[{"data":1,"prerenderedAt":2766},["ShallowReactive",2],{"post-arxiv-2506-13777v1":3,"site-info":1018,"related-en-arxiv-2506-13777v1":1114},{"zh":4,"en":678},{"id":5,"title":6,"author":7,"body":8,"cover":660,"coverAlt":6,"date":661,"description":662,"draft":663,"extension":664,"featured":663,"lang":665,"meta":666,"navigation":667,"path":668,"seo":669,"slug":670,"stem":671,"tags":672,"translationKey":670,"updated":676,"__hash__":677},"blog\u002Fblog\u002Fzh-TW\u002Farxiv-2506-13777v1.md","複雜城市系統的實體資訊AI綜述","Jecision",{"type":9,"value":10,"toc":634},"minimark",[11,22,27,31,69,72,76,80,206,208,212,216,324,326,330,334,339,359,363,383,387,407,411,431,435,449,453,467,471,485,489,503,505,509,513,517,537,541,578,582,598,602,613,615,619],[12,13,15,16],"h2",{"id":14},"論文位置httpsarxivorghtml250613777v1","論文位置：",[17,18,19],"a",{"href":19,"rel":20},"https:\u002F\u002Farxiv.org\u002Fhtml\u002F2506.13777v1",[21],"nofollow",[23,24,26],"h3",{"id":25},"a-survey-of-physics-informed-ai-for-complex-urban-systems","A Survey of Physics-Informed AI for Complex Urban Systems",[12,28,30],{"id":29},"_1-引言-introduction","1. 引言 (Introduction)",[32,33,34,42,48,54,60,63,66],"ul",{},[35,36,37,41],"li",{},[38,39,40],"strong",{},"背景","：物理與AI的融合是科學研究的新範式，2024年諾貝爾物理學獎和化學獎都授予AI相關成果，凸顯了這一趨勢。",[35,43,44,47],{},[38,45,46],{},"城市系統的複雜性","：傳統實體或純資料驅動方法難以應對城市系統的多維動態互動挑戰。",[35,49,50,53],{},[38,51,52],{},"核心問題","：如何將物理規律（如流體力學、熱傳導、社會力模型）與AI的數據驅動能力結合，以提升城市系統的預測準確性、可解釋性和決策能力。",[35,55,56,59],{},[38,57,58],{},"論文貢獻","：",[35,61,62],{},"提出新的PIAI方法分類法（三大典範、七種方法）；",[35,64,65],{},"系統性回顧PIAI在八大城市子系統的應用；",[35,67,68],{},"指出目前限制並規劃未來研究方向。",[70,71],"hr",{},[12,73,75],{"id":74},"_2-物理資訊ai方法-physics-informed-ai-methods","2. 物理資訊AI方法 (Physics-informed AI Methods)",[23,77,79],{"id":78},"三大典範與七種方法","三大典範與七種方法：",[81,82,83,102],"table",{},[84,85,86],"thead",{},[87,88,89,93,96,99],"tr",{},[90,91,92],"th",{},"範式",[90,94,95],{},"方法",[90,97,98],{},"核心思想",[90,100,101],{},"典型技術",[103,104,105,122,135,148,164,177,193],"tbody",{},[87,106,107,113,116,119],{},[108,109,110],"td",{},[38,111,112],{},"實體整合AI",[108,114,115],{},"1. PINN with Loss Function",[108,117,118],{},"將物理約束作為損失項嵌入神經網路",[108,120,121],{},"PINN, CPINN, PCNN",[87,123,124,126,129,132],{},[108,125],{},[108,127,128],{},"2. PINN with Weight Initialization",[108,130,131],{},"用物理知識初始化網路權重，提升訓練效率",[108,133,134],{},"遷移學習、自我監督預訓練",[87,136,137,139,142,145],{},[108,138],{},[108,140,141],{},"3. PINN with Architecture Design",[108,143,144],{},"將物理結構嵌入網路架構，增強可解釋性",[108,146,147],{},"對稱性編碼、哈密頓結構",[87,149,150,155,158,161],{},[108,151,152],{},[38,153,154],{},"物理-AI混合整合",[108,156,157],{},"4. Sequential Physics-AI Ensemble",[108,159,160],{},"物理與AI模型順序協作，分階段處理",[108,162,163],{},"殘差建模、管道架構",[87,165,166,168,171,174],{},[108,167],{},[108,169,170],{},"5. Parallel Physics-AI Ensemble",[108,172,173],{},"物理與AI模型並行處理，結果融合",[108,175,176],{},"動態加權、多模型輸出組合",[87,178,179,184,187,190],{},[108,180,181],{},[38,182,183],{},"AI整合物理",[108,185,186],{},"6. AI-Discovered Physics Model",[108,188,189],{},"用AI從資料自動發現物理規律",[108,191,192],{},"SINDy框架、符號迴歸",[87,194,195,197,200,203],{},[108,196],{},[108,198,199],{},"7. Neural Physics Model",[108,201,202],{},"以AI模組取代物理模型中難以建模的部分",[108,204,205],{},"神經網路取代湍流閉合項",[70,207],{},[12,209,211],{"id":210},"_3-複雜城市系統及涉及的物理定律-complex-urban-systems-and-involved-physical-laws","3. 