AI4X-AC 2026 Conference Schedule

The rapid expansion of the mRNA therapeutic landscape necessitates discovery platforms that can overcome the limitations of sparse historical data and traditional trial-and-error methodologies. This talk introduces LUMI-lab, a foundation model-driven self-driving laboratory that integrates a transformer-based 3D molecular model with robotic automation to navigate vast chemical spaces. By leveraging unsupervised pretraining on over 28 million structures, the platform enables data-efficient, few-shot learning to identify high-performing ionizable lipids for lung-specific mRNA delivery. LUMI-lab autonomously synthesized and screened over 1,700 lipid nanoparticles, leading to the discovery of halogenated lipid tails as a novel structural feature that enhances endosomal escape. These efforts culminated in the identification of LUMI-6, which achieved 20.3% gene editing efficacy in lung epithelial cells in vivo. Ultimately, LUMI-lab establishes a scalable framework for the autonomous discovery of complex biomaterials, bridging the gap between artificial intelligence and translatable RNA medicine.

Polymers exhibit complex behavior across multiple length scales, but traditional single-molecule experiments are slow and low-throughput, limiting statistical insight. In this talk, I will present an automated experimental platform for high-throughput observation and manipulation of individual polymer chains under nonequilibrium conditions. The platform integrates microfluidics and real-time feedback control to trap, stretch and observe individual molecules with minimal human intervention. These experiments generate rich and large datasets that capture nonequilibrium polymer dynamics at the single-molecule level. Using these datasets, we apply physics-guided deep learning to extract reduced coordinates and reconstruct energy landscapes, linking microscopic fluctuations to macroscopic behavior. This combination of automation and data-driven modeling enables quantitative studies of nonequilibrium polymer dynamics and lays the foundation for autonomous experimental exploration of soft matter systems.

High-throughput $\ce{CO_2}$ electrocatalysis requires reproducible experimental workflows that connect hardware, data, and decision-making. However, current laboratory systems are fragmented across heterogeneous instruments, vendor-specific data formats, and manual analysis procedures. Here, we present a Physical Unified Device Architecture (PUDA)-enabled infrastructure for AI-assisted $\ce{CO_2}$ electrocatalysis research. PUDA functions as a loop-closure architecture that converts laboratory instruments and robotic platforms into command-line-accessible, Large Language Model (LLM)-invokable physical endpoints. Through a unified middleware layer, electrochemical workstations, chromatographic instruments, mass-flow controllers, robotic systems, and characterization tools can be connected to AI agents for controlled execution, data acquisition, and workflow orchestration. This is integrated with a centralized data-processing platform that automatically synchronizes raw experimental files, parses them into structured formats, and enables traceable analysis through natural-language instructions. The current implementation supports a full spectrum data processing techniques ranging from electrochemical workstation, gas chromatography, high-performance liquid chromatography, mass-flow controllers, nuclear magnetic resonance, X-ray photoelectron spectroscopy, ultraviolet-visible spectroscopy, X-ray diffraction, and X-ray absorption spectroscopy. In $\ce{CO_2}$ electrocatalysis workflows, the system combines electrochemical, gas-product, liquid-product, and flow-rate data to perform calibration, concentration calculation, product quantification, and Faradaic efficiency analysis directly from raw files. This PUDA-enabled framework provides an AI-native experimental backbone for closed-loop catalyst screening and $\ce{CO_2}$ electrocatalysis test bedding. By standardizing physical tool access and raw-data processing, it improves throughput, reproducibility, and data traceability, while laying the foundation for autonomous optimization of catalysts and electrolyzer systems.

Human knowledge occupies a small, smooth island in a vast ocean of impossible ideas. Most modern AI systems are trained to move efficiently across that island, optimizing within the boundaries of smooth and reward-dense patterns humans have already discovered. The next generation of AI will not remain on that island. As we ask machines to explore spaces beyond current human understanding, the assumptions underlying conventional optimisation begin to break down. Outside known domains, the landscape is discontinuous, sparse, and dominated by hard constraints rather than gradual gradients: unwary explorers risk dropping off the edge of the map. This talk introduces our work on Negative Space Learning (NSL), an alternative approach to discovery-oriented AI systems in reward-sparse regimes. We argue that superhuman AI requires open-ended learning in environments we cannot scaffold. We do not know what a solution to a 500 IQ problem looks like, and thus we cannot reward-shape our way there. Instead we must provide survival constraints and let the system find its own path. Instead of optimising directly toward a predefined notion of success, NSL systems iteratively carve away regions of failure through interaction with environments and constraint violations.

