Join

State of the field

AI for science, field by field

For each discipline: what already works, what has been shown to work once, what is still a promise, and the newest work of the last three weeks. The sorting is our judgement; every result links to its evidence.

Compiled from 26 sources. A new edition is compiled early every morning.

The map

Where each field stands

Select a result to read the evidence behind it, or a field to see what is new in it.

FieldIn everyday useShown to workStill a promise
Engineering, Earth and climateWeather forecastingFlood forecastingChip floorplanningNone yetPINNs vs solvers
Chemistry and drug discoveryNone yetRentosertib trialMolecular propertiesBinding affinity
MaterialsInteratomic potentialsSelf-driving labsGenerative materials
Physics and astronomyLHC triggersFusion plasma controlGravitational wavesNone yet
Biology and medicineAlphaFold DBCell segmentationEnzyme designVirtual cell
Interdisciplinary and methodsNone yetAI co-scientistsAlgorithm searchLLMs doing research

Field 1 of 6

Engineering, Earth and climate

Forecasting is the clearest success: learned weather and flood models now run in operational services. Replacing numerical solvers with physics-informed networks has not yet beaten strong baselines.

Where it stands

Data-driven weather forecasting

EngineeringIn production

Operational since Feb 2025

Europe's main forecast centre runs a machine-learned global model in its operational service.

Evidence and sources

Evidence

ECMWF's AIFS became the first data-driven model in an operational forecast service at an international centre in February 2025, with an ensemble version from July 2025. GraphCast beat the physics-based model on 90% of 1,380 verification targets, and GenCast on 97.2% of 1,320.

Still open

Certification. The Met Office's FastNet beats its global model on error metrics and is still not operational, because physical consistency has to be shown first.

At Cambridge

Aardvark Weather, the first forecast learned end to end from raw observations, was led by Richard Turner in Engineering.

Sources45678

Flood forecasting in ungauged rivers

EngineeringIn production

More than 80 countries

A learned model gives five-day flood warnings where no river gauge exists.

Evidence and sources

Evidence

An LSTM model forecasts extreme floods in ungauged basins with five-day forecasts as reliable as the global reference system's same-day nowcasts, and it runs operationally through Google's Flood Hub in more than 80 countries.

Still open

Extremes outside the training record, and forecasts that national agencies can audit.

Sources9

Chip floorplanning by reinforcement learning

EngineeringIn production

3 TPU generations

In production at one company, and still disputed as science.

Evidence and sources

Evidence

Google reports using its reinforcement-learning placement method for TPU v5e, v5p and Trillium. Independent work found that stronger simulated-annealing baselines compete, and that the original paper withheld inputs needed to reproduce it.

Still open

An open benchmark with released inputs. Being in production says nothing about whether a method beats a strong baseline.

Sources1819

Physics-informed networks replacing solvers

EngineeringPromised

79% of 76 papers: weak baselines

On forward problems, mature numerical methods still win.

Evidence and sources

Evidence

Physics-informed neural networks did not beat the finite element method in time or accuracy on standard problems (2023), and 79% of 76 machine-learning papers on PDEs compared against a weak baseline (2024).

What would move it up

Using them where they help, for inverse problems and data assimilation, and comparing forward solves against strong solvers.

At Cambridge

Carola-Bibiane Schönlieb's group in DAMTP ran the comparison with the finite element method.

Sources4647

Field 2 of 6

Chemistry and drug discovery

An AI-discovered drug has completed a randomised phase 2a trial, but on most property-prediction tasks classical baselines still win, and co-folding affinity predictions have not held up in independent tests.

Where it stands

An AI-discovered drug in a randomised trial

ChemistryDemonstrated

Phase 2a, n = 71

An AI-derived target and an AI-generated molecule, tested in patients for twelve weeks.

Evidence and sources

Evidence

Rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis, completed a randomised phase 2a trial. The primary endpoint was safety. Lung function rose 98.4 mL at the top dose against a 20.3 mL fall on placebo, which was a secondary endpoint.

What would make it routine

A larger, longer trial with efficacy as the primary endpoint. Across AI-native pipelines, phase I success is 80 to 90%, while phase II is about 40%, in line with history.

Sources2122

Molecular property prediction

ChemistryPromised

Random forest 68%, GNN 32%

On realistic small-data tasks, a classical baseline still wins.