複雜城市系統及涉及的物理定律 (Complex Urban Systems and Involved Physical Laws)",[23,213,215],{"id":214},"七大子系統及其核心物理原理","七大子系統及其核心物理原理：",[81,217,218,231],{},[84,219,220],{},[87,221,222,225,228],{},[90,223,224],{},"子系統",[90,226,227],{},"物理定律\u002F模型",[90,229,230],{},"應用範例",[103,232,233,246,259,272,285,298,311],{},[87,234,235,240,243],{},[108,236,237],{},[38,238,239],{},"能源",[108,241,242],{},"電力潮流方程式、壓力平衡方程式",[108,244,245],{},"電網最佳化、再生能源整合",[87,247,248,253,256],{},[108,249,250],{},[38,251,252],{},"環境",[108,254,255],{},"高斯煙羽模型、SWMM水文模型",[108,257,258],{},"空氣品質預測、洪水模擬",[87,260,261,266,269],{},[108,262,263],{},[38,264,265],{},"經濟",[108,267,268],{},"網路流模型、Huff消費選擇模型",[108,270,271],{},"供應鏈最佳化、市場需求預測",[87,273,274,279,282],{},[108,275,276],{},[38,277,278],{},"交通",[108,280,281],{},"LWR交通流模型、跟馳模型",[108,283,284],{},"交通狀態估計、號誌控制",[87,286,287,292,295],{},[108,288,289],{},[38,290,291],{},"資訊",[108,293,294],{},"香農容量公式、射線追蹤模型",[108,296,297],{},"社群網路資訊傳播預測",[87,299,300,305,308],{},[108,301,302],{},[38,303,304],{},"公共服務",[108,306,307],{},"SIR流行病模型、社會力模型",[108,309,310],{},"疫情預測、人群疏散模擬",[87,312,313,318,321],{},[108,314,315],{},[38,316,317],{},"緊急管理",[108,319,320],{},"社會力模型、時空衰減模型",[108,322,323],{},"災害動力學建模、災後人口流動預測",[70,325],{},[12,327,329],{"id":328},"_4-實體資訊ai在複雜城市系統中的應用-piai-in-complex-urban-systems","4. 實體資訊AI在複雜城市系統中的應用 (PIAI in Complex Urban Systems)",[23,331,333],{"id":332},"分領域應用摘要","分領域應用摘要：",[335,336,338],"h4",{"id":337},"_41-能源系統","4.1 能源系統",[32,340,341,347,353],{},[35,342,343,346],{},[38,344,345],{},"發電","：PINN用於風力發電場重建、太陽能輸出預測；",[35,348,349,352],{},[38,350,351],{},"輸電","：結合基爾霍夫定律的神經網路加速最優潮流計算；",[35,354,355,358],{},[38,356,357],{},"用電","：整合價格-需求關係、熱力學模型預測EV充電需求。",[335,360,362],{"id":361},"_42-環境系統","4.2 環境系統",[32,364,365,371,377],{},[35,366,367,370],{},[38,368,369],{},"空氣","：擴散-平流方程式與神經網路結合提升污染物預測精度；",[35,372,373,376],{},[38,374,375],{},"水","：PINN求解Richards方程式模擬土壤水流動；",[35,378,379,382],{},[38,380,381],{},"碳","：實體資訊網路整合遙感資料估算碳儲量。",[335,384,386],{"id":385},"_43-交通系統","4.3 交通系統",[32,388,389,395,401],{},[35,390,391,394],{},[38,392,393],{},"狀態估計","：LWR\u002FARZ模型與深度學習融合，從稀疏感測器資料推斷交通狀態；",[35,396,397,400],{},[38,398,399],{},"流量預測","：時空微分方程網路（STDEN）建模潛在能量場驅動流量變化；",[35,402,403,406],{},[38,404,405],{},"控制","：物理約束強化學習用於信號燈和匝道控制最佳化。",[335,408,410],{"id":409},"_44-經濟系統","4.4 經濟系統",[32,412,413,419,425],{},[35,414,415,418],{},[38,416,417],{},"製造","：PINN解決熱傳導問題，優化工業流程；",[35,420,421,424],{},[38,422,423],{},"金融","：Black-Scholes方程式與神經網路結合用於選擇權定價；",[35,426,427,430],{},[38,428,429],{},"商品與POI","：結合規模定律與軌跡流程圖預測公司成長與使用者移動模式。",[335,432,434],{"id":433},"_45-資訊系統","4.5 資訊系統",[32,436,437,443],{},[35,438,439,442],{},[38,440,441],{},"資訊傳播","：基於熱傳導方程式建模社交網路資訊密度流；",[35,444,445,448],{},[38,446,447],{},"觀點演化","：神經ODE模擬社群網路中使用者觀點的動態變化。",[335,450,452],{"id":451},"_46-公共服務","4.6 公共服務",[32,454,455,461],{},[35,456,457,460],{},[38,458,459],{},"公共衛生","：EINN（流行病學資訊神經網路）融合多源資料提升疫情預測；",[35,462,463,466],{},[38,464,465],{},"公共安全","：社會力模型與AI結合優化人群模擬和車輛軌跡預測。",[335,468,470],{"id":469},"_47-緊急管理","4.7 緊急管理",[32,472,473,479],{},[35,474,475,478],{},[38,476,477],{},"災害預測","：PINN用於洪水峰值深度預測、地震前兆偵測；",[35,480,481,484],{},[38,482,483],{},"人群模擬","：物理知識增強的生成模型模擬災時人群移動。",[335,486,488],{"id":487},"_48-城市系統整體","4.8 城市系統整體",[32,490,491,497],{},[35,492,493,496],{},[38,494,495],{},"跨領域耦合","：圖神經網路耦合洪水動力學與交通模型，模擬極端天氣下的城市運作；",[35,498,499,502],{},[38,500,501],{},"多尺度整合","：從建築級熱模型到區域氣候模式的統一PIAI框架。",[70,504],{},[12,506,508],{"id":507},"_5-討論與開放問題-discussions-and-open-problems","5. 