The growing demand for computational power needed to develop and operate advanced artificial intelligence models is consuming large amounts of energy. Motivated by the efficacy, efficiency, and robustness of natural intelligence in biological information processing systems, neuromorphic engineering offers viable options for extremely low-energy cognitive computing in highly parallel, distributed architectures. The emergence of nanoscale memristors, combined with neuromorphic hardware, has generated hope for building ultra-dense circuit architectures to perform in-memory-computing. Current hybrid CMOS-memristor technologies allow for the implementation of large-scale neuromorphic systems with dense networks of memristors in the form of crossbars, performing synaptic connections between neural layers. These systems are especially suitable for implementing online learning algorithms such as STDP (Spike-Timing-Dependent Plasticity).

The rapid growth of materials informatics has enabled computational models to propose vast numbers of candidate materials with targeted properties. However, property optimization alone is not sufficient for real-world discovery: proposed materials must also be experimentally synthesizable under feasible conditions. This talk will focus on our recent work toward synthesizability-aware materials discovery, including machine learning models that predict the likelihood of inorganic materials synthesis and identify plausible synthesis pathways. I will highlight how explainable AI can reveal the structural, chemical, and process-related factors that govern synthesis feasibility, thereby transforming prediction models into practical design tools. I will further discuss how agentic AI systems can integrate materials design, synthesis planning, feedback from experiments, and autonomous decision-making to close the loop between computation and laboratory realization.

Data-driven methodologies are transforming the discovery of functional materials by overcoming the limitations of human bias and traditional trial-and-error approaches. This is particularly critical in rechargeable battery research, where exploring the massive compositional search space of multi-component electrolytes is impractical using conventional methods. To address this challenge, researchers have developed a high-throughput, automated experimental platform featuring a closed-type 36-well multi-channel electrochemical (MCE) cell module and a non-contact dispenser. This robotic setup significantly suppresses electrolyte evaporation and enables the rapid preparation and evaluation of over 400 samples per week. To achieve fully autonomous materials discovery, this high-throughput hardware is integrated with artificial intelligence (AI) using NIMO, a newly developed modular orchestration system. NIMO allows for flexible closed-loop workflows by connecting various AI algorithms with robotic modules. As a demonstration, the integrated system utilized Bayesian optimization to autonomously explore and optimize multi-component electrolytes for sodium-ion batteries. Through iterative closed-loop cycles, the AI successfully identified specific additive combinations that maximized the coulombic efficiency of \ce{Na3V2(PO4)3}/hard carbon batteries. Importantly, this was achieved entirely through data-driven exploration, demonstrating the power of autonomous optimization without reliance on prior human intuition

For most of its history, the scientific method has followed a familiar rhythm: hypothesize, design, experiment, learn, repeat. Recent advances in AI, machine learning, and automated systems enable us to ask deeper questions about how science changes when all of these technologies come together. In this session, Santiago Miret will introduce Lila Sciences and its vision for building an Operating System for Science that leverages a platform combining advanced AI reasoning models with autonomous robot-centric laboratories (AI Science Factories™) to drive new scientific discovery. Santiago will share concrete examples of how this closed-loop paradigm is accelerating discovery across materials science, chemistry, and the life sciences, including non-platinum catalysts for green hydrogen production to novel sorbents for carbon capture.

Scientific research begins with the literature: both human and AI scientists rely on prior work to synthesize ideas, identify gaps and contradictions, and formulate new hypotheses. Despite rapid advances in artificial intelligence (AI) for scientific prediction, automation, and self-driving laboratories (SDLs), this foundational step, reasoning over the scientific literature at scale, remains largely inaccessible to current systems. Existing approaches typically rely on localized retrieval of a small number of documents, limiting their ability to aggregate evidence, detect global trends, or reconcile conflicting results across thousands of studies. In this talk, I present an approach that leverages the native language of materials for literature-informed, autonomous scientific discovery. The framework treats the research literature as a first-class data modality and integrates it with machine learning and experimentation in an iterative, closed-loop system. Moving beyond isolated predictions, it enables global reasoning across prior studies, supports causal hypothesis generation, and continuously refines insights based on new evidence. I will demonstrate use cases in the design of metal–organic framework (MOF) materials for carbon capture and storage.

AI for science is transforming research across materials science, environmental science, semiconductor design, and many other scientific and engineering domains. However, many scientific challenges require more than data-driven learning alone; they demand a deep understanding of the geometry, physics, and interactions that govern real-world systems. This talk introduces Physical AI Systems, built upon physical world models that integrate geometric representations, mathematical formulations, physical laws, simulation-based analysis, and data-driven intelligence. These models provide a computational understanding of environments, structures, and processes, enabling complex phenomena to be represented, explored, and validated across multiple scales, operating conditions, and design parameters. By combining parametric modeling, physics-driven simulation, and AI techniques, our framework transforms geometric and physical information into actionable scientific knowledge. It supports prediction, optimization, reasoning, decision-making, and design exploration beyond traditional data-centric approaches, enhancing scientific understanding and accelerating engineering innovation in real-world applications.

This talk presents how we use machine learning to radically accelerate the discovery of next‑generation construction materials within URBAX, a green cement and kaolin-based platform. By combining experimental data, physics-informed features, and modern ML models, we explore previously uncharted regions of compositional and processing space. This allows us to design binders that reach international construction strength and durability standards in only 5 days, instead of the traditional 28-day qualification window. Beyond speed, the approach reveals non-intuitive formulations and trade-offs that link sustainability, performance, and cost, illustrating how AI can transform the way we engineer materials for the built environment and, more broadly, how data-driven discovery can reshape resource-intensive industries.

Nanoemulsions are a class of biphasic colloidal materials comprised of 20-500-nm oil droplets dispersed in a continuous aqueous phase. These materials are capable of compatibilizing materials with dissimilar hydrophilicities, leading to their widespread use across many industries, including personal care products, pharmaceuticals, and agriculture. Synthesizing nanoemulsions involves balancing ingredients (often 5-10 at a time) and processing parameters (e.g., temperature, mixing conditions) to achieve small and stable droplets. Higher order interactions dominate this process, causing many formulators to rely on heuristics when developing new formulations. Moreover, a lack of synthesis standardization complicates the ability to learn from existing data or replicate any observed trends. We have thus combined a chemistry-aware machine learning model (brain) and an automated synthesis platform (body) to produce a first-in-class self-driving lab (SDL) for nanoemulsions. The brain is a high-performing and flexible ML model with a multi-class classification head for emulsification success. After demonstrating competitive performance with baseline models, it was deployed within a custom in-house graphical user interface (GUI) supporting the end-to-end machine learning workflow and database management, enabling formulators to simulate experiments, interpret model predictions and query results. The body is an automated synthesis platform with integrated dispensing, heating and mixing unit operations. Each unit operation is a standalone module individually controlled by a central orchestrator, enabling modularity in platform design. Custom mounts and interfaces are used to improve throughput and versatility. Initial tests confirm successful end-to-end automated synthesis that outperforms manual methods in both speed and consistency. Future work will integrate the brain and body for fully autonomous operation and accelerated formulation discovery.

The activity–stability balance of the oxygen evolution reaction (OER) electrocatalysts is dictated by how interfacial oxidation states gate both O–O bond formation and metal dissolution, yet these coupled events are rarely accessible to atomistic simulations at electrified solid–liquid interfaces. Here we combine a Ru–O–H machine-learned interatomic potential with enhanced-sampling molecular dynamics to resolve oxygen evolution and Ru detachment on RuO (110) in explicit water across distinct Ru oxidation environments. We find that moderately over-oxidized, *O-ligated Ru centers promote an oxide-path mechanism (OPM) in which adjacent *O species couple with a low free-energy barrier, while a solvent-structured “water gap” suppresses nucleophilic attack and disfavors *OOH formation via the adsorbate evolution mechanism (AEM). As the surface is reduced toward Ru4+-like states and adjacent *O motifs are depleted, the O–O coupling barrier rises, and the dominant pathway shifts back to AEM. In parallel, Ru dissolution proceeds through an oxidation-assisted route: additional water ligation stabilizes higher-valent Ru (Ru5+-like) and lowers the detachment barrier by ~1 eV relative to direct dissolution, rendering coordinatively unsaturated sites more vulnerable than bridge-bonded Ru. Guided by this mechanistic map, we develop an electrochemical pre-activation protocol that biases RuO2 toward an optimally over-oxidized interfacial state, enhancing OER performance while limiting Ru loss. These results establish Ru oxidation state as a unifying descriptor for controlling competing OER and degradation pathways at electrochemical interfaces.

Artificial intelligence has advanced rapidly, yet its impact in biomedicine remains limited by the quality and physiological relevance of the experimental models used to generate training data. At the Self-Driving Lab for Human Organ Mimicry (SDL6), we are developing human tissue models that better represent human biology and support autonomous experimentation. Our work combines label-free quality control of induced pluripotent stem cell (iPSC) maintenance and expansion, automated fabrication and functional evaluation of engineered cardiac microtissues, and robotic generation of vascularized multicellular co-culture systems spanning multiple organ types. Together, these capabilities enable reproducible production, long-term culture, and quantitative characterization of complex human tissue models. By integrating advanced biological models with automation and data-driven analysis, SDL6 aims to improve experimental reproducibility, generate content-rich biological datasets, and accelerate the development of predictive platforms for drug discovery, disease modeling, and regenerative medicine.

