The Scientific Challenge Atlas
Mapping the machine-accessible frontier of science
We are seeing flashes of AI’s potential in Mathematics. Other fields, and the communities that support them, are still trying to understand the scale and character of what can be accomplished.
Our hypothesis is that a meaningful fraction of important, unsolved problems in science are limited by our ability to reason and compute. The Scientific Challenge Atlas develops a map of important open scientific problems that are machine-accessible (i.e. limited by Cognition & Computation alone). This question has not been addressed in an organized fashion.
Some entries may be solved by exhaustive synthesis. Others may require a new proof, model comparison, bound, or conceptual reframing.
What cognition & compute-limited means, operationally
Partition scientific problems into three regimes below. Most of the near-term yields will come from Regimes A and B, and combinations thereof. Note that existing data may be valuable.
Inside the atlas
Regime A is pure reasoning: the answer is latent in existing data and theory, and what’s missing is synthesis a human never had time for: an overlooked mechanism, a conjecture provable from known structure, a re-analysis of old data with better methods.
Regime B is compute-bound search over a trusted model: a validated forward model exists and the bottleneck is intelligent inverse design over a combinatorial space, where cognition decides what is worth computing and HPC or surrogates actually run the physics.
Outside the atlas
Regime C is measurement-limited: existing data underdetermines the answer, or the models aren’t trustworthy enough, and one genuinely need new, possibly expensive/time consuming experiments.
Why?
What the atlas would do
The Scientific Challenge Atlas is a public registry of important open scientific questions in the intelligence-reachable part of the landscape. Some entries may be solved by exhaustive synthesis. Others may require a new proof, model comparison, bound, or conceptual reframing. The common feature is that the needed information about the world is already available.
The atlas would:
Show the span of meaningful scientific advances that strong AI could plausibly make without first waiting for new experiments or observations.
Identify challenges whose solution may now be limited mainly by compute, search, and verification.
Estimate the rough reasoning, search, coding, and compute burden required for each advance.
Provide a public, disciplined way to evaluate the significance of AI for meaningful scientific advances.
Precedent
What the Erdős problems showed
This idea is inspired by the role that Erdős problems have played in mathematics. Paul Erdős posed many important problems in compact form. Thomas Bloom’s Erdős Problems database now lists 1,217 problems, with 555 marked solved as of July 24, 2026.
1,217
Problems now listed in Thomas Bloom’s Erdős Problems database
555
Marked solved as of July 24, 2026
9 / 353
Open Erdős problems resolved by a large-scale formal-proof effort, at a cost of a few hundred dollars per problem
Recent AI efforts have shown why this matters. On May 21, 2026, OpenAI announced an AI-assisted breakthrough on Erdős’s planar unit distance problem. Another large-scale formal-proof effort resolved 9 of 353 open Erdős problems at a cost of a few hundred dollars per problem. The broader lesson is that a good public problem list creates a focal point for the community to attack challenges. Science needs an analogous map.
Criteria for entry
The Form
A useful feature of present advances in AI in mathematics is that they do not require resolving difficult debates about topics like the nature of knowledge or the character of human creativity. They provide the question, the answer shape, and the verification criterion (a sound mathematical proof).
The atlas aims for the same simplicity. Problems that cannot be stated with those criteria, for example because there is no plausible verification path, are excluded from the atlas.
Submit a problem
Every atlas entry follows the same structure: a stated question, an answer shape, and a verification path. The LaTeX template lays out these sections so a proposed problem arrives in a form the editors can assess directly. Draft your problem in the template and send it to the address below.
Submit a Solution
Solved one of the open problems? A complete solution should follow the problem’s verification path so an independent group can reproduce and check it. Send your solution package: write-up, code, data, and any certificates, to the address below, naming the problem it answers.
Contributors
Collaborating institutions
The Scientific Challenge Atlas thus far has been developed in collaboration with the following institutions. If you or your institution would like to contribute, please contact us.
University of Michigan
Ann Arbor, Michigan
Open AI
San Francisco, California
Los Alamos National Laboratory
Los Alamos, New Mexico
Argonne National Laboratory
Lemont, Illnois