討論與開放問題 (Discussions and Open Problems)",[23,510,512],{"id":511},"_51-方法選擇指南","5.1 方法選擇指南",[514,515,516],"p",{},"選擇PIAI方法需考慮：",[32,518,519,525,531],{},[35,520,521,524],{},[38,522,523],{},"物理知識可用性","（強約束 vs 弱約束）；",[35,526,527,530],{},[38,528,529],{},"資料可用性","（豐富 vs 稀疏）；",[35,532,533,536],{},[38,534,535],{},"任務目標","（即時性、準確性、可解釋性）。",[23,538,540],{"id":539},"_52-當前限制與未來方向","5.2 當前限制與未來方向",[32,542,543,549,555,561,566,569,572,575],{},[35,544,545,548],{},[38,546,547],{},"理論保障不足","：PINN的收斂性和泛化性缺乏嚴格證明；",[35,550,551,554],{},[38,552,553],{},"運算效率","：訓練成本高，難以即時部署；",[35,556,557,560],{},[38,558,559],{},"知識表示","：缺乏對因果、規則等非數值知識的整合；",[35,562,563,59],{},[38,564,565],{},"未來方向",[35,567,568],{},"發展多解析度網路架構；",[35,570,571],{},"結合主動學習與遷移學習；",[35,573,574],{},"探索PIAI與LLM的融合；",[35,576,577],{},"建構城市數位孿生和通用世界模型。",[23,579,581],{"id":580},"_53-關鍵發現","5.3 關鍵發現",[32,583,584,595],{},[35,585,586,587,590,591,594],{},"城市系統是PIAI的天然試驗場，因其兼具",[38,588,589],{},"物理規律性","與",[38,592,593],{},"人類活動不確定性","；",[35,596,597],{},"資料豐富的任務偏向AI主導方法，而決策支援任務更需要物理可解釋性。",[23,599,601],{"id":600},"_54-開放問題","5.4 開放問題",[32,603,604,607,610],{},[35,605,606],{},"如何為高維複雜環境中的PIAI建立理論根基？",[35,608,609],{},"如何建構統一框架整合確定性、隨機性和基於Agent的模型？",[35,611,612],{},"如何實現跨城市子系統的可遷移PIAI模型？",[70,614],{},[12,616,618],{"id":617},"_6-結論-conclusion","6. 結論 (Conclusion)",[32,620,621,628,631],{},[35,622,623,624,627],{},"PIAI透過融合物理與AI，為城市系統建模提供了",[38,625,626],{},"更可靠、可解釋、自適應","的解決方案。",[35,629,630],{},"論文提出的分類法與應用綜述為後續研究提供了清晰路線圖。",[35,632,633],{},"未來需跨領域合作，推動PIAI在真實城市環境中的部署與驗證。",{"title":635,"searchDepth":636,"depth":636,"links":637},"",2,[638,643,644,647,650,653,659],{"id":14,"depth":636,"text":639,"children":640},"論文位置：https:\u002F\u002Farxiv.org\u002Fhtml\u002F2506.13777v1",[641],{"id":25,"depth":642,"text":26},3,{"id":29,"depth":636,"text":30},{"id":74,"depth":636,"text":75,"children":645},[646],{"id":78,"depth":642,"text":79},{"id":210,"depth":636,"text":211,"children":648},[649],{"id":214,"depth":642,"text":215},{"id":328,"depth":636,"text":329,"children":651},[652],{"id":332,"depth":642,"text":333},{"id":507,"depth":636,"text":508,"children":654},[655,656,657,658],{"id":511,"depth":642,"text":512},{"id":539,"depth":642,"text":540},{"id":580,"depth":642,"text":581},{"id":600,"depth":642,"text":601},{"id":617,"depth":636,"text":618},"\u002Fimages\u002Fblog\u002F2506.13777.1.webp","2025-09-16","城市系統的物理資訊 AI 綜述：整理這篇 survey 如何把物理約束接進都市尺度的預測模型，以及它點出的資料、尺度與可解釋性缺口。",false,"md","zh-TW",{},true,"\u002Fblog\u002Fzh-tw\u002Farxiv-2506-13777v1",{"title":6,"description":662},"arxiv-2506-13777v1","blog\u002Fzh-TW\u002Farxiv-2506-13777v1",[673,674,675],"paper","urban","survey",null,"MX39x2Lyu08AZhatQPB9H_RGdiocApt5nsn4PPt8w_c",{"id":679,"title":680,"author":7,"body":681,"cover":660,"coverAlt":1009,"date":661,"description":1010,"draft":663,"extension":664,"featured":663,"lang":1011,"meta":1012,"navigation":667,"path":1013,"seo":1014,"slug":670,"stem":1015,"tags":1016,"translationKey":670,"updated":676,"__hash__":1017},"blog\u002Fblog\u002Fen\u002Farxiv-2506-13777v1.md","Physics-informed