Evidence and sources

Evidence

Across 58 industry-curated datasets, a random forest won 68% of Polaris tasks against 32% for Chemprop, the standard graph neural network. The 2025 deep-learning win came from pretraining on classical descriptors.

What would move it up

Reporting a descriptor baseline and a scaffold or time split by default.

At Cambridge

Jonathan Goodman's group in Chemistry works on chemical informatics and machine learning for structure assignment.

Sources4142

Binding affinity from co-folding models

ChemistryPromised

Not yet enough for leads

Developer benchmarks look strong, and an independent test on 38,000 compounds did not.

Evidence and sources

Evidence

Boltz-2's developers report affinity accuracy approaching free-energy calculations at a thousandth of the cost. An independent evaluation on 16,780 and 21,702 compounds found it lacks the energetic resolution needed to pick leads (2026).

What would move it up

Prospective, blinded evaluation run by someone other than the developers.

Sources4445

Field 3 of 6

Materials

Machine-learned interatomic potentials are routine tools. Self-driving labs work but still need trustworthy automated analysis, and generative discovery at scale is unproven.

Where it stands

Machine-learned interatomic potentials

MaterialsIn production

F1 0.93 on Matbench Discovery

Universal potentials are now an off-the-shelf tool for screening and molecular dynamics.

Evidence and sources

Evidence

Trained on about 1.5 million Materials Project configurations, MACE-MP-0 runs stable dynamics for solids, liquids, interfaces and a small protein out of the box. On Matbench Discovery, universal potentials are the strongest model class, and the best F1 rose from about 0.6 in 2023 to 0.93 in 2026.

Still open

Quantitative accuracy for defects, interfaces and glasses still needs fine-tuning on a handful of DFT points, and the benchmark itself is close to its ceiling.

At Cambridge

Gábor Csányi's group in Engineering built GAP, MACE and MACE-MP-0. Matbench Discovery was first-authored in the Cavendish.

Sources121314

Self-driving laboratories

MaterialsDemonstrated

36 of 40 confirmed

Robots run the experiments. Judging the results turned out to be the hard part.

Evidence and sources

Evidence

Liverpool's mobile robot ran 688 experiments in eight days and found photocatalysts six times more active (2020). The A-Lab's claimed syntheses were re-analysed in a 2026 Nature correction: 36 of 40 were confirmed, four were inconclusive, and one compound was removed because it had been in the training data.

What would make it routine

Trustworthy automated phase identification, so that a synthesis can be scored without a crystallographer.

At Cambridge

Ross King's Robot Scientist work is in Chemical Engineering and Biotechnology.

Sources2627

Generative materials discovery at scale

MaterialsPromised

736 made

Millions of candidate crystals, a few hundred made, and known phases found in disguise.

Evidence and sources

Evidence

GNoME predicted 2.2 million structures, about 381,000 of them stable, and 736 had been made independently. MatterGen's one synthesised compound was later matched to a known disordered phase already in its training set.

What would move it up

Predicting synthesisability and disorder, and not only stability at zero kelvin.

At Cambridge

Anthony Cheetham in Materials Science co-wrote the most cited critique. Chris Pickard's group works on structure searching.

Sources37383940

Field 4 of 6

Physics and astronomy

Neural networks run inside the LHC's live data path. Learned plasma control and second-scale gravitational-wave inference have each been shown, but are not yet routine.

Where it stands

Machine learning inside the LHC

PhysicsIn production

Under 100 ns per decision

Neural networks select and label collisions in the detectors' live data path.

Evidence and sources

Evidence

CMS has run an anomaly-detection autoencoder in its Level-1 trigger since May 2024, deciding in under 100 nanoseconds on an FPGA. ATLAS runs its transformer flavour tagger, GN2, on collision data, where it improves charm-jet rejection 3.5-fold at 70% b-efficiency.

Still open

Robustness to the gap between simulation and data as the High-Luminosity LHC raises collision rates five- to sevenfold.

At Cambridge

The Cavendish high-energy physics group works on ATLAS.

Sources1011

Learned control of fusion plasmas

PhysicsDemonstrated

19 coils, one policy

Reinforcement learning has shaped and stabilised real tokamak plasmas.

Evidence and sources

Evidence

One deep reinforcement-learning policy drove all 19 magnetic coils of the TCV tokamak and held snowflake and droplet configurations (2022). Another avoided tearing instabilities on DIII-D in ITER-like conditions (2024).

What would make it routine

Routine operation on a new machine. DeepMind and Commonwealth Fusion Systems plan this for SPARC with the differentiable TORAX simulator.

