Explore the research

Physics

Can AI move beyond rediscovering known rules to make predictions that new experiments confirm?

18%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Theoretical novelty20%
Simulation efficiency40%
Novel prediction20%
Experimental confirmation0%
Experiment efficiency25%
Equation recovery0%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Equation recovery

0%estimated progress

What we track

Correct models / all held-out benchmark systems, with equivalent forms recognized and noise, missing variables, and budgets specified.

Why this estimate?

The quantified equation-recovery benchmark predates 2022, so it anchors the starting point. This outcome's baseline does not establish a comparable post-anchor increment.

Advancements

AI Feynman recovers all 100 benchmark physics equationsApril 15, 2020 revised preprint. AI Feynman recovered all 100 equations in its Feynman benchmark, versus 71 for the comparison method.
Φ-SO symbolic regression enforces physical units to recover textbook equationsMarch 6, 2023. The Φ-SO deep reinforcement learning framework generates only unit-consistent expressions and reports state-of-the-art recovery on Feynman Lectures and other textbook equations when noise exceeds 0.1%, remaining robust at 10% noise.
LLM-SR uses language models to discover scientific equations as programsApril 29, 2024. LLM-SR combines language-model hypotheses with evolutionary search and, on four benchmark problems designed to limit recitation, found physically accurate equations that outperformed symbolic regression baselines, especially out of domain. ICLR 2025 Oral.
LLM-SRBench: the best equation-discovery system reaches 31.5% symbolic accuracyApril 14, 2025. On a benchmark of 239 problems across four scientific domains built to prevent memorization, the best-performing system reached only 31.5% symbolic accuracy.

Read the dated findings and sources

Status: Known-law recovery benchmark. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

April 15, 2020 revised preprint: AI Feynman recovered all 100 equations in its Feynman benchmark, versus 71 for the comparison method. AI Feynman.

Interpretation and tracking value

This historical benchmark is not a current universal success rate under realistic missing variables and noise. Algebraic equivalence and benchmark familiarity matter.

Next comparable measurement

Freeze fresh systems, noise and missing-variable levels; compare equivalent equations under identical time/compute budgets and report all failures. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Experiment efficiency

25%estimated progress

What we track

Experiments, facility time, and total cost needed to resolve a prespecified scientific question.

Why this estimate?

Agents can explore simulated systems, but physical facility time, cost and independent resolution of scientific questions remain unmeasured.

Advancements

AI search finds gravitational-wave detector designs that outperform planned onesDecember 6, 2023 (arXiv); published in Physical Review X in 2025. An AI system searching detector configurations found new interferometer topologies that outperform current next-generation designs under realistic constraints; more than 50 solutions were released as a public Gravitational Wave Detector Zoo.
Princeton AI controller forecasts tokamak tearing instabilities 300 ms aheadMay 10, 2024 (article date; study published February 2024). A reinforcement-learning controller trained on past DIII-D experiments forecast tearing mode instabilities up to 300 milliseconds in advance, enough time to adjust operating parameters and avoid them.
Deep Loop Shaping cuts LIGO control noise by 30 to 100 timesSeptember 4, 2025. Google DeepMind reported in Science that an AI control method reduced noise in LIGO's most unstable feedback loop by 30 to 100 times compared with existing controllers, tested on the real system at the Livingston observatory.
SciExplorer agents investigate simulated physical systemsSeptember 29, 2025 preprint; 2026 journal publication. SciExplorer studies agents investigating simulated mechanical, wave and quantum many-body systems.
Sandia self-driving lab finds a light-steering method in 300 experiments29 January 2026. A self-driving lab at Sandia National Laboratories ran 300 experiments in about five hours and found a way to steer spontaneous light emission, with up to a four-fold gain in peak directivity across a 74 degree field of view; published in Nature Communications.

Read the dated findings and sources

Status: Simulated exploration progress; facility-cost metric unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 29, 2025 preprint; 2026 journal publication: SciExplorer studies agents investigating simulated mechanical, wave and quantum many-body systems. SciExplorer.