AI for complex urban systems",{"type":9,"value":682,"toc":1002},[683,697,701,704,707,714,718,831,838,842,936,939,950,954,957,977,980,984,987,993,999],[514,684,685,688,689,693,696],{},[38,686,687],{},"Paper"," — ",[17,690,692],{"href":19,"rel":691},[21],"arXiv:2506.13777",[38,694,695],{},"Title"," — A Survey of Physics-Informed AI for Complex Urban Systems",[12,698,700],{"id":699},"why-this-survey-exists","Why this survey exists",[514,702,703],{},"Cities are the awkward middle ground of modelling. Pure physics cannot capture human behaviour; pure data-driven models cannot explain themselves, and they break exactly when conditions change — which is when a city most needs a prediction.",[514,705,706],{},"Physics-informed AI (PIAI) is the attempt to have both. Its selling point is a model that respects conservation laws and heat transfer while still learning from data.",[514,708,709,710,713],{},"The survey's contribution is organisational: ",[38,711,712],{},"a taxonomy of three paradigms and seven methods",", a review of applications across eight urban subsystems, and an honest list of what is still missing.",[12,715,717],{"id":716},"the-taxonomy","The taxonomy",[81,719,720,736],{},[84,721,722],{},[87,723,724,727,730,733],{},[90,725,726],{},"Paradigm",[90,728,729],{},"Method",[90,731,732],{},"Core idea",[90,734,735],{},"Typical techniques",[103,737,738,751,764,777,791,804,818],{},[87,739,740,743,746,749],{},[108,741,742],{},"Physics-integrated AI",[108,744,745],{},"1. PINN with loss function",[108,747,748],{},"physics constraints as a loss term",[108,750,121],{},[87,752,753,755,758,761],{},[108,754],{},[108,756,757],{},"2. PINN with weight initialisation",[108,759,760],{},"initialise weights from physical knowledge",[108,762,763],{},"transfer learning, self-supervised pretraining",[87,765,766,768,771,774],{},[108,767],{},[108,769,770],{},"3. PINN with architecture design",[108,772,773],{},"bake physical structure into the network",[108,775,776],{},"symmetry encoding, Hamiltonian structure",[87,778,779,782,785,788],{},[108,780,781],{},"Hybrid",[108,783,784],{},"4. Sequential ensemble",[108,786,787],{},"physics and AI cooperate in stages",[108,789,790],{},"residual modelling, pipeline architectures",[87,792,793,795,798,801],{},[108,794],{},[108,796,797],{},"5. Parallel ensemble",[108,799,800],{},"run both, fuse the outputs",[108,802,803],{},"dynamic weighting, model combination",[87,805,806,809,812,815],{},[108,807,808],{},"AI-integrated physics",[108,810,811],{},"6. AI-discovered physics",[108,813,814],{},"let AI find the governing law",[108,816,817],{},"SINDy, symbolic regression",[87,819,820,822,825,828],{},[108,821],{},[108,823,824],{},"7. Neural physics model",[108,826,827],{},"replace the hard-to-model part with a network",[108,829,830],{},"neural closures for turbulence",[514,832,833,834,837],{},"The useful distinction is ",[38,835,836],{},"where physics enters",": as a penalty, as an initialisation, as architecture, as a co-model, or not at all (physics emerges from data).",[12,839,841],{"id":840},"where-it-is-used","Where