The stochastic optimization of a loss function during the training of deep neural networks shares many similarities with classical thermodynamical systems. By analysing stochastic differential equations that describe the evolution of (scale-invariant) neural network during training we derive the characteristics of its stationary state. Surprisingly it becomes very similar to ideal gas law. Following this similarity one may define analogues of temperature, pressure, and volume for neural networks. Using those analogies we establish various thermodynamic potentials such as Gibbs and Helmholtz free energies and show that they are minimized during training under popular training protocols.

There has been widespread adoption of machine learning (ML) technologies to discover new insights of the complex relationships among diverse parameters in various additive manufacturing (AM) techniques. The collaborative synergy between ML and AM holds the potential to achieve interesting function and drive new generation design of AM‐printed parts. In our work, we demonstrated such synergy for 3D printed soft robotics, with optimization and control enabled by ML. We will first present the overall potential of ML in materials and processes of AM, before presenting 2 interesting examples of soft robotics. We developed a 3D printing‐enabled artificially innervated smart soft gripper with variable joint stiffness where ML was used for joint angle prediction. Furthering the research for intelligent system, we recently reported a soft continuum robot that integrates a conductive polymer composite directly into a node-based lattice structure for intrinsic sensing, coupled with a neural network for near-real-time shape reconstruction. This approach enables near-real-time proprioception without compromising flexibility, with potential applications in biomedical manipulation and remote inspection.

The development of composite materials for sustainable separations, resource recovery, and energy technologies is often constrained by the vast multidimensional design spaces associated with composition, structure, processing conditions, and operating environments. Traditional trial-and-error experimentation is increasingly inadequate for navigating these complex parameter spaces and identifying optimal material architectures. Closed-loop autonomous laboratories offer a transformative paradigm by integrating artificial intelligence, automated experimentation, high-throughput characterization, and adaptive decision-making into self-driving discovery platforms. In this talk, I present a vision for the closed-loop autonomous discovery of functional membranes and two-dimensional (2D) materials for resource recovery and energy applications. The proposed framework combines machine learning-driven experiment planning, automated synthesis and fabrication, real-time characterization, and active learning to iteratively optimize material performance with minimal human intervention. Particular emphasis is placed on membrane-based ion separations, critical mineral recovery, electrocatalytic materials, and self-assembled 2D architectures. Several representative case studies will be discussed, including AI-guided optimization of 2D materials-based membranes for selective lithium extraction from complex brines and battery recycling streams, for $\ce{CO_2}$ capture and conversion, and closed-loop development of functional materials for selective recovery of precious metals from electronic waste. We further outline opportunities for integrating computer vision, robotic experimentation, digital twins, and autonomous scientific agents to control and optimize non-equilibrium self-assembly processes in real time. By coupling intelligent decision-making with automated experimentation, autonomous laboratories can accelerate the discovery of high-performance materials while reducing experimental co

Modern AI systems are increasingly aligned to human instructions and preferences. Yet alignment alone does not guarantee that AI understands the people, cultures, contexts, and realities in which it operates. A model can be helpful, harmless, and fluent—and still fail to connect meaningfully with society. This challenge emerges from a broader evolution of AI—from symbolic representations and curated annotations, to statistical learning, large-scale representation learning, and alignment—each expanding what machines can learn from human-generated data. This talk argues that the next frontier of AI is grounding: connecting AI not only to language and data, but also to human perception, interaction, emotions, culture, and domain knowledge. Grounding shifts the focus from generating plausible outputs to understanding context, adapting to diverse communities, and supporting real-world decisions. Drawing on examples from speech, language, and multimodal AI, I will illustrate how grounding can be achieved through human interaction, multicultural reasoning, and real-world deployment. Case studies include MERaLiON, the first multimodal large language model for Southeast Asia; SingaKids AI Tutor, a multilingual learning companion for children learning Malay, Mandarin, and Tamil; and Siu Dai, a telebot supporting patients in chronic disease management. As AI becomes embedded in education, healthcare, government, and everyday life, the challenge is no longer merely to align AI with human preferences, but to ground AI in human experience. The future of AI in society depends on it.

What drives people to have trust in using artificial intelligence (AI)? How does the institutional environment shape social trust in AI? This talk addresses these questions to explain the role of institutions in allowing AI-based technologies to be socially accepted. In this talk, social trust in AI is situated in three institutional entities, namely, the government, tech companies, and the scientific community. It is posited that the level of social trust in AI is correlated to the level of trust in these institutions. The stronger the trust in the institutions, the deeper the social trust in the use of AI. The talk draws on a cross-country survey in East Asia to show convincing evidence of how institutions shape social trust in AI and its acceptance. It reveals that trust in institutions is positively associated with trust in AI technologies. Trust in institutions is based on perceived competence, benevolence, and integrity. A higher level of trust in AI technologies leads to a higher level of intention to use these technologies. This has profound implications on the governance of AI in society. By taking into account institutional factors in the planning and implementation of AI regulations, we can be assured that social trust in AI is sufficiently founded.

