Workshops

Buildings are vitally important to our economy, energy systems, and daily lives. Nationally, the commercial sector accounts for nearly 20% of primary energy use, over a third of electricity use, and nearly $200B in annual energy expenditures. 17% of our GDP is attributed to the real estate and construction industries alone. As we enter the AI era, we are asking more of our buildings than ever before – that they operate as efficient cost centers, provide low-energy, comfortable indoor environments, and increasingly, be dynamically coordinated with the power grid. This talk will highlight current trends in the industry, and research on the potential for AI to advance building performance across the lifecycle. With a focus on software tools for the operational phase, Dr. Granderson will present recent findings on how AI is being used in today's commercial products and services. She will conclude with highlights of ongoing research underway at Lawrence Berkeley National Laboratory, spanning advances in AI-ready data infrastructure, autonomous control and operation, and robotics for inspection and maintenance.

The process of finding new materials, optimal for a given application, is lengthy, often unpredictable, and has a low throughput. The integration of machine-learning and artificial intelligence techniques in discovery workflows can drastically accelerate the cycle, delivering a throughput otherwise impossible with first-principles methods. In this talk, I will describe two alternative computational strategies to achieve such acceleration. Firstly, I will discuss a family of machine-learning models for predicting materials quantities, with a specific example in the Curie temperature of ferromagnets. These use solely the chemical composition of a compound as feature and experimental data as target. The creation of models of this type can be drastically accelerated by the automated extraction of experimental data from published scientific literature, and by AI agentic schemes for the design of materials descriptors. Then, I will describe how an accurate description of the structure of materials, which is amenable to be used with machine learning, can offer a quantum-chemistry-accurate description of local properties at virtually no computational costs. This is based on our developed cluster expansion formulated in terms of Jacobi-Legendre polynomials. Such general framework allows us to construct, on the same footing, extremely accurate force fields as well as converged DFT charge densities. These can both be integrated in material-discovery workflows generating unprecedented speedup.

Quantum computing has the potential to contribute to challenging problems in chemistry and materials science, but realizing that potential in the near term will require close integration of quantum computing with advances in classical computing and AI. In this talk, we will briefly review how AI can improve key components of quantum computation, including circuit optimization and quantum-state preparation. We will then present ADAPT-GQE, a generative AI framework that learns to synthesize ground-state preparation circuits for electronic-structure calculations. We demonstrate the approach on imipramine, a tricyclic antidepressant that provides a challenging, chemically relevant target for computational modelling in drug-stability protocols. ADAPT-GQE achieves order-of-magnitude reductions in circuit-generation time compared with ADAPT-VQE while maintaining comparable or improved state-preparation accuracy. These results establish a pathway toward automated quantum circuit synthesis for utility-scale quantum computational chemistry.

Open research questions, barriers to deployment, and priorities for the next five to ten years.

What evidence is needed to move an AI-enabled energy technology from research toward funding, adoption, and deployment.

Topic tables on the major research challenges for the next five to ten years, the advances needed to address them, and how to show that they work, organized around three areas: what we build (materials, devices, and manufacturing), how we operate it (planning, forecasting, control, and coordination), and how we defend it (reliability, restoration, and security).

Traditionally, decision making begins with people gathering information and then choosing how to act. Digital platforms are fundamentally changing this process. They increasingly determine what information people see and use that information to shape their decisions. These platforms influence how we navigate cities, consume energy, shop online, interact socially, allocate resources, vote on candidates, and participate in markets. AI is dramatically expanding their capabilities, enabling them to learn from data, interact continuously with users, and influence the decisions of millions of people and autonomous agents. Platforms such as navigation and e-commerce systems observe users’ actions and develop detailed information about the state of the system. To influence decisions effectively, however, they must also understand users’ private preferences, objectives, constraints, and valuations. Platforms can learn the probability distribution of these characteristics from accumulated data, while advertisements, recommendations, and other interactions allow them to refine their estimates. These estimates are imperfect, particularly when preferences change with context. Effective personalization therefore requires platforms to elicit private information in real time. This emerging architecture gives rise to a new control paradigm based on information markets. Rather than directly controlling the actions of autonomous agents, an information market designs the information each agent receives to influence collective behavior and achieve a desired system objective. It must also create appropriate incentives for agents to reveal their private information. The monetization of information introduces a new feedback loop involving both information and payments, effectively creating a market for data. Designing such a market requires the simultaneous design of elicitation mechanisms, personalized information, and learning algorithms in a distributed environment. This paradigm presents both promise and peril. Carefully designed information can improve individual decisions, coordinate collective behavior, and enhance system performance. The same architecture, however, can enable platforms to manipulate the information people see, exploit information asymmetries, and exercise market power. Recent class action lawsuits against major e-commerce platforms illustrate the risks of using informational power to influence prices, demand, and consumer behavior. This talk will examine how information can be designed to produce desired outcomes, when such design serves a societal objective, and when it becomes a mechanism for manipulation. Understanding this duality is essential to building AI-enabled platforms that realize the benefits of personalized information while limiting the risks of concentrated informational power.