Interpretation and tracking value

Choosing informative simulated experiments is not measured reduction in physical facility time or dollars. Simulator access and human setup must be counted.

Next comparable measurement

Preregister a scientific question; compare physical experiment counts, beam/robot time and full cost to a matched expert design strategy. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Experimental confirmation

0%estimated progress

What we track

Predicted effects confirmed / all preregistered effects tested, with effect sizes, uncertainty, controls, and independent replication.

Why this estimate?

Agreement with a simulator does not supply the required independent physical confirmation. No qualifying confirmed-effect fraction is established in the baseline.

Advancements

External labs independently made 736 of GNoME's predicted materialsNovember 29, 2023. Google DeepMind reported that 736 of the crystal structures predicted by GNoME had been independently created in laboratories, and that Berkeley Lab's autonomous A-Lab synthesized more than 41 new materials using its stability predictions.
MatterGen-designed TaCr2O6 is synthesized; bulk modulus within 20% of targetJanuary 16, 2025. A material generated by Microsoft's MatterGen for a target bulk modulus of 200 GPa, TaCr2O6, was synthesized at the Shenzhen Institutes of Advanced Technology and measured at 169 GPa, a relative error below 20%; the structure matched apart from Ta/Cr compositional disorder.
SciExplorer is evaluated on simulated physical systemsSeptember 29, 2025 preprint; 2026 journal publication. SciExplorer's evaluated environments include simulated physical systems.

Read the dated findings and sources

Status: Simulation evidence; independent physical confirmation unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 29, 2025 preprint; 2026 journal publication: SciExplorer's evaluated environments include simulated physical systems. SciExplorer.

Interpretation and tracking value

Simulator agreement cannot be counted as independent experimental confirmation of a new physical effect. No qualifying confirmed/all-preregistered-effects fraction is extracted.

Next comparable measurement

Register all candidate effects, including failures; commission independent controlled measurements and report effect sizes and uncertainty. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Novel prediction

20%estimated progress

What we track

Quantitative predictions registered before data collection that distinguish the proposed model from the strongest established alternatives.

Why this estimate?

Simulation-based hypothesis generation is a precursor, but predictions novel to physics and registered before new physical observations are not established.

Advancements

LLNL's CogSim model gave the NIF ignition shot a 50.2% chance beforehandJune 21, 2023 (article date; shot fired December 5, 2022). Before the December 5, 2022 shot, Lawrence Livermore's machine-learning cognitive simulation model gave the design a 50.2% chance of ignition, against 17.3% for the previous record shot. The shot produced 3.15 MJ from 2.05 MJ of laser energy, consistent with the predicted yield distribution.
GNoME predicts 2.2 million new crystals, 380,000 of them stableNovember 29, 2023. Google DeepMind's GNoME graph network predicted 2.2 million new crystal structures, of which 380,000 are the most stable, raising the number of known stable materials to 421,000.
SciExplorer explores systems unknown to the agentSeptember 29, 2025 preprint; 2026 journal publication. SciExplorer demonstrates exploration of systems initially unknown to the agent.

Read the dated findings and sources

Status: Hypothesis generation in simulation; prospective novelty baseline unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 29, 2025 preprint; 2026 journal publication: SciExplorer demonstrates exploration of systems initially unknown to the agent. SciExplorer.

Interpretation and tracking value

Unknown to an agent does not mean unknown to physics. A model fitted to simulator outputs does not establish a registered new prediction tested on future physical data.

Next comparable measurement

Timestamp discriminating quantitative predictions before collection; identify strongest alternative models and preregister rejection thresholds. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Simulation efficiency

40%estimated progress

What we track

Time, energy, and dollars per solution at a fixed independently checked error tolerance, including training and setup costs.

Why this estimate?

Fast atmospheric simulation demonstrates strong bounded execution and testable output. Unmatched error tolerances and excluded full costs limit broader autonomous efficiency claims.