Sources232425

Gravitational-wave inference in a second

PhysicsDemonstrated

About 1 s per merger

Neural posterior estimation analyses a neutron-star merger fast enough to point telescopes.

Evidence and sources

Evidence

DINGO-BNS performs full inference on a binary neutron-star merger in about a second, against about an hour for the fastest conventional method, and localises it on the sky about 30% more precisely than low-latency approximations (2025).

What would make it routine

Adoption in the live alert pipeline, with coverage checks, since simulation-based posteriors can be over-confident.

At Cambridge

Will Handley's group works on nested sampling and Bayesian machine learning for cosmology.

Sources3233

Field 5 of 6

Biology and medicine

Structure prediction and image analysis are infrastructure, and designed enzymes have been demonstrated. Predicting how cells respond to perturbation still loses to linear baselines.

Where it stands

AlphaFold Protein Structure Database

BiologyIn production

3.4 M users

Predicted structures served to millions of people who are mostly not machine-learning researchers.

Evidence and sources

Evidence

The database holds more than 260 million predictions, including protein complexes since 2026, and has been used by over 3.4 million people in 190 countries. AlphaFold 2's accuracy in the blind CASP14 assessment made it credible at once.

Still open

Dynamics. One predicted structure is one frame of a movie. BioEmu samples equilibrium ensembles to about 1 kcal/mol, and current models still favour ground states.

At Cambridge

Hosted by EMBL-EBI on the Wellcome Genome Campus at Hinxton, just south of Cambridge.

Sources123

Cell segmentation and connectomics

BiologyIn production

139,255 neurons mapped

Learned segmentation is routine in microscopy and made whole-brain wiring maps possible.

Evidence and sources

Evidence

Cellpose segments cells across imaging types with no retraining or parameter tuning. The FlyWire wiring diagram of an adult fly brain, 139,255 neurons, exists because machine-learned segmentation made it tractable.

Still open

Scale. The MICrONS millimetre of mouse visual cortex already holds more than 200,000 cells and half a billion synapses.

At Cambridge

The MRC Laboratory of Molecular Biology develops cryo-EM reconstruction methods in the Scheres group.

Sources151617

De novo enzyme design

BiologyDemonstrated

41 of 41 active sites

Designed enzymes that work, found by testing fewer than 96 sequences.

Evidence and sources

Evidence

RFdiffusion2 scaffolded all 41 benchmark enzyme active sites, against 16 of 41 for earlier methods, and produced active enzymes for three catalytic mechanisms from fewer than 96 designs tested each.

What would make it routine

Use by groups outside the developers' labs, on targets they chose themselves.

Sources20

Predicting how cells respond to perturbation

BiologyPromised

Below linear baselines

The virtual cell is the field's hardest test, and deep models do not yet pass it.

Evidence and sources

Evidence

Five foundation models and two other deep models did not beat simple linear baselines on single or double gene perturbations (2025). Across 400 pretrained single-cell models and 6,400 experiments there was no clear data-scaling law (2026).

What would move it up

Benchmarks that can be failed. The 2026 Virtual Cell Challenge asks for zero-shot prediction in six unseen cell lines.

At Cambridge

Sarah Teichmann's group at the Cambridge Stem Cell Institute works on single-cell genomics with machine learning.

Sources343536

Field 6 of 6

Interdisciplinary and methods

Search with machine-checkable answers, in algorithms and mathematics, is closest to routine. AI co-scientists have early bench results; language models on research-level problems remain far from reliable.

Where it stands

AI co-scientists

Cross-cuttingDemonstrated

Two Nature papers, back to back

Agent systems proposed hypotheses that then held up at the bench.

Evidence and sources

Evidence

Google's AI co-scientist and FutureHouse's Robin were published back to back in Nature in 2026. Robin proposed ripasudil for dry macular degeneration, and the co-scientist proposed drug-repurposing candidates for leukaemia. Both held up in cell experiments that humans ran.

What would make it routine

Independent replication, and evaluation that the developers do not run.

At Cambridge

Boris Bolliet's Denario applies multi-agent systems to cosmology in the Cavendish.

Sources282930

Algorithms found by search

Cross-cuttingDemonstrated

48 multiplications

A language-model search found a faster matrix product that anyone can check.

Evidence and sources

Evidence

AlphaEvolve multiplies 4×4 complex matrices with 48 scalar multiplications, the first improvement on Strassen's 49 in 56 years, and improved about a fifth of the more than 50 open problems it was tried on.

