Physicist Nathaniel Craig has proposed a possible definition of the Standard Model on a lattice, a grid of points replacing continuous space; he credits OpenAI’s GPT-6 Astra AI model with a leading role in the proof
Physicist Nathaniel Craig has proposed a possible definition of the Standard Model on a lattice, a grid of points replacing continuous space; he credits OpenAI’s GPT-6 Astra AI model with a leading role in the proof
In a preprint dated October 6, Craig, a visiting scientist at OpenAI, constructed a measure on a four-dimensional lattice: a rule for summing contributions from particles. On November 15, he described the work on the OpenAI blog.
Standard Model predictions are usually calculated as small corrections to a calculation in which particles do not interact. This approach cannot yield the proton’s mass, which arises from interactions among the particles inside it. Physicists therefore replace spacetime with a lattice, on which the theory can be rigorously defined and calculations performed on a computer.
Electrons and quarks have left-handed and right-handed versions, mirror images of each other like a pair of gloves, and the weak interaction distinguishes between them. The most natural lattice rules give a left-handed particle a right-handed partner, one “you didn’t ask for,” as Craig puts it, producing the wrong set of particles. A solution developed in the 1990s removes these partners but leaves a choice of phase, a multiplicative factor in the summation rule. After any closed cycle of changes to the force fields, this factor must compensate for the phase accumulated by the particle states. Otherwise, the result would depend on the path taken as the fields changed.
Physicist Martin Lüscher found a phase for a version of electromagnetism that distinguishes left from right. He continuously reduced the field to zero while keeping it admissible, with a small field strength on every elementary square of the lattice. Fields of the weak and strong interactions can leave the admissible region during this process, and a phase for the full set of forces had remained elusive for more than 25 years.
Craig proposed representing the same fields on progressively finer lattices. Comparing particle states on successive lattices gives part of the phase correction. When the contributions are summed, the intermediate phases cancel, leaving the phase accumulated on the original lattice divided by the phase on the finest lattice, which approaches unity. The sum converges because the leading terms cancel among the particles in one generation of the Standard Model.
Craig had been unable to prove that the method worked in every required case. According to his account, Astra proposed how to transfer the fields to finer lattices and showed that the comparisons on those lattices converge. Three specialists in lattice field theory read the proof. The result is a theorem: for admissible fields on a finite lattice, the summation rule is smooth and independent of how the fields are described.
Craig and his colleagues proposed checks of the method, and Astra carried them out. One involved writing the proof in Lean, a program that checks each step under the stated assumptions. It accepted all 16 modules without errors. AI agents checked that the formalization matched the written proof. The preprint tests the physical validity of the method through numerical calculations in a two-dimensional analogue and a second, independent construction of the measure.
The method, Craig writes, is “much more an answer in principle than an answer in practice.”
Columbia University mathematician Peter Woit called the result a modest advance on Lüscher’s work. Woit notes that, to his knowledge, Craig had never previously published papers on lattice gauge theories, while Astra is taking the role of a mathematician coauthor. He asks whether other physicists and mathematicians will try using AI to work beyond their own specialties. According to Woit, the line of research Craig is pursuing began in the 1980s, was developed by a small number of researchers in the 1990s, and was then largely abandoned as unpromising:
“Directing the power of AI toward advancing a complicated proposal that stalled for good reason is unlikely to lead to anything interesting.”