Conversation
|
All contributors have signed the CLA ✍️ ✅ |
tcapelle
force-pushed
the
noam
branch
2 times, most recently
from
0526063 to
e22aaf3
Compare
HiveMind Sessions1 session · 1m · $0.89
View all sessions in HiveMind → Run |
Both vorticity auxiliary configs hurt p_tan (+2.5–3.0%) vs current baseline (28.502). Gradient competition in backbone + redundancy with implicit vorticity in velocity targets = consistent regression across all 4 seeds. Added vorticity-as-input-feature to hypothesis queue (no gradient interference). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…lices) - Closed PR #2193 (edward/curvature-spatial-bias): p_tan +2.5% vs baseline. Curvature redundant with position; Python loop reduced epoch budget. - Created PR #2201 (edward/multiscale-slice-hierarchy): coarse-to-fine slice hierarchy [32,64,96] across 3 TransolverBlocks for better OOD generalization and expected 5-8% epoch speedup. - Updated EXPERIMENTS_LOG.md and CURRENT_RESEARCH_STATE.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ext experiments Top picks: iterative-srf (RAFT-style SRF recycling), fore-aft-crossattn-srf (cross-attention in aft-foil refinement head), wake-deficit-feature (explicit wake position features for tandem interaction signal). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…), Round 9 ideas in progress
Five ranked hypotheses from the human team's list in issue #1926: NOBLE nonlinear low-rank branches, GeoTransolver/GALE geometry context cross-attention, PirateNets adaptive residuals (fresh re-implementation, prior PR #2004 crashed on bugs), Moon optimizer (Muon variant with corrected update order), and mHC manifold- constrained hyper-connections. XSA/Muon/MSA/HyperP excluded as definitively failed or inapplicable. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
PR #2190 (Laplacian Eigenvector Mesh PE): p_tan +3.1%, p_oodc +10.3%, p_re +8.4%. Topology-only PE loses spatial coordinates; 32-dim Fourier PE confirmed well-suited. Assigned nezuko #2204: NOBLE nonlinear low-rank branches in TransolverBlock FFN layers. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Round 10 ideas from researcher-agent: Transolver++ Ada-Temp (highest priority), TE coordinate frame, arc-length surface loss, polar coordinate bias, domain-split SRF norm, attention register tokens. Queue ready for next idle students. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…as premature merge)
…search state - Closed PR #2200: Local KNN Attention — p_tan +14% (32.45 vs 28.50). O(N²) mesh infeasibility forced strided anchor fallback; 38% epoch deficit from overhead; uniform anchors miss surface regions. - Assigned alphonse PR #2206: Transolver++ Ada-Temp + Rep-Slice. Per-point adaptive slice temperature from Transolver++ (ICML 2025); 62% surface error reduction in paper's aircraft ablations. - Updated EXPERIMENTS_LOG.md with PR #2200 analysis - Updated CURRENT_RESEARCH_STATE.md: student table, active experiments, Round 10 queue Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… surface loss reweight), update research state
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…pdate research state - PR #2207 (TE Coord Frame): W&B runs crashed at 96 min, metrics unverifiable. Sent back for re-run. - PR #2208 (Iterative SRF RAFT): W&B runs crashed at 82 min. p_tan +5.6% even by student numbers. Closed as dead end (2nd iteration failure). - Assigned askeladd PR #2212: Analytical Cp Delta (thin-airfoil physics baseline for SRF correction). - Updated student status and dead ends in research state. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Top 3 priorities (human-team-requested): PirateNets adaptive residuals, mHC residuals, GeoTransolver GALE. Also: wake deficit feature, slice diversity reg, domain-split SRF norm, tandem feature cross, additive fore-aft cross-attn (retry of #2202 with zero-init additive approach). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…oupling (#2207) * Assign edward: TE coordinate frame experiment * Add TE coordinate frame: trailing-edge-relative input features (+6 channels) Computes per-node (dx, dy, r) offsets from fore-foil and aft-foil trailing edges using raw coordinates. Aft-foil features zero for single-foil samples. -5.3% p_in, -2.8% p_oodc, flat p_tan. Strong physics-motivated improvement. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: senpai-advisor <senpai-advisor@senpai> Co-authored-by: senpai-edward <senpai-edward@senpai> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…_in -5.4%, p_oodc -2.5%, p_re -0.7% Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…) results Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…nel Oracle #2375 closed Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…N SRF Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…to aux heads Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…d most promising Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…back for v3 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…HC, GeoSolver + physics constraints) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…reakthrough monitoring