What would make it routine

Where answers are machine-checkable this is close to routine. The open question is how far beyond maths and code it reaches.

At Cambridge

Challenger Mishra in Computer Science works on AI-driven mathematical discovery.

Sources31

Language models doing research

Cross-cuttingPromised

5.7% on CritPt

Olympiad gold, while research-level physics problems remain mostly unsolved.

Evidence and sources

Evidence

On CritPt, 71 research problems written by more than 50 physicists, the best model scored 5.7%, or about 10% with tools. Frontier models are reported by their developers at 87.7% on olympiad physics theory.

What would move it up

Verifiers for open-ended research questions, which do not exist yet.

Sources43

Why things climb

Three things decide how fast a result climbs. The model is rarely the bottleneck.

  1. Verifier

    A cheap verifier

    Something that tells you an answer is right without a year of lab work: a proof checker, a compiler, tomorrow's weather.

    Where one exists, progress is fast and uncontested. Maths, code and weather forecasting all have one.

  2. Benchmark

    A blind benchmark

    Held out, run by someone else, and possible to fail. CASP's blind, biennial format is why AlphaFold was believed at once.

    Where benchmarks leak, scores rise and nothing ships. Leakage errors have been found in 329 papers across 17 fields.

  3. Loop

    A short feedback loop

    The time between a prediction and finding out it was wrong: seconds for a compiler, days for a forecast, months for a synthesis, years for a trial.

    This is why materials and drug discovery stall at demonstrated despite strong models.

How long until you find out you were wrongOrders of magnitude, for illustration
Compiler or proof checkerseconds
Weather forecastdays
Materials synthesisweeks to months
Clinical trialyears

Sources248

Before you believe it

Six questions to ask of any AI-for-science claim

Ask them at the next talk you go to, including ours.

  1. Q1

    What was the baseline?

    Without a strong classical baseline the result is unevaluated. Look for a random forest, a linear model or the finite element method.

  2. Q2

    Who held out the test set?

    Blind and external, or the authors' own split? Near-duplicate molecules and crystals make random splits meaningless.

  3. Q3

    How many things were actually made?

    Candidate counts cost nothing. Ask how many were synthesised, measured or dosed.

  4. Q4

    Who checks the answer, and how fast?

    A compiler, a proof checker and a forecast verify themselves. A synthesis takes months. The gap predicts the level.

  5. Q5

    Is the number in the paper?

    Some widely repeated speed-ups do not appear in the papers they are credited to, and developer-reported numbers are sometimes contradicted by independent tests.

  6. Q6

    What would failure have looked like?

    If no outcome would have counted as failure, it was a demonstration, not an experiment.

Sources47483745

Sources

References

Every number on this page comes from one of these.

  1. 1

    EMBL-EBI. AlphaFold Protein Structure Database: complexes added, 16 Mar 2026 (3.4 million users in 190 countries); more than 260 million predictions, 24 Sep 2026. alphafold.ebi.ac.uk

    Cited in AlphaFold DB

  2. 2

    Jumper et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2

    Cited in AlphaFold DB, Why things climb

  3. 3

    Lewis et al. Scalable emulation of protein equilibrium ensembles with generative deep learning (BioEmu). Science 389, eadv9817 (2025). doi:10.1126/science.adv9817

    Cited in AlphaFold DB

  4. 4

    ECMWF. AIFS Single operational from 25 Feb 2025; AIFS ENS operational from 1 Jul 2025. www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ensemble-ai-forecasts-become-operational

    Cited in Weather forecasting

  5. 5

    Lam et al. Learning skillful medium-range global weather forecasting (GraphCast). Science 382, 1416–1421 (2023). doi:10.1126/science.adi2336

    Cited in Weather forecasting

  6. 6

    Price et al. Probabilistic weather forecasting with machine learning (GenCast). Nature 637, 84–90 (2025). doi:10.1038/s41586-024-08252-9

    Cited in Weather forecasting

  7. 7

    Daub et al. FastNet, from the Alan Turing Institute and the Met Office. arXiv:2509.17658 (2025); Met Office news release, 8 Jul 2026. arxiv.org/abs/2509.17658

    Cited in Weather forecasting

  8. 8

    Allen et al. End-to-end data-driven weather prediction (Aardvark Weather). Nature 641, 1172–1179 (2025). doi:10.1038/s41586-025-08897-0