Advancements

GraphCast makes 10-day global forecasts in under a minuteNovember 14, 2023. Google DeepMind's GraphCast produces a 10-day forecast in under one minute on a single Google TPU v4 machine and beat ECMWF's HRES on more than 90% of 1,380 test variables; an HRES forecast can take hours on a supercomputer.
NeuralGCM simulates 70,000 days in 24 hoursJuly 22, 2024. NeuralGCM-1.4° simulated 70,000 days in 24 hours on one TPU; the cited X-SHiELD comparison simulated 19 days using 13,824 CPU cores.
GenCast beats the ECMWF ensemble on 97.2% of targetsDecember 4, 2024. Google DeepMind's GenCast outperformed ECMWF's ENS on 97.2% of 1,320 targets, and 99.8% at lead times over 36 hours; one 15-day forecast in its ensemble takes about 8 minutes on a single Google Cloud TPU v5.
Microsoft's Skala functional reaches hybrid-level accuracy at semi-local DFT costJune 17, 2025. Skala, a deep-learning exchange-correlation functional, reports an error of 2.8 kcal/mol on the GMTKN55 main-group chemistry benchmark, outperforming state-of-the-art hybrid functionals while keeping the lower computational cost of semi-local density functional theory.

Read the dated findings and sources

Status: Measured throughput; equal-error comparison incomplete. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

July 22, 2024: NeuralGCM-1.4° simulated 70,000 days in 24 hours on one TPU; the cited X-SHiELD comparison simulated 19 days using 13,824 CPU cores. Neural general circulation models, Nature.

Interpretation and tracking value

Different resolutions, hardware and model formulations prevent interpreting this as a universal fixed-error speedup. Training, initialization and energy are excluded from that throughput figure.

Next comparable measurement

Compare at a common physical error tolerance; include training/setup amortization, energy and full cost per accepted solution. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Theoretical novelty

20%estimated progress

What we track

Independent expert classification of genuinely new explanatory content versus interpolation, reparameterization, or rediscovery.

Why this estimate?

Recent recovery of known laws demonstrates constructive exploration, but does not establish new explanatory theory or independent novelty evaluation.

Advancements

Emory machine-learning method identifies non-reciprocal forces in dusty plasmaJuly 30, 2025. Emory physicists reported in PNAS that a neural-network model trained on laboratory data from a dusty plasma identified new features of non-reciprocal forces between particles in a many-body system, using a method the authors describe as interpretable.
GPT-5.2 Pro conjectures a gluon amplitude formula later proved by an OpenAI model12 February 2026. The preprint 'Single-minus gluon tree amplitudes are nonzero' shows that amplitudes often assumed to vanish are nonzero in half-collinear configurations. It states that its key formula, Eq. (39), was first conjectured by GPT-5.2 Pro and then proved by a new internal OpenAI model, and was checked by hand using the Berends-Giele recursion.
NYU Abu Dhabi AI system rediscovers particle-physics symmetries from 1950s-60s dataMarch 31, 2026. Given particle data from the 1950s and 1960s and no knowledge of the period's mathematical tools, an AI system rediscovered baryon number, isospin and charm, the Eightfold Way classification, and Regge trajectories; published in the Journal of High Energy Physics. This is rediscovery of known physics.
AI-Newton recovers Newtonian mechanics and gravitationNovember 4, 2026. AI-Newton reports recovering established mechanics concepts and laws, including Newtonian dynamics and gravitation.

Read the dated findings and sources

Status: Recent rediscovery demonstration. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 4, 2026: AI-Newton reports recovering established mechanics concepts and laws, including Newtonian dynamics and gravitation. AI-Newton, Physical Review Research.

Interpretation and tracking value

Recovering known principles is useful but does not establish new explanatory content. No qualifying new-theory count is assigned from this paper.

Next comparable measurement

Require independent expert review, explicit comparison with known equivalent formulations and testable consequences beyond existing theory. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note