Agenda

This opening session will set the stage for the workshop by tracing the conceptual foundations of Generative AI, illuminating both its historical lineage and the methodological advances that have resulted in its major breakthroughs in recent years, making it a transformative force for revolutionizing science and engineering. Via a series of talks, this session will provide a conceptual and technical framing of GenAI methods, highlighting key developments and themes that will resonate throughout the workshop. It will also lay the foundation for the sessions that follow, which will explore emerging challenges and opportunities in method development and testbed design, as well as domain-specific applications in ecology, agriculture, materials science, manufacturing, and the life and health sciences.

Talk 1: Generative AI: Historical Roots, Enabling Advances, and Emerging Capabilities (Anuj Karpatne, Aryan Deshwal)

We will describe the historical roots and key methodological advances that have catalyzed recent progress in GenAI,  including deep representation learning, self-supervised learning and context-sensitive architectures like transformers. We will discuss how current GenAI developments have been made possible by the cross-pollination of ideas from multiple fields, including energy-based models, Bayesian frameworks, auto-regressive models, latent space models, forward-inverse modeling, and adversarial optimization. This cross-pollination of ideas has led to a rich landscape of modern generative AI methods including diffusion models, GPT, variational autoencoders, normalizing flows, flow matching, GANs.

We will present a two-dimensional organization of this landscape : 1. How different generative AI methods approach learning the underlying probability distribution of data instances (explicit vs. implicit)? (2) What family of neural network architectures they employ (encoder-decoder, encoder only, or decoder only)? Additionally, we will present the new capabilities that are being unlocked with modern GenAI methods such as in-context learning, few shot/zero shot learning, and plug-and-play guidance, that underpin the rise of foundation models. The talk will conclude with a discussion of how these GenAI methods and capabilities are beginning to be leveraged in scientific and engineering applications.

Talk 2: Opportunities for scientific discovery with Reinforcement Learning and GenAI (Jana Doppa)

The talk will explore recent advances in the GenAI landscape that is the growing interplay between reinforcement learning (RL) and GenAI.  In the context of LLMs, RL has already become a critical component in enabling reasoning, planning, and goal-directed behavior, especially when combined with generative architectures in tasks that require decision-making under uncertainty. In particular, RL has seen popular recent use in post-training of LLMs, including RL from human feedback and RL from verifiable rewards that led to models like ChatGPT and DeepSeek R1/OpenAI O1, respectively. We will discuss potential untapped opportunities of RL combined with GenAI for scientific discovery. We will highlight the unique challenges in scientific applications that require development of new ideas in RL. 

Talk  3: Opportunities for Advancing GenAI Inspired by Needs in Science and Engineering (Anuj Karpatne, Aryan Deshwal)

This talk will discuss open challenges and capabilities required for leveraging GenAI in scientific discovery and how scientific domains can provide the next frontier for fundamental advances in GenAI. This includes incorporating scientific knowledge through the emerging field of knowledge-guided machine learning (KGML), generating and validating scientific hypotheses, supporting principled decision-making, experimental design, and providing scientifically meaningful explanations. We will provide an overview of the rich landscape of research happening in KGML through a multi-dimensional structure covering varying forms and properties of scientific knowledge, frameworks for incorporating knowledge in AI/ML models, and scientific use-cases motivating applications of KGML. We will also discuss how recent advances in KGML have the potential to create entirely new ways of building generative AI and Foundation models for science and engineering, that are anchored in scientific laws, offer out-of-distribution generalizability even with limited labels, promote scientific interpretability and trustworthiness, and enable scientifically controlled generation of outputs necessary for accelerating scientific discovery.

The talk will conclude by highlighting ecosystem requirements identified in the 2024 AI4Science workshop report such as access to AI-ready benchmark datasets, standardized evaluation metrics and protocols, cyber-infrastructure for building and deploying GenAI solutions, and trans-disciplinary training, engagement, and collaborations between AI and domain scientists. These priorities will serve as a springboard for the deeper exploration that will take place in the panel discussion that follows.

Scientific discovery places unique demands on both GenAI methods and the testbeds that support their development and evaluation. GenAI models must not only generate realistic outputs, but also respect physical laws, integrate domain knowledge, and support scientific reasoning. Testbeds must go beyond generic benchmarks to include domain-specific datasets that are unbiased, representative, and capable of supporting rich exploration across model architectures, training regimes, and scientific use cases. They must also span multiple technology readiness levels and AI stack layers—from hardware and software platforms to application support environments and downstream integration with scientific workflows.

This panel will examine key challenges that arise in the scientific use of GenAI, including out-of-distribution generalization, scientifically grounded explanations, symbolic reasoning, uncertainty quantification, hypothesis validation, and the high dimensionality and computational cost of scientific models. The expectation of reproducibility and repeatability adds further demands that current GenAI systems are not yet equipped to meet. Building the next generation of GenAI methods and testbeds that explicitly address these challenges will be essential for scaling scientific innovation and enabling reliable, trustworthy advances across disciplines.

This panel will examine the transformative potential of GenAI to advance ecological research, with applications ranging from modeling aquatic systems, monitoring ecosystem dynamics, and analyzing carbon fluxes. It will highlight the distinctive strengths of GenAI compared to traditional process-based models, particularly in capturing complex nonlinear ecological interactions automatically from data. The discussion will also address key limitations of current GenAI approaches in ecology, including challenges posed by sparse, multi-resolution, and multi-source data, the spatial and temporal variability of environmental systems; and the difficulty of transferring models and knowledge from well-observed regions to data-poor areas.

This panel will explore use-inspired Generative AI research across agriculture and environmental systems, including crop production, forestry, soil health, and hydrological modeling. The discussion will highlight opportunities to enhance productivity, resilience and sustainability in the face of increasing environmental variability and resource pressures. Panelists will address distinctive challenges such as agricultural exceptionalism, spatial heterogeneity, limited economic margins, and complex interactions between natural and human systems. A key focus will be on advancing GenAI approaches that incorporate domain knowledge to improve generalization, robustness, and interpretability—delivering actionable insights for regenerative agriculture, sustainable forestry, and effective management of land and water resources. These topics reflect growing interest in AI-driven solutions that support rural economy, agricultural systems and long-term stewardship.

A series of 3-minute lightning talks by UMN researchers.

Dan Boley (Computer Science & Engineering): Guiding and Disentangling Signals in Environmental and Neuro-Sciences, video recording

Caiwen Ding (Computer Science & Engineering): Mullti-Agent LLM-Guided Hierarchical Chiplet Design with Adaptive Verification and Validation, video recording

Qizhi He (Civil, Environmental, and Geo- Engineering): Physics-Constrained Differentiable Modeling and Inverse Design with Latent Machine Learning, 

Zac McEachran (Soil, Water, and Climate): Knowledge-Guided Machine Learning for Flood Forecasting, video recording

Shancong Mou (Industrial & Systems Engineering): Physics-Informed Neural ODE with Heterogeneous Control Inputs Quality Prediction of Composite Adhesive Joints, video recording

Murti Salapaka (Electrical and Computer Engineering): Machine Learning Enabling Single Molecule Physics, video recording

Viven Sharma (Earth & Environmental Sciences): GeoSimGPT: LLM assisted AI agent for automating geodynamic simulation workflows, video recording

Ju Sun (Computer Science & Engineering): Accelerating Materials Discovery by Learning with Physics-Informed Constraints, video recording

Suo Yang (Mechanical Engineering): Predicting Critical Parameters of Hydrocarbon Species using Bidirectional Encoder Representations from Transformers

Li Wang (Math): Measure theoretic approaches for uncertainty propagation, video recording

This panel examines challenges and opportunities for GenAI in advanced design and manufacturing systems, with a focus on energy efficiency (both in AI model training and inference, and plant operations), generative methods for process and design space exploration, automation (including embodied AI, robotics), safety (including human-robot and multi-robot environments), and security (including physical as well as cybersecurity). These challenges manifest themselves in the form of verifiable models including test-pattern generation and generative test programs, low-power devices and circuits, robust cyber-physical operation, interactions between a large number of AI agents across the design and manufacturing chain, secure and privacy-aware frameworks for execution and analysis of process data, and need for AI standards across the supply chain. 