it is used",[81,843,844,857],{},[84,845,846],{},[87,847,848,851,854],{},[90,849,850],{},"Subsystem",[90,852,853],{},"Physical law \u002F model",[90,855,856],{},"Example application",[103,858,859,870,881,892,903,914,925],{},[87,860,861,864,867],{},[108,862,863],{},"Energy",[108,865,866],{},"power flow equations, pressure balance",[108,868,869],{},"grid optimisation, renewable integration",[87,871,872,875,878],{},[108,873,874],{},"Environment",[108,876,877],{},"Gaussian plume, SWMM hydrology",[108,879,880],{},"air quality, flood simulation",[87,882,883,886,889],{},[108,884,885],{},"Economy",[108,887,888],{},"network flow, Huff choice model",[108,890,891],{},"supply chain, demand forecasting",[87,893,894,897,900],{},[108,895,896],{},"Transport",[108,898,899],{},"LWR traffic flow, car-following",[108,901,902],{},"state estimation, signal control",[87,904,905,908,911],{},[108,906,907],{},"Information",[108,909,910],{},"Shannon capacity, ray tracing",[108,912,913],{},"information diffusion in social networks",[87,915,916,919,922],{},[108,917,918],{},"Public services",[108,920,921],{},"SIR epidemic, social force",[108,923,924],{},"outbreak forecasting, crowd evacuation",[87,926,927,930,933],{},[108,928,929],{},"Emergency",[108,931,932],{},"social force, spatio-temporal decay",[108,934,935],{},"disaster dynamics, post-disaster mobility",[514,937,938],{},"Highlights from the review: PINNs reconstructing wind farms and predicting solar output; Kirchhoff-constrained networks accelerating optimal power flow; diffusion–advection equations fused with neural networks for pollutant forecasting; LWR\u002FARZ models combined with deep learning to estimate traffic state from sparse sensors; neural ODEs modelling opinion dynamics; epidemiological neural networks fusing multi-source data for outbreak prediction.",[514,940,941,942,945,946,949],{},"The last section is the most interesting: ",[38,943,944],{},"cross-domain coupling"," — graph neural networks linking flood dynamics with transport models to simulate a city under extreme weather — and ",[38,947,948],{},"multi-scale integration",", from building-level thermal models up to regional climate.",[12,951,953],{"id":952},"the-open-problems","The open problems",[514,955,956],{},"The survey is refreshingly blunt about what is unsolved:",[32,958,959,965,971],{},[35,960,961,964],{},[38,962,963],{},"Weak theory"," — convergence and generalisation of PINNs lack rigorous proofs",[35,966,967,970],{},[38,968,969],{},"Compute"," — training cost is high, which rules out real-time deployment",[35,972,973,976],{},[38,974,975],{},"Knowledge representation"," — causal and rule-based (non-numeric) knowledge is not integrated",[514,978,979],{},"Proposed directions include multi-resolution architectures, active and transfer learning, combining PIAI with LLMs, and building urban digital twins or general world models.",[12,981,983],{"id":982},"what-i-take-from-it","What I take from it",[514,985,986],{},"Two observations in the paper are worth more than the taxonomy itself:",[514,988,989,992],{},[38,990,991],{},"Cities are the natural testbed for PIAI"," precisely because they combine physical regularity with human unpredictability. