

I will discuss the emerging need to massively use and deploy Artificial Intelligence (AI) techniques in materials and devices design and engineering, by overviewing our journey into the research of innovative advanced materials using AI in the field of interconnects technologies. I will point towards the revolution of fully describing as-grown materials with AI-driven reconstruction of digital models/twins with atomistic modelling, hence bringing the accuracy of first principles methods to unprecedented multimillions atoms models. This opens a tremendous capability to correlate in-depth atomistic-scale features of as-grown materials with resulting local and global chemical, physical and devices properties, offering reverse engineering and AI-driven co-piloting for industrial innovation strategies.

Particle physics has a rich history of developing advanced simulators, ranging from modelling proton collisions to creating virtual detectors. We are now integrating Artificial Intelligence through "Neural Simulation Based Inference," which enables the analysis of complex phenomena that cannot be computed analytically from first principles.

Machine learning (ML) is opening new possibilities for designing advanced control systems in chemical processes. While model predictive control (MPC) is the gold standard for such applications, it traditionally relies on linear empirical models, which may not capture the inherent nonlinearity of chemical systems. Nonlinear first-principles models offer better accuracy but are often too complex for practical implementation. ML tools, such as neural networks, provide a data-driven way to model nonlinear dynamics efficiently and improve MPC performance. In this talk, we will share our recent work on integrating ML with MPC. We will outline: (a) a general framework for using neural networks in nonlinear system modeling for MPC; (b) theoretical results on RNN generalization and closed-loop stability based on statistical learning theory; and (c) new approaches to tackle challenges like data scarcity, model uncertainty, and high dimensionality. Applications in chemical and pharmaceutical processes will be presented to illustrate the impact of our methods on next-generation manufacturing.

At Gero, we’ve spent years training physics-based AI models on some of the largest longitudinal human datasets, measured over years. This lets us see not just how single diseases work, but how they interact, how aging accelerates them, and how some diseases, in turn, speed up aging. Our models suggest that in long-lived mammals like humans and dogs, aging itself is mostly an entropic process — driven by accumulated damage that’s hard to reverse. But we’ve also shown that many diseases share common biological failures, and those failures are what we can target. If you think this sounds like science fiction, think about what GLP-1 drugs did recently. One drug class now treats diabetes, obesity, heart failure, and more. This is the first proof that the “super-drug” concept works. And we believe this is just the beginning. Aging itself may be the next frontier.

We discuss some recent work on constructing stable and interpretable macroscopic dynamics from trajectory data using deep learning. We adopt a modelling approach: instead of generic neural networks as functional approximators, we use a model-based ansatz for the dynamics following a suitable generalisation of the classical Onsager principle for non-equilibrium systems. This allows the construction of macroscopic dynamics that are physically motivated and can be readily used for subsequent analysis and control. We discuss applications in the analysis of polymer stretching in elongational flow. Moreover, we will also discuss some algorithmic challenges associated with learning (macroscopic) dynamics for scientific applications.

The transition from laboratory discovery to commercial manufacturing remains a critical challenge in photovoltaic technology development. This talk presents our comprehensive vision for self-driving laboratories (SDLs) that not only accelerate development but fundamentally reshape how we approach photovoltaic innovation. We begin by demonstrating our recent achievements in roll-to-roll manufacturing of perovskite solar cells, where our SDL platform enabled the fabrication of modules with 11% efficiency under ambient conditions, processing over 11,800 devices in 24 hours. This unprecedented throughput, combined with our cost projections of ~0.7 USD/W, establishes a new benchmark for scalable photovoltaic manufacturing. However, the true potential of SDLs extends far beyond automation. We will discuss our perspective on the evolution of SDL paradigms - from current focused optimization approaches to future systems capable of autonomous hypothesis generation and testing. This includes our ongoing work towards hybrid perovskite solar cells (HPSCs) through cross-domain optimization, where SDLs simultaneously consider materials properties, device architectures, and manufacturing constraints. Our latest innovations in integrating biomass-derived materials and developing novel device structures demonstrate how SDLs can unlock previously unexplored pathways in photovoltaic development. We propose a roadmap for future SDL development that emphasizes the democratization of these technologies, integration of manufacturing concepts from the outset, and the potential for autonomous agents to accelerate discovery.