    Cited in Weather forecasting

  9. 9

    Nearing et al. Global prediction of extreme floods in ungauged watersheds. Nature 627, 559–563 (2024). doi:10.1038/s41586-024-07145-1

    Cited in Flood forecasting

  10. 10

    CMS Collaboration. AXOL1TL anomaly-detection autoencoder in the Level-1 trigger menu from May 2024; test-crate study, arXiv:2312.10009. arxiv.org/abs/2312.10009

    Cited in LHC triggers

  11. 11

    ATLAS Collaboration. Transforming jet flavour tagging at ATLAS (GN2). Nature Communications 17, 541 (2026). doi:10.1038/s41467-025-65059-6

    Cited in LHC triggers

  12. 12

    Batatia, Kovács, Simm, Ortner, Csányi. MACE: higher order equivariant message passing neural networks for fast and accurate force fields. NeurIPS 35 (2022). arxiv.org/abs/2206.07697

    Cited in Interatomic potentials

  13. 13

    Batatia et al. A foundation model for atomistic materials chemistry (MACE-MP-0). arXiv:2401.00096; J. Chem. Phys. 163, 184110 (2025). doi:10.1063/5.0297006

    Cited in Interatomic potentials

  14. 14

    Riebesell et al. A framework to evaluate machine learning crystal stability predictions (Matbench Discovery). Nature Machine Intelligence 7, 836–847 (2025); live leaderboard, Oct 2026. doi:10.1038/s42256-025-01055-1

    Cited in Interatomic potentials

  15. 15

    Stringer, Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods 18, 100–106 (2021). doi:10.1038/s41592-020-01018-x

    Cited in Cell segmentation

  16. 16

    Dorkenwald et al. and the FlyWire Consortium. Neuronal wiring diagram of an adult brain. Nature 634, 124–138 (2024). doi:10.1038/s41586-024-07558-y

    Cited in Cell segmentation

  17. 17

    MICrONS Consortium. Functional connectomics spanning multiple areas of mouse visual cortex. Nature 640, 435–447 (2025). doi:10.1038/s41586-025-08790-w

    Cited in Cell segmentation

  18. 18

    Mirhoseini et al. A graph placement methodology for fast chip design. Nature 594, 207–212 (2021); addendum, Nature 634, E10 (2024). doi:10.1038/s41586-021-03544-w

    Cited in Chip floorplanning

  19. 19

    Cheng, Kahng et al. An updated assessment of reinforcement learning for macro placement. arXiv:2302.11014; Markov. The false dawn. arXiv:2306.09633. arxiv.org/abs/2302.11014

    Cited in Chip floorplanning

  20. 20

    Ahern et al. Atom-level enzyme active site scaffolding using RFdiffusion2. Nature Methods 23, 96–105 (2026). doi:10.1038/s41592-025-02975-x

    Cited in Enzyme design

  21. 21

    Xu et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine 31, 2602–2610 (2025). doi:10.1038/s41591-025-03743-2

    Cited in Rentosertib trial

  22. 22

    Jayatunga et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today 29, 104009 (2024). doi:10.1016/j.drudis.2024.104009

    Cited in Rentosertib trial

  23. 23

    Degrave et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602, 414–419 (2022). doi:10.1038/s41586-021-04301-9

    Cited in Fusion plasma control

  24. 24

    Seo et al. Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature 626, 746–751 (2024). doi:10.1038/s41586-024-07024-9

    Cited in Fusion plasma control

  25. 25

    Google DeepMind. Bringing AI to the next generation of fusion energy, with Commonwealth Fusion Systems, 16 Oct 2025. deepmind.google/blog/bringing-ai-to-the-next-generation-of-fusion-energy

    Cited in Fusion plasma control

  26. 26

    Burger et al. A mobile robotic chemist. Nature 583, 237–241 (2020). doi:10.1038/s41586-020-2442-2

    Cited in Self-driving labs

  27. 27

    Szymanski et al. An autonomous laboratory for the accelerated synthesis of inorganic materials (A-Lab). Nature 624, 86–91 (2023); Author Correction, Nature 650, E1 (2026). doi:10.1038/s41586-025-09992-y

    Cited in Self-driving labs

  28. 28

    Gottweis et al. Accelerating scientific discovery with Co-Scientist. Nature 655, 487–496 (2026). doi:10.1038/s41586-026-10644-y