As AI grows more capable and more widely adopted, it quietly becomes one of the largest social-engineering projects in history: billions of people will heed its advice on questions personal and professional. That makes one design choice decisive - which human values, society model, and assumptions we build into it. The natural target are human preferences and broader culture. They are accessible, they fit the commercial logic of giving customers what they want, and they wear the color of democratic legitimacy. I will argue this target is a trap. The values evolution and cultural history installed in us were optimized for short-term survival and tribal cohesion, not for the flourishing of large, technologically empowered societies - and they routinely fail us. We reward the yes-man and call sycophancy the joke of our field. We pull strings for our relatives and keep our country corrupt and poor. We chase income, status, and engagement - then lament the solitude. We let moral judgments quietly overwrite empirical ones. Much of the dysfunction in the world is a human-alignment failure; train AI to reproduce our preferences and we amplify it. But AI is not only a poison - it can be a cure. The case is not for aligning AI to who we are, but to who we wish we were. Strip away the in-the-moment impulse and what is striking is not that every culture wants identical things, but that every culture distinguishes its impulses from its aspirations - and the aspirations, where they are written down, converge far more than the impulses do: to be honest when a lie is easier (integrity), correct rather than flattering (competence), fair to the stranger as to our own kin (benevolence), faithful to the person we will be in ten years. We know what the better version looks like; we just cannot reliably be it. AI need not carry the baggage that defeats us - no tribe to favor, no ego to defend, no present bias unless we build one in. And for the first time we have the tools to realize these aspirations rather than merely state them: we can inspect an AI's reasoning and re-run it under counterfactuals; we can judge it against outcomes - did the business get built, the patient recover, the friendship survive - instead of against applause; and we can select for virtue, keeping the version that stays honest under pressure, as we rarely manage to do with human power. The better angel of our nature need not be a fiction we tell ourselves. It can be something we specify, measure, and build.

This talk will discuss recent advances at the interface of AI and physics, focusing on tensor learning methods for compressing quantum interactions in materials. We will show that such compression can accelerate first-principles calculations of electronic interactions by orders of magnitude and enable previously inaccessible many-body calculations, including accurate summation of all Feynman diagrams providing a numerically exact solution of the polaron problem. These advances open new directions for predictive calculations of transport phenomena and nonequilibrium dynamics in materials.

Tensor networks have long been a central computational tool in many-body physics, and more recently have emerged as a powerful framework for simulating quantum computers. In this talk, I will describe how tensor networks can also be used as function representations for fast generative modeling, where training and inference can be done through linear algebraic operations. I will also discuss how this opens new opportunities for enhancing physics simulations.

Memristive devices are promising candidates for future embedded non-volatile memory technologies, particularly as conventional embedded Flash faces increasing scaling and integration challenges in advanced CMOS nodes. Silicon-oxide memristors are especially attractive in this context because their material system is closely aligned with silicon CMOS technology, avoiding the need to introduce additional non-standard switching materials into the integration flow. This makes them a relevant platform for exploring dense, scalable and manufacturable memory solutions, while also offering prospects for emerging computing paradigms such as neuromorphic hardware. This talk presents CMOS-integrated silicon-oxide memristors implemented in a arrays of 1T1R cells, where each resistive switching element is addressed by a CMOS selector transistor. Their electrical behaviour is assessed under realistic operating conditions, with emphasis on retention and endurance as key reliability metrics for embedded memory operation. Retention measurements evaluate the stability of programmed resistance states over time, while endurance measurements assess repeated switching between high- and low-resistance states, providing insight into the practical operating window of the devices. These results are discussed in terms of their relevance for reliable memory programming, readout and cycling in circuit environments. In parallel, the talk introduces a SPICE-compatible compact model based on the memdiode framework, where current transport and memory-state evolution are described through coupled equations fitted to experimental data. The model reproduces DC and pulsed operation and enables circuit-level simulations of write, read and cycling schemes in hybrid CMOS-memristor circuits. This provides a link between measured device behaviour and circuit design, allowing the impact of device properties, variability and programming conditions to be explored before implementation. Finally, the talk will b

As self-driving laboratories expand the scope and complexity of experimental processes they can perform, ensuring safe operation becomes critical. I will highlight practical limitations of current AI and robotics systems in real laboratory settings and discuss opportunities for building more trustworthy and robust autonomous scientific platforms.

