Foundation Models &
Agentic Methods for
Astrophysical Simulation

This talk will introduce knowledge-guided machine learning (KGML), a rapidly growing area of research in AI/ML where scientific knowledge is deeply integrated in machine learning frameworks to produce scientifically grounded, interpretable, and generalizable solutions even on out-of-distribution data. This talk will present the landscape of research in KGML by characterizing previous research in this area in terms of the nature and format of scientific knowledge used, the form of knowledge-ML integration explored, and the method for incorporating scientific knowledge in ML for diverse scientific use-cases. The talk will illustrate KGML methodologies in the context of a variety of use cases spanning multiple scientific disciplines including aquatic sciences, biodiversity science, agriculture, virology, fluid dynamics, and physics. This includes modeling the quality of water in lakes across the US and discovering novel biological traits of organisms linked with evolution from biodiversity images. The talk will conclude with a discussion of emerging opportunities in KGML especially in the light of recent advances in generative AI, Foundation models, and agentic AI with potential applications in a broad range of scientific disciplines.

Next generation transient discovery surveys require accurate real-time analysis of evolving transients for followup prioritization from other surveys. SELDON (Supernova Explosions Learned by Deep ODE Networks) is a transient foundation model designed with this capability in mind, and more. Our model not only provides classifications and forecasts for the future evolution of transient light-curves – providing estimates for optimal follow-up observing windows – but also approximates the underlying spectral energy distribution, redshift, and even some characteristics of the observing instrument itself. The architecture itself offers a flexible approach to multi-modal astrophysical data analysis, with physically motivated decoding built upon both intrinsic physical properties of transient evolution as well as the physical engineering behind the observing instrument itself, and time-invariant latent representations of transients that can be adapted for downstream tasks.

The name "foundation model" has become widespread. In astrophysics it usually means unsupervised representation learning at scale: models like AION learn structured latent spaces from heterogeneous survey data, enabling similarity search for rare objects, outlier detection. Outside of astrophysics, the word may mean something stronger: a language model can be pointed at almost any task, reasoning about it, writing code, acting over long horizons. In this talk I will present two separate projects on the trouble foundation models have telling what they know from what they have absorbed.

The first addresses the fact that a data-driven representation model encodes instrument systematics in the same space as the physics. We train a dual-encoder architecture with a counterfactual generation objective, learning what a galaxy would look like through another instrument. On cross-matched HSC and DESI Legacy images, we show how this separates intrinsic properties from instrumental distortions, and its effect on downstream tasks.

The second project was designed to answer the question: can a language model discover physics it was never taught? Static question-and-answer benchmarks are especially poorly suited to this: they reward retrieving a known answer, whereas the scientific method resembles more an interactive loop of proposing a hypothesis, designing an experiment to test it, and revising it. We built DiscoverPhysics to evaluate that loop directly. It presents an agent with simulated worlds whose laws deviate from our own, and the agent must run experiments, observe raw trajectories, and submit both an explanation and an implementation of the law it inferred. The strongest models pass only half the worlds, and predictive accuracy turns out to be a poor proxy for understanding. I will close with early results on turning DiscoverPhysics into a reinforcement learning environment, where the simulator provides verifiable rewards (RLVR) for training models to reason about physics.

Machine-learning surrogates for fluid dynamics are typically trained for a particular physical regime, discretization, and sampling protocol. VICON is a vision in-context operator network that learns across multiple fluid systems with a single set of weights. At inference time, a small number of example input-output pairs from a new system are provided as context, allowing VICON to implicitly identify the system dynamics and forecast its evolution without fine-tuning. The model uses patch-wise tokenization of two-dimensional fields and is trained jointly across incompressible Navier-Stokes and two compressible-flow regimes, including an effectively inviscid regime closely related to the Euler systems used in astrophysical hydrodynamics. This in-context formulation also accommodates varying timestep strides and missing frames without retraining or interpolation, with substantially less accuracy degradation than fixed-timestep baselines.

I will close with recent work extending this direction from learned simulation models to scientific agents: Foam-Agent, which automates OpenFOAM workflows from natural-language specifications, and SimulCost, which studies the computational cost of agents tuning simulation parameters.

Science advances by formulating and testing hypotheses, collecting and analyzing data, and drawing conclusions—yet much of a scientist’s time is spent in tasks such as coding analyses, writing and revising text, reviewing the literature, and learning new concepts. Can recent advances in AI help reclaim some of that time? In this seminar, I will show how large language models and AI agents may help scientists with these tasks. I will first describe what AI agents are and discuss their applications in science. Next, I will present Denario, a complex, publicly available, multi-AI-agent system designed to function as a research assistant. Developed and evaluated by a diverse team of scientists, mathematicians, and philosophers, Denario is an interdisciplinary tool capable of generating ideas, searching the literature, developing research plans, writing and executing code, crafting plots, and drafting and reviewing scientific papers. I will then show how a swarm of Denario agents can operate in parallel: developing scientific ideas, formulating hypotheses, designing research plans, writing and running code, analyzing results, generating plots, drafting papers, and producing referee-style evaluations that guide subsequent iterations. This swarm-based workflow allows many research directions to be explored simultaneously, with each agent refining its own line of inquiry while contributing to a broader, structured map of ideas, results, and evaluations.

I will also present and discuss some of the papers already generated by Denario in disciplines such as astrophysics, biology, biophysics, biomedical informatics, chemistry, machine learning, material science, mathematical physics, medicine, neuroscience, planetary physics, and quantum physics. I will close by discussing how tools like Denario may help researchers accelerate scientific research and invite the audience to a broad discussion on the benefits and risks of this technology.

AI agents that autonomously conduct research risk breaking a delicate balance in science. The scientific paper is a lossy format: methodological choices and assumptions compressed into a few pages, with peer review and community scrutiny providing the warrant that results are sound. This trade-off worked because the rate of new work stayed roughly matched to the community's capacity to scrutinize it. Agents now threaten that balance, accelerating production far faster than scrutiny can keep up and deepening an already well-documented trust and reproducibility crisis. But the same technology is also an opportunity: agents can document exactly how results are produced at a level of detail that was previously too onerous to maintain by hand, if we build the infrastructure to capture it. Closing that gap is what we're building toward at Lightcone Research, a new initiative between UC Berkeley and the French National Centre for Scientific Research (CNRS). We are building a new standard ensuring that conforming analyses are not only reproducible by default, but also inspectable, verifiable, and extendable — every result traceable to the assumptions that produced it, decisions substitutable without rewriting entire pipelines, and analyses able to cite and build upon one another. Done right, this is what allows AI-powered science to remain trustworthy and continue to compound. Because infrastructure of this kind has to be built with and for the community that relies on it, our project is open-source, collaborative, and designed for shared stewardship from the start. In this talk, I will survey the current state of agentic AI in science and the pitfalls it is exposing, introduce the Lightcone Research project and its agentic research stack, as well as share lessons from working with UC Berkeley early testers and reproducing flagship cosmology results.