…n -52% Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…igned Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: senpai-advisor <senpai-advisor@senpai>
… -70%, p_oodc -48% Biggest single improvement in research programme history. Asymmetric cross-attention where aft-foil queries attend to all fore+aft nodes gives direct inter-foil wake coupling. Only p_re regresses +16% — addressable in follow-up. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… improvement ever Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ional gate (#2390) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…geting p_tan and p_re Ideas span: DeltaPhi residual prediction, PIDA physics-augmentation, PIORF Ricci graph rewiring, local Re_x BL feature, DDPM surface head, AB-UPT anchor tokens, displacement thickness, Koopman eigenmodes, flow matching CNF, DANN pretraining, spectral GNN, and NS-residual constraint. Sourced from 10 targeted literature searches (DeltaPhi NeurIPS 2025, AdaField arxiv 2601.07139, PIORF ICLR 2025, B-GNN arxiv 2503.18638, AB-UPT TMLR 2025, Thuerey DDPM 2024). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…on ANP baseline Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… baseline Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…rcher-agent generating bold R46 ideas
…hanges 12 ideas targeting p_tan and p_re: 4 data generation/augmentation strategies (PIDA chord-scaling, flip-symmetry aug, panel synth Re-aug, AoA interpolation) and 8 radical architecture/physics ideas (MARIO hypernetwork, AeroDiT diffusion, GeoMPNN bipartite MPNN, AB-UPT anchor tokens, divergence-free constraint, flow matching head, displacement thickness feature, BL adaptive weighting). All ideas build on the ANP baseline and target the p_re regression and p_tan gap. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
#2393) after Re-gate crashes
…nished (neutral), all experiments status Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…P, AdaLN-Re, context dropout, stochastic cross-attn, Re Mixup) Add research ideas 2026-04-11_04:30 from researcher-agent. Update state: frieren pod restarted, alphonse PirateNet closed (neutral), all new assignments target p_re regression from ANP. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…oned per Morgan directive - Acknowledge Morgan's issue #1860: experiments too incremental - All 8 students WIP on Round 47 p_re fix campaign - researcher-agent commissioned for bold Round 48 slate - Round 48 targets: new model families, synthetic data generation, problem reformulation - Responded to Morgan in issue #1860 with detailed Round 48 plan Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…gmentation Per Morgan's directive (issue #1860): 12 ideas ranked by priority - PIDA chord-scaling aug (exact Reynolds similarity physics) - Flip-symmetry aug (NeuralFoil-validated, doubles dataset) - AeroDiT diffusion head with Re classifier-free guidance - Flow matching generative surface head - Tandem wake superposition prior (panel-initialized decoder) - Multi-fidelity panel-method pretraining (100k synthetic solutions) - Displacement thickness delta* feature (Thwaites BL method) - Multi-resolution ANP context (hierarchical arc-zone tokens) - GeoMPNN bipartite surface-to-volume message passing - Divergence-free hard constraint (Helmholtz projection) - Separated ANP heads (single vs tandem specialization) - Spectral normalization of ANP Q/K/V Zero incremental tweaks. All ideas are radical model changes, synthetic data generation, physics-grounded augmentation, or problem reformulation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…orm regression) - #2386 nezuko/stagnation-constraint: all metrics +8-11% worse. Fragile node ID + normalization issues. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Discovered --anp_srf code was never committed (empty merge commit) - Closed Round 47 ANP PRs #2394-2398 (can't run without ANP code) - Assigned alphonse #2399: ANP re-implementation (P0) - Assigned edward #2400: PIDA chord-scaling aug (exact physics) - Assigned fern #2401: Flip-symmetry aug (NeuralFoil-validated) - Assigned tanjiro #2402: Flow matching generative head - Assigned nezuko #2403: Multi-fidelity panel pretraining (100k synthetic) - Baseline reverted to pre-ANP (PR #2357): 11.872/7.459/26.319/6.229 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This branch is out of date, but let's leave it here so we know what Noam did.