As AI shifts from a tool we use to a partner we direct, the challenge for universities is no longer teaching individuals to prompt a model — it is learning, as institutions, to work alongside increasingly capable AI systems. Drawing on Northeastern’s effort to weave AI readiness through its curriculum, faculty roles, and governance, this talk separates three competencies too often collapsed into one. Literacy is understanding what these systems are. Fluency is applying them skillfully within a discipline. Agency is the disposition to frame problems, direct and interrogate AI, and own the outcomes — to orchestrate human–agent teams toward human ends. As AI grows more capable, the frontier of human learning moves from fluency to agency. I will offer a practical framework for what each requires, and the structures that make them real rather than aspirational.

Human flourishing, which has a deep history in philosophy and religion, requires developing the abilities and living under conditions in which a good or decent life is securable. Accounts of human flourishing and human well-being usually feature prominently the skills and habits of mind necessary for persons to know, critically reflect upon, and understand the empirical world, as well as the abilities to form and revise conceptions of the good and of other-regarding morality. In addition to personal capabilities, humans must also live under conditions that are safe, adequately resourced, and well maintained by the appropriate institutions to enable them to effectively apply these skills and flourish. However, the potential for individuals to live good lives is heavily dependent on the social systems and institutions that define what is possible. These systems and institutions can nurture human flourishing or undermine it. Educational systems, including higher education, are designed to advance multiple elements of well-being, perhaps most notably, knowledge and understanding. The skills and abilities necessary for knowledge and understanding are developed in the first decades of life but must be supported over the course of a lifetime. For those fortunate enough to continue their education into their twenties, institutions of higher education have a special obligation to see that these abilities continue to develop, both generally and in ways specific to particular professions and callings. An urgent question for higher education is when and how AI use can degrade or deepen these abilities and what it means to advance knowledge and understanding as AI enters and seemingly takes over the chat.

The culture of K-12 schools and institutions of higher ed typically reinforce a larger societal focus of students on educational outcomes. When adolescent students care more about work completion, GPAs, and college and job acceptances than they do about what they are meant to learn, why wouldn’t they use genAI tools as a shortcut whenever they can? The ethics of cheating can’t be the only answer because, simply, it’s not enough to stem the tide when genAI is an everywhere tsunami. However, when we reframe the goal of learning as human development, a different pathway emerges for considering the educational use of AI in adolescence (ages 12 to 25). In this talk, we will explore recent neuroscience research demonstrating the developmental impact on adolescents of what Dr. Mary Helen Immordino-Yang calls “transcendent thinking.” When comparing this kind of emotionally infused intellectual engagement in high school and undergraduate coursework to the act of crawling in young children, a larger argument for the strategic and careful use of AI by students and teachers takes shape. As adolescents rush ahead with AI tools that are, in many ways, running amok themselves, we must wrestle with how our youth are being cheated of powerful developmental opportunities to have emotions about ideas and make meaning of their world, themselves, and who they are becoming as young adults.

For several decades, education policymakers have focused on preparing young people for the labor market: the human capital theory of education. In recasting education purposes, the OECD high-performing education systems are looking to link education to the ancient ideal of human flourishing, while rebalancing it in the service of a broader idea: to nurture in all of us a suite of distinctive human capacities, that equip us not only to flourish as individuals but also to contribute to flourishing societies and economies, in harmony with the planet. The development of education for human flourishing coincides with rapid advances in AI. In the near term, AI threatens to undermine human meaning, human agency, and human security. In the longer term (which may not be very long at all), superintelligent AI could threaten the survival of humanity itself. Education for human flourishing, in broadening and rebalancing human capabilities, restoring meaning to human lives and creating fair and sustainable models for the future, may be our best shot at controlling it.

When people think of applying generative AI in education, they often place the AI in the role of an expert tutor, but this dynamic can offload too much cognitive work to the AI at the expense of student learning. In this talk, I argue that collaborative learning interactions between students and agents might be a more effective metaphor in some learning scenarios. I outline the different kinds of collaborative roles agents might fill with students, such as an agent acting as a protege, and when these roles might be appropriate. I then discuss considerations in computational modeling of these human-agent dialogic processes, as well as the technical foundations for creating theory-grounded dialogic interventions with generative AI.