Molecular foundation models are becoming a key interface between molecular data, domain knowledge, and scientific decision-making. In this talk, I will present our recent line of work on molecular foundation models, progressing from pretraining–finetuning on molecular graphs to language-model-based molecular intelligence. I will first revisit how large-scale self-supervised graph pretraining, represented by GROVER, learns chemically meaningful molecular representations from unlabeled molecules, and how graph topology induced optimal transport improves efficient finetuning under scarce labels. I will then discuss the role of expressive architectures for molecules and biomolecules, including graph transformers, cross-dependent GNNs, hypergraph learning, and symmetry-aware models for protein interactions. Finally, I will describe our recent efforts to connect molecular foundation models with LLMs through hierarchical graph tokenization, and to evaluate LLMs in executable molecular workflows via MolViBench. I will close with ongoing work toward cross-domain tokenization for 3D atomic systems, aiming at a more general foundation model across molecules, proteins, and materials.

Improved Climate and Weather models are essential for humanity in a changing climate that causes more extreme weather events. Generative AI (GenAI) has become extremely popular in recent times with the advent of LLMs like ChatGPT and image/video generation tools like Stable Diffusion, DALL-E, etc. GenAI methods have also shown great promise in the Climate and Weather domain. Early work with GANs has been augmented with newer generative methods, such as denoising diffusion models and score-based generative models. This talk introduces the NVIDIA Earth-2 program for climate and weather modelling, discusses the features of these latest models and their potential applications. We will also share the latest trends in generative AI for climate and weather modelling.

We explore (i) how we can use pre-existing knowledge (e.g., physics) to improve AI systems, and (ii) how we can possible 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 general circulation models. On the second topic, we present some results on the use of explainable AI for Earth System applications in the context of extreme events. We conclude with an overall perspective bridging the two topics and grounded in the digital twin paradigm.

We present neural network methods for solving non-smooth PDE-constrained optimization problems. Our investigation focuses on three challenging categories: (1) optimization with non-smooth regularization, (2) optimal control of PDEs involving interfaces, and (3) optimal control of elliptic variational inequalities. For each category, we develop tailored neural network algorithms that exploit the specific mathematical structure of the underlying problem. The principal advantages of our methods are that they are mesh-free, thus avoiding grid generation challenges; computationally scalable to high dimensions and complex domains; and straightforward to implement. Extensive numerical experiments demonstrate their computational efficiency and accuracy on benchmark problems.

Quantum AI represent the next generation of computing, where quantum technology meets artificial intelligence to unlock the unprecedented computational power. It enables faster analysis smarter predictions and breakthrough innovation across industries, transforming how complex challenges are solved. The talk will focus on how Quantum AI and Autonomous Labs can power smarter labs, accelerate simulations, optimize experiments, and enable autonomous scientific discovery workflows across materials, chemistry, pharma, and deep-tech R&D.

Dr. Peichen Zhong is an Assistant Professor at the Department of Materials Science and Engineering, National University of Singapore. He obtained a BS in Physics from the University of Science and Technology of China (USTC) in 2018, followed by a PhD in Materials Science from UC Berkeley in 2023. He then completed the postdoctoral work at Lawrence Berkeley National Laboratory (LBNL) and Bakar Institute of Digital Materials for the Planet (BIDMaP). He was awarded the 2023 Rising Stars in Materials Science and Engineering by CMU/MIT/Stanford, the BIDMaP Emerging Scholar Fellowship from the College of Data Science, Computing and Society (CDSS) at UC Berkeley, and the AI2050 Early Career Fellowship by Schmidt Sciences.

Machine learning interatomic potentials (MLIPs) are transforming the landscape of computational chemistry and materials science. MLIPs enable atomistic simulations that combine the fidelity of computationally expensive quantum chemistry with the scaling power of AI. Yet, developers and computational chemists working at this intersection face a persistent challenge: a lack of robust, Pythonic tools for GPU-accelerated atomistic simulation.  In this talk, we present NVIDIA ALCHEMI Toolkit and NIM microservices which allow users to flexibly compose atomistic simulation workflows with ease. These simulations workflows are accelerated by optimized batch kernels from ALCHEMI Toolkit-Ops. We will also highlight partner and customer success stories across verticals, from catalysts to beauty products.

We present recent progress in the development of a GPU-accelerated Gaussian-type orbital (GTO) integral engine for high-performance electronic structure calculations. Our implementation enables efficient Hartree–Fock and DFT calculations for large molecular systems, achieving speedups of 15–66× over PySCF and 1.5–2.5× over GPU4PySCF in benchmark tests. The integral engine is designed with automatic differentiation in mind, allowing it to support emerging differentiable quantum chemistry workflows. Beyond molecular calculations, the framework is extensible to solid-state systems, including both three-dimensional and two-dimensional materials. These developments provide a flexible and efficient foundation for scalable, differentiable, GPU-native quantum chemistry simulations.