    Cited in AI co-scientists

  29. 29

    Ghareeb et al. A multi-agent system for automating scientific discovery (Robin). Nature 655, 497–505 (2026). doi:10.1038/s41586-026-10652-y

    Cited in AI co-scientists

  30. 30

    Villaescusa-Navarro, Bolliet et al. Denario, a modular multi-agent research assistant. arXiv:2510.26887 (2025). arxiv.org/abs/2510.26887

    Cited in AI co-scientists

  31. 31

    Novikov et al. AlphaEvolve: a coding agent for scientific and algorithmic discovery. arXiv:2506.13131 (2025). arxiv.org/abs/2506.13131

    Cited in Algorithm search

  32. 32

    Dax et al. Real-time inference for binary neutron star mergers using machine learning (DINGO-BNS). Nature 639, 49–53 (2025). doi:10.1038/s41586-025-08593-z

    Cited in Gravitational waves

  33. 33

    Hermans et al. A trust crisis in simulation-based inference? Your posterior approximations can be unfaithful. TMLR (2022). arxiv.org/abs/2110.06581

    Cited in Gravitational waves

  34. 34

    Ahlmann-Eltze, Huber, Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods 22, 1657–1661 (2025). doi:10.1038/s41592-025-02772-6

    Cited in Virtual cell

  35. 35

    DenAdel et al. Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance. Nature Methods 23, 1447–1457 (2026). doi:10.1038/s41592-026-03120-y

    Cited in Virtual cell

  36. 36

    Adduri et al. Virtual Cell Challenge 2026: benchmarking zero-shot generalization across cellular contexts. Cell 189, 5827–5830 (2026). doi:10.1016/j.cell.2026.08.004

    Cited in Virtual cell

  37. 37

    Merchant et al. Scaling deep learning for materials discovery (GNoME). Nature 624, 80–85 (2023). doi:10.1038/s41586-023-06735-9

    Cited in Generative materials, Six questions

  38. 38

    Cheetham, Seshadri. Artificial intelligence driving materials discovery? Perspective on the article: Scaling deep learning for materials discovery. Chemistry of Materials 36, 3490–3495 (2024). doi:10.1021/acs.chemmater.4c00643

    Cited in Generative materials

  39. 39

    Zeni et al. A generative model for inorganic materials design (MatterGen). Nature 639, 624–632 (2025). doi:10.1038/s41586-025-08628-5

    Cited in Generative materials

  40. 40

    Juelsholt. Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset. Materials Horizons 13, 5672–5679 (2026). doi:10.1039/d6mh00268d

    Cited in Generative materials

  41. 41

    Burns et al. Deep learning foundation models from classical molecular descriptors (CheMeleon). arXiv:2506.15792 (2025). arxiv.org/abs/2506.15792

    Cited in Molecular properties

  42. 42

    Wognum et al. A call for an industry-led initiative to critically assess machine learning for real-world drug discovery (Polaris). Nature Machine Intelligence 6, 1120–1121 (2024). doi:10.1038/s42256-024-00911-w

    Cited in Molecular properties

  43. 43

    Zhu, Tian, Yang et al. CritPt, a benchmark of 71 research-level physics challenges. arXiv:2509.26574 (2025). arxiv.org/abs/2509.26574

    Cited in LLMs doing research

  44. 44

    Passaro et al. Boltz-2: towards accurate and efficient binding affinity prediction. bioRxiv (2025). doi:10.1101/2025.06.14.659707

    Cited in Binding affinity

  45. 45

    Wan et al. Reliability of AI methods in drug discovery: evaluation of Boltz-2 for structure and binding affinity prediction. J. Chem. Theory Comput. 22, 7811–7824 (2026). doi:10.1021/acs.jctc.6c01334

    Cited in Binding affinity, Six questions

  46. 46

    Grossmann, Komorowska, Latz, Schönlieb. Can physics-informed neural networks beat the finite element method? arXiv:2302.04107 (2023). arxiv.org/abs/2302.04107

    Cited in PINNs vs solvers

  47. 47

    McGreivy, Hakim. Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations. Nature Machine Intelligence 6, 1256–1269 (2024). doi:10.1038/s42256-024-00897-5

    Cited in PINNs vs solvers, Six questions

  48. 48

    Kapoor, Narayanan. Leakage and the reproducibility crisis in machine-learning-based science. Patterns 4, 100804 (2023). doi:10.1016/j.patter.2023.100804

    Cited in Why things climb, Six questions