The discussion will also explore the impact of automation on workforce development, strategies for upskilling, and the need for supporting infrastructure—including data provenance, training data ecosystems, and cross-platform interoperability.

Materials discovery to processing for desired functionality provides a remarkably rich set of applications and challenges for GenAI. The associated AI models scale from sub-atomic properties at femtoseconds to meso-scale models that scale to meters and years, and beyond. They leverage extensive datasets and first-principles models to generate new insights into structure, dynamics, and function, stimulus (e.g., materials under stress, energetics), and applications. Emerging applications in quantum, bio-functional, and energetic materials represent emerging frontiers in diverse domains. This panel explores the challenges and opportunities for GenAI in materials modeling and design. It explores the interfaces between complex first-principles models and traditional AI frameworks, the need for extreme spatio-temporal scaling and abstraction, novel representations that facilitate explainable inference, and exploration of high-dimensional design spaces. 

The dream of precision health is to develop a data-driven, continuous learning system where new health information is instantly incorporated to optimize care delivery and accelerate biomedical discovery. The confluence of technological advances and social policies has led to rapid digitization of multimodal, longitudinal patient journeys, such as electronic health records (EHRs), imaging, and multiomics. Our overarching research agenda lies in advancing multimodal generative AI for precision health, where we harness real-world data to pretrain powerful multimodal patient embedding, which can serve as digital twins for patients. This enables us to synthesize multimodal, longitudinal information for millions of cancer patients, and apply the population-scale real-world evidence to advancing precision oncology in deep partnerships with real-world stakeholders such as large health systems and life sciences companies.

This panel explores the unique challenges and transformative potential of Generative AI in life sciences and healthcare, particularly in high-impact use-cases that demand precision, personalization, and trustworthiness. Unlike traditional AI systems that optimize for population-level outcomes, GenAI in this domain must support individualized reasoning in addition to population health, rare outcome modeling, and the integration of domain-specific prior knowledge. Key topics include causal and interventional inference, model introspection and explainability, and strategies to address bias, confounding, and uncertainty in complex biological data. The panel will also examine foundational issues such as data accessibility, privacy, portability, and the need for fair, transparent, and compliant AI systems. These challenges and opportunities reflect growing national efforts to promote safe, equitable, and scientifically grounded AI innovation in the life sciences.

Federal agencies have long played a critical role in advancing research at the intersection of artificial intelligence, science, and engineering. This panel convenes representatives from agencies such as NASA, NIST, NSF, the U.S. Army, the Department of Energy, and the U.S. Forest Service to reflect on their experiences supporting mission-aligned research across domains including health, materials, manufacturing, and national security.  The discussion will explore opportunities and challenges related to Generative AI in these contexts, including topics such as scientific foundation models, interdisciplinary collaboration, data and infrastructure needs, and workforce development. Drawing on their unique vantage points, panelists will share observations on how the research landscape is evolving and how the broader scientific community might engage with emerging directions.  The session aims to catalyze conversation between researchers and agency-affiliated experts, offering insights that can help shape the future of AI-enabled discovery and innovation.

This concluding panel will reflect on key insights and recurring themes that have emerged across the workshop. Panelists will discuss shared challenges and domain-specific adaptations, critical gaps, promising directions, and methodological needs at the intersection of Generative AI and scientific discovery. The session will also focus on shaping a post-workshop report that captures the community’s vision for advancing GenAI in scientific domains—highlighting priorities for foundational research, infrastructure, evaluation, and collaboration. Participants will be invited to contribute ideas for framing, structuring, and dissemination of the report, with the goal of informing funding agencies, policymakers, and the broader stakeholder community.