Measurement theory guides the development of tests and the interpretation of their scores. For a long time, psychometric methods (i.e., the statistical tools used to analyze test response data) have provided ways to empirically evaluate score reliability, flag questions with quality or fairness concerns, and estimate student proficiency, skill mastery, and other cognitive and non-cognitive characteristics with quantified uncertainty. As educational assessment moves beyond standardized tests with simple, structured responses (e.g., multiple-choice) toward more complex tasks, measurement theory principles remain useful, but many statistical models built for questions’ ordinal scores cannot be applied directly to unstructured data (e.g., log files, open-ended responses). How can we extend psychometric tools as behavioral evidence moves beyond right/wrong or ordinal question scores? In this presentation, I will share several projects that use test-taking process data to address testing and educational questions, including test reliability (improving the accuracy of student proficiency scores), question fairness (understanding why a question was flagged for bias against a subgroup), and performance gaps between demographic groups (identifying problem-solving behaviors that explain the performance gap of students with learning disabilities and their peers). I will also discuss some future extensions for diagnostic assessment with complex tasks, where students’ intermediate work and solution paths may provide evidence about specific skills.

Effective tutoring requires a model of what a student knows, a way to measure how that knowledge changes, and realistic predictions of how the student will respond. This talk presents a line of work connecting these three goals in open-ended math tutoring dialogues. We first discuss how large language models (LLMs) can be used to identify knowledge components and assess response correctness from dialogue, enabling knowledge tracing, a common student modeling technique, in dialogues. We then detail a joint item-response framework that treats tutor-posed tasks in dialogues as formative assessment items alongside standard assessment items, and show that this framework improves item parameter estimation and student response prediction. We finally explore LLM-based student simulation in tutoring dialogues. We discuss evaluation criteria across linguistic, behavioral, and cognitive dimensions, which reveal that prompt-only simulators often create plausible but unfaithful students. We then show that grounding the student simulation task on their prior item response and dialogue history, learning how to summarize the history through reinforcement learning, substantially improves student simulation quality.

A live Open to Debate recording in Shriver Hall.

As the impact of artificial intelligence (AI) continues to proliferate, computer architects must assess and mitigate its energy demands. This talk will survey strategies for mitigating the energy used by AI computation and datacenter infrastructure, drawing on data and experiences from industrial, hyperscale systems. First, we analyze the energy implications of super-linear AI growth. Second, we re-think datacenter infrastructure and define a solution space for sustainable computation with renewable energy, utility-scale batteries, and job scheduling. Finally, we develop strategies for datacenter demand response, incentivizing workloads to modulate power usage in ways that reflect their performance goals. In summary, the talk provides a broad perspective on sustainable computing and outlines the many remaining directions for future work.

Two fundamental dimensions of “daily life” are the activities we pursue as workers and as democratic citizens. The development of AI technology, with the threat it poses in terms of a massive reduction of job opportunities for humans, forces us to re-examine the values that are realised through work and the extent to which they can be achieved through non-work activities, such as play. Relatedly, we need to re-examine the connection between work and active democratic citizenship: to what extent does a healthy democratic ethos depend upon widespread participation on the part of citizens in work?

Despite persistent criticism, AI is now undoubtedly part of everyday life including healthcare. The opportunities for AI and related technology to benefit healthcare delivery and operations and sheer limitless, ranging from decision support systems for scheduling and triage to autonomous surgery. In this talk, I will highlight some of our recent advances in AI-assisted healthcare showcasing some of the transformative potential this technology holds for delivering better care.

Robotic assisted surgery (RAS) systems incorporate highly dexterous tools, hand tremor filtering, and motion scaling to enable a minimally invasive surgical approach, reducing collateral damage and patient recovery times. However, current state-of-the-art telerobotic surgery requires a surgeon operating every motion of the robot, resulting in long procedure times and inconsistent results. The advantages of autonomous robotic functionality have been demonstrated in applications outside of medicine, such as manufacturing and aviation. A limited form of autonomous RAS with pre-planned functionality was introduced in orthopedic procedures, radiotherapy, and cochlear implants. Efforts in automating soft tissue surgeries have been limited so far to elemental tasks such as knot tying, needle insertion, and executing predefined motions. The fundamental problems in soft tissue surgery include unpredictable shape changes, tissue deformations, and perception challenges. The research goal is to transform current manual and teleoperated robotic soft tissue surgery to autonomous robotic surgery, improving patient outcomes by reducing the reliance on the operating surgeon, eliminating human errors, and increasing precision and speed. This presentation will discuss our novel strategies to overcome the challenges encountered in soft tissue autonomous surgery. Presentation topics will include a robotic system for supervised autonomous laparoscopic anastomosis and our latest work in end-to-end imitation learning of surgical tasks and procedures, which we recently published in Science Robotics.

Machine Learning and Causal Inference for Extreme Event Attribution in Climate Science (Cassie Chou, Department of Biostatistics); Towards Sustainable Distributed Computing: Integrating Energy Costs and Benefits for Optimal Growth in Traffic Autonomy (Junyue Jiang, Department of Civil and Systems Engineering); Bridging the AI Literacy Chasm: Initial Insights from “Project X: Algebra Engineering Lab” in Baltimore City High Schools (Alexis Daniels, Center for Educational Outreach).

A welcome and brief introduction to the workshop, followed by the keynote on universal deep learning potentials and machine-learning-guided materials discovery.