{"id":"35ff932b-eef4-46e3-8e59-fd928d3ff4c0","arxiv_id":"2607.23501","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"PINNs without initial conditions recover verifiable three-body periodic orbits from sparse noisy data, with training data—not init distribution—controlling which families emerge across seed ensembles.","lead":"Physics-informed neural nets trained on sparse noisy data, with no initial conditions, recover genuine periodic three-body orbits—including families never shown in training. The training data, not the weight-init distribution, steers which family appears; recovered states refine to catalog-matched solutions.","discovery_kind":"new_application","skeptic_critique":null,"referee_report":null,"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new result is real and narrow: train a PINN on 90 noisy samples with no initial conditions, and a non-trivial fraction of seeds land on genuine periodic three-body orbits—including families that were not in the data. They back this with 300 seeds, two χ² tests (data source moves the distribution, p<0.001, V=0.339; Glorot Uniform vs Normal does not, p=0.620), and external checks. The figure-eight recovered from Lagrange data matches Li–Liao I.A.1 to seven digits in T*; a BHH-type orbit from figure-eight data closes to δ_T ~ 10^{-9} after least-squares refinement. Code is public. That is more than most inverse-PINN demos deliver.\n\nWhat they do well: they separate seed from sampling distribution from data source instead of hand-waving “multi-modal loss.” Forward ablations (second-order residuals, fixed Fourier features, percentile RAR, trainable C) are honest and tabled. Euler/Lagrange T* hit the exact central-configuration formulas. They flag their own limits—no landscape theory, equal-mass planar only, PINNs still slow on ordinary IVPs.\n\nSoft spots, in proportion. BHH and “other” IDs lean on T* (plus geometry) after refinement; the authors keep the “-like” label and say continuation was not done. If T* collides, some cross-family percentages are mis-sorted. That weakens the exact 23–25% and the χ² contrast a bit, but it does not erase the two tightly verified case studies or the fact that refined states close. Only two init families and two data sources were swept; loss weights and noise level are free parameters. None of this looks load-bearing against the main claim.\n\nMath and citations are fine—Newton, classical reductions, Zou et al., Li–Liao, Broucke are used correctly; no circular self-validation. Who it is for: people who need candidate periodic ICs when they have sparse observations and no good guess, and PINN/scientific-ML readers interested in multi-basin inverse problems. Not a dynamical-systems foundations paper.\n\nI would send it to referees. Worth engaging if you work on inverse celestial mechanics or multi-solution PINNs; skim the case studies and the χ² tables if you only need the headline.","headline":"PINNs without ICs really do spit out catalog-checkable three-body periodic families from sparse noisy data, with a clean χ² split showing training data (not Glorot U vs N) shifts the family mix.","tokens_in":23327,"tokens_out":598,"would_cite":true,"duration_ms":10821,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Physics-informed neural networks recover periodic three-body orbits from sparse noisy data with no initial conditions, including families never seen in training.","keywords":["physics-informed neural networks","inverse problems","unknown initial conditions","celestial mechanics","periodic orbit families","multi-modal loss landscape","three-body problem"],"falsifier":"Re-run the same seed ensembles and classify every refined orbit by full numerical continuation into known families; if cross-family rates collapse or the chi-squared contrast between training-data sources disappears, the claim that data steers recovery of distinct genuine families fails.","tokens_in":23449,"feed_emoji":"🪐","tokens_out":924,"duration_ms":35634,"temperature":0.7,"pith_summary":"Finding periodic orbits in the chaotic three-body problem usually demands an initial guess already close to the target. This paper shows that a physics-informed neural network, given only sparse noisy observations and no initial conditions, can still recover genuine periodic orbits—and that roughly one run in four lands on a family that was not in the training data. Controlled seed ensembles and chi-squared tests separate two influences: which family a single run finds is seed-dependent, but the aggregate mix of families shifts when the training data source changes and does not shift when the weight-initialization distribution is swapped. Recovered states refine to true periodic solutions that match published catalogs, including the figure-eight choreography from Lagrange data and a Broucke–Hadjidemetriou–Hénon orbit from figure-eight data. A sympathetic reader cares because the method supplies candidate initial conditions in regions where no good guess exists to seed ordinary continuation or gradient search.","feed_headline":"PINNs find hidden three-body orbits without initial guesses","feed_subtitle":"About one run in four recovers a family absent from training; the data, not the init scheme, steers which","key_machinery":"The under-determined inverse PINN loss—ODE residual plus data fit, with initial conditions removed—whose multi-modal landscape lets different random seeds settle in different orbit-family basins. Enabling pieces, each validated on forward problems, are a second-order ODE residual, fixed-frequency Fourier features, percentile-based adaptive collocation, and a trainable residual scaling parameter C.","core_discovery":"PINNs trained on sparse, noisy observations without initial conditions recover periodic orbits of the gravitational three-body problem, including families absent from the training data. Across two 100-seed ensembles, 23–25% of runs converge to such families. Changing the training data source significantly shifts the distribution of recovered families, whereas switching between Glorot Uniform and Glorot Normal does not. The recovered orbits are verifiable: refined states close as genuine periodic solutions and match catalogued families, including the figure-eight (Li–Liao I.A.1) from Lagrange data to seven digits in the scale-invariant period T*.","pith_inferences":["The same data-shaped multi-basin pattern may appear in other multi-stable oscillators; the Duffing system, flagged as an open test by the authors, is the natural next check.","If mode-connectivity or Hessian probes show separated basins whose relative volumes track the training data, the statistical picture would gain a geometric basis the paper leaves open.","Moderate observation noise may act as a regularizer that keeps several basins reachable; a controlled noise sweep would test whether the cross-family fraction is noise-dependent."],"forward_implications":["Ensembles of PINNs on sparse observations can generate candidate initial conditions without a prior guess for conventional search.","To diversify recovered families, vary the training data source rather than the Glorot initialization variant.","Cross-family recovery is bidirectional: Lagrange data can yield the figure-eight, and figure-eight data can yield BHH-type orbits.","The niche is inverse and exploratory settings; the method is not a substitute for integrators on well-posed initial-value problems."],"fun_headline_variants":["PINNs recover unseen three-body orbits from sparse noisy data","Training data steers which three-body orbit family PINNs find","One in four PINN runs finds three-body families absent in training","PINNs locate periodic three-body orbits without initial guesses","Data source, not weight init, shifts recovered three-body families"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That matching the scale-invariant period after numerical refinement, plus simple geometry for the classical cases, is enough to name non-classical families without independent continuation checks.","fun_headline_variants_meta":{"raw":{"variants":["PINNs recover unseen three-body orbits from sparse noisy data","Training data steers which three-body orbit family PINNs find","One in four PINN runs finds three-body families absent in training","PINNs locate periodic three-body orbits without initial guesses","Data source, not weight init, shifts recovered three-body families"]},"model":"grok-4.5","effort":"low","cost_usd":0.002333,"raw_usage":{"total_tokens":1110,"prompt_tokens":973,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":23328000,"prompt_tokens_details":{"text_tokens":973,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":61,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":973,"tokens_out":76,"duration_ms":2940,"temperature":1.0,"reasoning_tokens":61,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T20:41:22.438851+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same seed ensembles and classify every refined orbit by full numerical continuation into known families; if cross-family rates collapse or the chi-squared contrast between training-data sources disappears, the claim that data steers recovery of distinct genuine families fails.","supporting_citations":[],"review_version":1}