Ultrafast spectroscopy provides direct access to nonequilibrium dynamics in materials, resolving electronic, vibrational, and structural processes on femtosecond timescales. In this talk, I present how these time-resolved measurements can be combined with data-driven approaches to extract meaningful insight from complex datasets, which is a central strategy at the Institut Courtois. By integrating machine learning and statistical analysis, we identify patterns, reduce dimensionality, and uncover relationships between structure, dynamics, and function. This interdisciplinary framework connects chemical mechanisms, physical behavior, and computational modeling, enabling more efficient interpretation and guiding materials discovery. I highlight recent examples demonstrating how the synergy between ultrafast experiments and data-centric methods accelerates understanding and supports the identification of materials with targeted properties, along with key challenges and future directions. I will also present an overview of opportunities to participate in the scientific program of the Institut Courtois.

AI-driven materials discovery platforms can now propose candidate materials faster than traditional evidence pipelines can reliably evaluate them. This creates a new bottleneck: generating trustworthy physical evidence fast enough to guide models, experiments, and subsequent decisions. In this talk, I will present CatPlat, a high-throughput simulations platform to generate high-fidelity data for validating new materials. CatPlat provides reusable computational building blocks for constructing, executing, and analysing large-scale simulation campaigns, while recording detailed provenance for every calculation. This makes it a versatile and verifiable data generation platform for training downstream models and guiding experimental decisions. I will demonstrate the utility of CatPlat by showing how it enabled closed-loop discovery of novel $\ce{CO_2}$-to-methanol catalysts with a 95% improved yield within six months.

Artificial intelligence has shown tremendous potential for scientific and engineering discovery, yet many practical applications remain constrained by sparse measurements, limited simulation data, and the need to satisfy physical laws. In this talk, I will present recent advances in physics-informed AI for inverse estimation and generative design. Topics include uncertainty-aware parameter identification from sparse sensing data using reduced-order physics models, as well as physics-constrained diffusion models for inverse design. These examples demonstrate how embedding physical knowledge into learning and optimization processes can improve data efficiency, reliability, and generalization, providing new opportunities for next-generation scientific and engineering discovery.

Machine learning force fields (MLFFs) have emerged as powerful tools for molecular dynamics (MD) simulations of complex materials, offering near-ab initio accuracy at significantly reduced computational cost. In solid electrolyte research, MLFFs enable us to study critical phenomena, including phase transition behaviors and ionic transport mechanisms, providing fundamental insights into these materials. However, a fundamental challenge persists: when MLFF-based MD simulations yield unexpected or even unphysical results, determining whether discrepancies originate from could be difficult due to the black-box nature of MLFFs. Discrepancies could come from limitations of MLFF frameworks, insuffciently sampled training datasets, or overlooked physical mechanisms. In this talk, I will share our experience in using MLFF for MD simulations in several solid electrolyte materials. Specifically, we (try to) address the following key questions: 1) Why our MLFFs work well for cubic-LLZO but always fails in tetragonal-LLZO? 2) Why do our MLFFs exhibit good training metrics yet still fail during long-time or large-scale production MD simulations? 3) How much training data is required to develop a reliable MLFF for solid electrolytes, and is highly accurate DFT reference data essential? 4) What are the trade-offs involved in explicitly including long-range interactions in terms of force field accuracy and performance? We hope our successes and failures (in particular failures) of applying MLFFs to MD simulations in solid electrolyte materials may provide valuable insights for the community, offering practical lessons for extending MLFFs to other complex material systems.

Accelerating the discovery and development of new polymers is a crucial step towards meeting pressing societal challenges. Data-driven approaches are a powerful part of this process but require high-quality and consistent experimental data spanning a meaningful portion of composition space. Unfortunately, polymers are an especially difficult material class in the context of informatics due to the high importance of processing and large variations in protocols. Further, it is often very difficult to compare polymers that are synthesized using different methods. Here, we report our work to springboard polymer discovery and development through the combination of modular chemistry and modular automation. First, we discuss the electrodeposition of polymer networks (EPoN), a flexible approach to depositing and modifying polymer films. This process involves the attachment of electroactive crosslinking groups onto polymers or prepolymers and subsequently depositing them as conformal and self-limiting films. We have shown this method is capable of realizing a wide array of homopolymer and block copolymer functional films under comparable conditions. To turn this process into a platform for polymer discovery, we combine it with the polymer analysis and discovery array (PANDA), a low-cost robotic system that is able to dispense fluid, perform electrochemistry, and optically characterize EPoN films in custom well plates. The PANDA is part of a larger automated distributed equipment network (DEN) wherein samples can be automatically moved to different stations for structural, analytical, and functional characterization including Raman and UV-Vis spectroscopy, contact angle measurement, and mechanical characterization including adhesion and stiffness. We discuss our initial campaigns to use the combination of EPoN and automation to seek PFAS-replacing non-wetting polymers and optimize materials for carbon capture. Ultimately, the combination of modular chemistry and automation al

The proliferation of artificial intelligence and machine learning technologies across the physical sciences has catalyzed a paradigm shift in research methodologies, enabling unprecedented opportunities for research acceleration. Pharmaceutical research organizations are increasingly leveraging these capabilities to transform traditional approaches to drug development. This presentation will examine MSD Development Sciences and Clinical Supply's strategic digital transformation initiative, specifically addressing the design, development, and deployment of AI/ML-enabled experimental design platforms within our research ecosystem. We will present our systematic approach to integrating contextualized data, predictive models, optimization algorithms, and advanced experimental planning tools to enhance research efficiency and decision-making accuracy.

