
Share
A new Stanford HAI conference wants to give "data physicists" and "phystatistians" a proper venue, tackling the weird statistical problems that come from building AI on physics data instead of images and text.
Physics has been doing big data before big data was cool. Decades of precision measurement, massive detector arrays, and simulation pipelines meant the field was running statistical inference at scale long before deep learning showed up. Now Stanford's Center for Decoding the Universe (C4DU) is launching PhyStat 2027, a conference built specifically for the people working at the seam between physics, statistics, and modern AI tooling.
The framing matters here. This isn't a general-purpose "AI for science" event. It's targeted at a specific and growing research niche: people who've had to adapt machine learning methods to datasets and problems that don't look like the standard ML benchmark suite. Think particle collision data, astrophysical surveys, or outputs from ab initio simulations, not neatly labeled images or scraped text corpora.
That distinction is doing a lot of work in the conference's pitch. Standard deep learning infrastructure, from convolutional nets to transformers, was largely developed and tuned against image and language data. Physics datasets often have different structure entirely: high-dimensional detector readouts, sparse event data, simulation outputs where you know the generative process but inference is still hard. The organizers point to the rise of simulation-based inference as a direct consequence, a class of methods for doing statistical inference when you can simulate your model but can't write down a tractable likelihood function. That's a real departure from how most ML research treats the inference problem, and it's a big part of why this community has started to coalesce around its own terminology and venues.
The organizing committee reads like a cross-section of the physics-statistics overlap: Benjamin Nachman and Risa Wechsler from Stanford and SLAC's C4DU, Louis Lyons and Olaf Behnke from the long-running CERN PhyStat effort, Michael Mahoney spanning Lawrence Berkeley's Scientific Data Division and UC Berkeley statistics, plus representatives from the American Statistical Association's sections on AI and physical/engineering sciences, and the American Astronomical Society's working group on astroinformatics and astrostatistics. It's a deliberately interdisciplinary lineup, pulling from particle physics, astronomy, and mainstream statistics departments rather than any single academic silo.
Structurally, PhyStat 2027 is modeled directly on the Machine Learning and Physical Sciences Workshop (MLPS) that's run at NeurIPS since 2018. That's a sensible template: MLPS has spent years proving there's a real audience for rigorous cross-disciplinary work that doesn't fit cleanly into either a physics journal or a mainstream ML venue. PhyStat is borrowing its peer review structure almost wholesale.
Here's how submissions will work:

That last point is worth flagging. Given how much of this research currently lives scattered across particle physics collaborations, astronomy working groups, and statistics departments, an open submission model could genuinely help surface work that wouldn't otherwise find a shared audience.
The conference sits under C4DU's broader mission, which is explicitly about pulling together researchers across scientific disciplines to tackle big questions about the universe using advanced statistical and computational methods. PhyStat 2027 is effectively a mechanism for formalizing a community that's already been meeting informally at places like the long-standing CERN PhyStat workshop series, now with a dedicated academic conference and a review process instead of ad hoc gatherings.
Specific dates, venue, and submission deadlines for PhyStat 2027 itself haven't been announced yet. Stanford HAI says details are forthcoming. What has been announced is a lead-in seminar series: an AI Measurement Science (AIMS) seminar running November 13, 2026, followed by a recurring Wednesday series in CoDa E160 through at least early December 2026, featuring a rotating cast of Stanford speakers. Think of these as the warm-up act, building momentum and community before the main conference.
The core bet here is that physics has genuinely different statistical problems than the ones mainstream ML research optimizes for, and that those problems deserve dedicated methodology, not just repurposed computer vision architectures. Simulation-based inference is the clearest example: when your forward model is a physics simulator rather than a labeled dataset, your whole approach to likelihoods, uncertainty quantification, and inference has to change.
For practitioners working on physics-adjacent AI problems, whether that's detector data, cosmological surveys, or materials simulations, this is worth tracking even before the full program drops. The organizing committee is heavy on people who've spent years building statistical tooling for large physics collaborations like CERN's PhyStat group, so the methodological bar is likely to be high. And because submissions are open to all, it's a lower-friction entry point than trying to publish in a physics-specific journal or squeeze a niche methods paper into a general ML conference.
Watch for the full call for papers and concrete dates, which Stanford HAI says are coming soon. If the MLPS comparison holds, expect this to become a recurring fixture rather than a one-off, especially if the current AIMS seminar series successfully builds a local community ahead of the 2027 event.
Tags
Original Sources
2027 Conference on Physics and AI (PhyStat 2027) | Stanford HAI
↗ https://hai.stanford.edu/events/2027-conference-on-physics-and-ai-phystat-2027
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
13 November 2026
14 articles
Related Articles

Tether Evo's New Research Tackles BCI's Calibration Problem with Cross-Subject Models
Models & Research · 5 min

Why Some Engineers Think the Ocean Floor Is the Future of Data Centers
Environment & Climate · 5 min

Europe's Digital ID Push Tests a New Privacy Tool: Zero-Knowledge Proofs
Policy & Regulation · 5 min
Related Articles

Tether Evo's New Research Tackles BCI's Calibration Problem with Cross-Subject Models
Models & Research · 5 min

Why Some Engineers Think the Ocean Floor Is the Future of Data Centers
Environment & Climate · 5 min

Europe's Digital ID Push Tests a New Privacy Tool: Zero-Knowledge Proofs
Policy & Regulation · 5 min
More Stories
© 2026 Cedar & Bloom. All rights reserved.