Transition state (TS) search is key in chemistry for elucidating reaction mechanisms and exploring reaction networks. The search for accurate 3D TS structures, however, requires numerous computationally intensive quantum chemistry calculations due to the complexity of potential energy surfaces. We developed an object-aware SE(3) equivariant diffusion model (OA-ReactDiff) that satisfies all physical symmetries and constraints for generating sets of structures – reactant, TS, and product – in an elementary reaction. By learning the joint distribution of reactant, TS, and product, OA-ReactDiff can sample novel reactions that go beyond empirical expectation of chemists. With an initial guess adopting prior knowledge in chemical reactions and optimal transport framework, we developed React-OT, a model tailored for TS generation. Provided reactant and product, React-OT generates a TS structure in sub-seconds instead of hours, which is typically required when performing quantum chemistry-based optimizations. The generated TS structures achieve a median of 0.05 Å root mean square deviation compared to the true TS. We envision the proposed approach useful in constructing large reaction networks with unknown mechanisms.

The breakthrough of crystal structure prediction has helped to solve formidable problems of compound prediction and prediction of stable molecules/clusters. The vast body of new unusual compounds and phenomena were predicted using these methods and then confirmed experimentally. All this required an explanation, and this has stimulated the development of new concepts. I will discuss several cases: 1. Discovery of anomalous compounds under pressure, such as Na3Cl, NaCl7 and highest-temperature superconductors known to date – H3S, YH6, CaH6, ThH10, LaH10. 2. Discovery of counterintuitive phenomena at high pressure – formation of transparent insulating phase of sodium and chemical reactivity of helium. 3. Rationalization of these and other phenomena based on newly developed scales of electronegativity and chemical hardness. 4. Prediction of stable molecules – the formalism and its applications. In particular, I shall discuss the results on molecules and crystalline allotropes of sulfur, phosphorus and boron. Chemical diversity of hydrocarbons will be explained, as well as unusual molecules in the C-H-N-O system.

TBC

The trajectory of scientific research worldwide is guided by long-term goals, spanning the spectrum from curiosity and fundamental discoveries in physics to the applied challenges of enhancing materials and devices for a wide array of applications. However, the execution and assessment of daily research efforts typically hinges on multiobjective reward functions, which can be evaluated either during or at the conclusion of an experimental campaign. Although this concept is tacitly acknowledged within the scientific community, the implementation of autonomous experimental workflows in automated laboratories necessitates the formulation of robust reward functions and their seamless integration across various domains. Should these reward functions be universally established, the entirety of experimental efforts could be conceptualized as optimization and decision making problems. Here, I will present our latest advancements in the development of autonomous research systems based on electron and scanning probe microscopy, as well as for automated materials synthesis based on reward driven workflows and reward integration across domains. We identify several categories of reward functions that are discernible during the experimental process, including imaging optimization, fundamental physical discoveries, the elucidation of correlative structure-property relationships, and the optimization of microstructures. The operationalization of these rewards function on autonomous microscopes is demonstrated, as well as strategies for human in the loop intervention. Utilizing these classifications, we construct a framework that facilitates the integration of multiple optimization workflows, demonstrated through the synchronous orchestration of diverse characterization tools across a shared chemical space, and the concurrent navigation of costly experiments and models that adjust for epistemic uncertainties between them.

Rapid accumulation of experimental data, coupled with the slow speed of its analysis, presents a significant challenge in modern chemistry and materials science. A substantial portion of experimental results remains underutilized, categorized as "sleeping" or "lost" data, which hinders innovation and slows down progress. In several data-intensive research projects in chemistry, only 10% of the data is analyzed manually, while around 90% of the data remains not analyzed. Artificial intelligence (AI) offers an efficient solution by enabling the systematic retrieval, organization, and analysis of such data, leading to improved decision-making and accelerated discovery.

The first milestone of AI for Science is to create a comprehensive data platform. However, the development of foundational database in polymer material science has significantly lagged behind. To overcome this barrier, we have constructed a massive computational database encompassing a vast chemical space of polymers, using RadonPy, a fully automated pipeline tool for all-atom classical molecular dynamics and first-principles-based computational experiments. The foundation models pre-trained on this computational database have demonstrated scalable Sim2Real transfer in a wide range of real-world predictive tasks through fine-tuning. This presentation will discuss several aspects of scaling laws for Sim2Real transfer learning and their practical applications for the discovery of new polymers.

