{"paper":{"title":"Policy-DRIFT: Dynamic Reward-Informed Flow Trajectory Steering","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A conditional flow matching model steers turbulent flow states to cut drag by 49 percent while using 37 times less energy than deep reinforcement learning.","cross_cats":[],"primary_cat":"physics.flu-dyn","authors_text":"Abhijeet Vishwasrao, Atharva Mahajan, Ricardo Vinuesa, Yuning Wang","submitted_at":"2026-05-13T18:34:46Z","abstract_excerpt":"Skin-friction drag induced by wall-bounded turbulent flows accounts for a substantial fraction of energy consumption across commercial aerospace, wind energy, and marine transport. Its active reduction is one of the highest-value targets in engineering fluid dynamics. Deep reinforcement learning (DRL) has emerged as the leading approach for real-time flow control, yet its performance ceiling is set not by algorithmic capability but by reward structure, the naive scalar objective does not optimally reflect the underlying physics. Policy-DRIFT bypasses this ceiling by relocating reward informati"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Policy-DRIFT achieves 49% drag reduction approaching the theoretical upper bound, which is ≈16% higher than the DRL benchmark, while consuming 37× less actuation energy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The conditional flow matching model constructs a physically-grounded manifold of realisable flow states that spans multiple control regimes without introducing unphysical artifacts or missing important dynamics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Policy-DRIFT combines conditional flow matching with terminal reward guidance and decoupled DRL to achieve 49% drag reduction in Re_tau=180 channel flow, 16% above DRL benchmarks and with 37 times less actuation energy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A conditional flow matching model steers turbulent flow states to cut drag by 49 percent while using 37 times less energy than deep reinforcement learning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7fc3b2c74b65febbc748175a2bd56c7f1decde8bbd9cc502f4e3833a5d3d0f96"},"source":{"id":"2605.14022","kind":"arxiv","version":1},"verdict":{"id":"88546f67-b178-49fe-bed8-8a0d1ea315da","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:37:49.382162Z","strongest_claim":"Policy-DRIFT achieves 49% drag reduction approaching the theoretical upper bound, which is ≈16% higher than the DRL benchmark, while consuming 37× less actuation energy.","one_line_summary":"Policy-DRIFT combines conditional flow matching with terminal reward guidance and decoupled DRL to achieve 49% drag reduction in Re_tau=180 channel flow, 16% above DRL benchmarks and with 37 times less actuation energy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The conditional flow matching model constructs a physically-grounded manifold of realisable flow states that spans multiple control regimes without introducing unphysical artifacts or missing important dynamics.","pith_extraction_headline":"A conditional flow matching model steers turbulent flow states to cut drag by 49 percent while using 37 times less energy than deep reinforcement learning."},"references":{"count":12,"sample":[{"doi":"","year":null,"title":"M. Beneitez, A. Cremades, L. Guastoni, and R. Vinuesa. Improving turbulence control through explainable deep learning.arXiv preprint arXiv:2504.02354,","work_id":"c66f111a-854b-4d4b-83fb-c03313b95bc0","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Se (3)-stochastic flow matching for protein backbone generation","work_id":"f8adff70-705d-40c6-922d-4f1e4aa71302","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"H. Harder, A. Vishwasrao, L. Guastoni, R. Vinuesa, and S. Peitz. Efficient probabilistic surro- gate modeling techniques for partially-observed large-scale dynamical systems.arXiv preprint arXiv:2511.","work_id":"e3910382-cccd-4f0a-bc77-c85a8b78e41a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Planning with Diffusion for Flexible Behavior Synthesis","work_id":"38b2c635-b754-412a-a8f5-dfcf3e405c95","ref_index":4,"cited_arxiv_id":"2205.09991","is_internal_anchor":true},{"doi":"","year":null,"title":"Continuous control with deep reinforcement learning","work_id":"41a65444-c819-4303-a1f1-b075aa86d40c","ref_index":5,"cited_arxiv_id":"1509.02971","is_internal_anchor":true}],"resolved_work":12,"snapshot_sha256":"3b9d49a76d94c5a5538daeec3744507a1874eb791078f80f18377e897baa12b1","internal_anchors":4},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}