Atomistic simulation is undergoing a fundamental shift, driven by the rapid maturation of machine learning interatomic potentials (MLIPs). Foundation models now deliver near-quantum accuracy across the periodic table, yet the surrounding simulation infrastructure has struggled to keep pace. Most engines remain CPU-centric legacy code, leaving GPUs underutilized and bottlenecked by neighbor list construction, dispersion corrections, electrostatics, and per-system serial execution.  We present the NVIDIA ALCHEMI Toolkit — a GPU-first, composable stack designed to accelerate the full MLIP workflow, from training and fine-tuning to production simulation. ALCHEMI Toolkit-Ops provides batched, differentiable primitives (neighbor lists, DFT-D3, PME/Ewald electrostatics, and MD integrators) with PyTorch and JAX bindings that drop into any existing MLIP pipeline. The Toolkit is designed to be easy to use — researchers can compose and scale multi-stage batched pipelines in a few lines of Python. We will demonstrate a few concrete use cases and highlight adoption across verticals and discuss how the community can shape the roadmap.

Inverse design of sodium-ion cathode materials with machine learning interatomic potentials capable of efficiently and accurately describing various site-disordered phases within the phase space. https://openreview.net/forum?id=xJeKm7k3cL

What if you could design, train, and validate an entire robotic chemistry workflow before running a single real experiment? This talk presents Matterix, a GPU-accelerated digital twin framework that captures key physical and operational elements of a chemistry lab, including robotic manipulation, liquids and powders, heat transfer, and known reaction kinetics, within a photorealistic, physics-grounded environment. A modular semantics engine bridges continuous dynamics and discrete logical states, enabling workflow modeling across multiple levels of abstraction. On top of this foundation, Matterix enables training of robotic skills and control policies directly in simulation, which can then be transferred to real systems. We demonstrate sim-to-real deployment in chemistry lab automation, reducing reliance on costly experimentation and enabling workflows to be designed, tested, and validated entirely in silico.

Scientific modeling and computation traditionally rely on structured mathematics and hand-designed algorithms. In this talk, I propose a new perspective: treating both modeling and computation as processes operating within the space of natural language. I will introduce two complementary approaches that realize this vision. The first uses symbolic learning based on tree structures to generate mathematical expressions, where modeling is performed by constructing symbolic trees and computation is governed by operator rules. The Finite Expression Method (FEX) exemplifies this approach by discovering interpretable, high-accuracy solutions to PDEs and physical systems. The second approach employs large language models (LLMs) for automatic code generation and reasoning to translate scientific problem descriptions into formal mathematical models and executable solvers to solve these problems. As an example, the OptimAI framework demonstrates how multi-agent LLM collaboration enables reliable end-to-end optimization problem modeling and solving. Together, these methods point toward a unified paradigm where symbolic and language models form the foundation for interpretable, scalable scientific discovery and computation.

At the heart of materials science studies for next generation materials is an idea that we want to be studying real materials doing real things, often in real devices in real time. In practice, this presents a number of key data analysis and interpretation challenges because it implies we are studying ever more complicated samples, often in complex heterogeneous environments and in time-resolved operando setups, and we are interrogating our data for more and more subtle effects such as microstructures and evolving defects and local structures. Advanced data analysis algorithms and software are essential for the success of this enterprise.  In this talk I will describe various developments that leverage the power of artificial intelligence (AI), principally machine learning (ML), to aid in this task. Some of these powerful tools are clearly ready to be applied more broadly in the community and others are still in the future but look very promising.  They include unsupervised and supervised machine learning approaches, conventional ML and deep neural networks, including generative models, as well as approaches for autonomous time-resolved experimentation. I will lay out various aspects of ML in structure science using examples from our own work.

The plane wave basis is widely used in Galerkin approximation, due to its periodicity and computational advantage, where the fast Fourier transform (FFT) can be applied. However, since its spatial resolution is uniform, the number of basis functions required can be excessive for problems with rapidly varying local features. We propose an adaptive basis called flow-distorted plane wave (FDPW), where the bijection of a normalizing flow is used to distort the problem domain, hence achieving adaptive resolution. We apply FDPW to Kohn-Sham density functional theory (DFT) calculations to solid-state systems, demonstrating improved speed and memory usage.