The implementation of large-scale universal quantum computation represents a challenging and ambitious task on the road to quantum processing of information. In recent years, an intermediate approach has been pursued to demonstrate quantum computational advantage via non-universal computational models. A relevant example for photonic platforms has been provided by the Boson Sampling paradigm which is known to be computationally hard while requiring at the same time only the manipulation of the quantum photonic resources via standard optics and single-photon detection. Beside quantum computational advantage demonstrations, a promising direction towards possibly useful applications can be found in the field of quantum machine learning, considering the currently unexplored intermediate scenario between non-adaptive linear optics and universal photonic quantum computation. We discuss the implementation of quantum machine learning by adding "adaptivity" to a Boson Sampling platform, based on universal programmable integrated photonic circuits. We will discuss how Adaptive Boson Sampling is a viable route towards dimension-enhanced quantum machine learning with optical devices.

The explosive growth of AI reasoning capabilities brings ambitious goals such as solving unproven mathematical conjectures or generating complex verified software into the realm of possibility. But when such synthesized proofs may span hundreds of pages and thus stretch or transcend human comprehension, who will verify and establish trust in them? The Lean theorem prover is an open-source tool for authoring and checking machine-readable proofs of arbitrary logical statements, reducing the trust bottleneck to the correctness of the tool and the statement -- a reduction by orders of magnitudes. In this talk, I will describe the fundamentals of this approach and how AI is benefiting from Lean and vice versa already today.

Since the publication of the classic paper by Chen et al (2018 NeurIPS), there has been an explosion of applications of Neural network and operator methods for solving ODEs (NODEs) and more recently PDEs. In particular, neural operator methods are becoming the preferred architectures for application to the development of neural Atmospheric General Circulation Models for weather and climate prediction. In this presentation I will discuss the close relationship between neural methods, dynamical systems, and Bayesian inference in application to diagnosing the dynamics and causal relationships of observed and simulated climate teleconnections. More specifically, I will discuss the so-called "loss of hyperbolicity" linked with the alignment of dynamical vectors characterizing the emergent dynamics of persistent synoptic weather events due to flow instabilities. I then show how dynamical systems approaches to identifying the hyperbolic splitting of the tangent space defined by the leading physical modes, closely mirrors mapping methods employed in NODEs. Lastly, I show how further generalization to normalizing flows connects to Bayesian structure learning methods in application to inferring Granger causal relationships from climate data.

In this talk, I am going to present the status of memristor-based neuromorphic computation. I will present the state-the-art performance and the current challenges to overcome in the next five years. In this talk I will discuss two different strategies on how to overcome such problems. The first one consists of integrating two-dimensional materials at the back-end-of-line of silicon microchips, and we will present the performance of on-chip memristors made of multilayer hexagonal boron nitride. The second strategy will consist of unconventional biasing of established fully CMOS-compatible electronic devices to produce outstanding neuromorphic responses. I will also discuss the main technical problems to be faced in the next years when following each strategy, and I will provide some recommendations on how to solve them.

Formalizing mathematics into machine-checkable logic is essential for advancing scientific rigor and enabling powerful AI reasoning. However, the process of translating informal mathematical text into formal languages remains a major bottleneck. This talk explores the challenge of autoformalization—the automated conversion of natural mathematical language into formal logic—through the lens of Euclidean geometry, one of the oldest and most foundational domains in mathematics. I will present insights from our recent work on LeanEuclid and PyEuclid, which demonstrate how modern Large Language Models (LLMs), combined with formal methods, can help bridge the gap between informal and formal mathematical reasoning.

Aurora is a foundation model for the Earth system pretrained on a large and diverse collection of geophysical data. The key ability of Aurora is that the model can be fine-tuned to produce forecasts for a wide variety of environmental forecasting applications, often matching or even outperforming state-of-the-art traditional approaches at a fraction of the computational cost. In this second part of a two-part on Aurora, I will explore the model architecture in more detail. I will also discuss pretraining, the scaling behaviour of the model, and how this enables high-performance weather forecasting, with a particular focus on tropical cyclone track forecasting.

Recent advances in generative models and reinforcement learning have enabled automatic theorem proving beyond the capabilities of classical methods. Scaling up data, model size, and inference compute has undeniably allowed us to prove more theorems. But is scaling alone enough to solve mathematical theorem proving? If not, what essential components might still be missing?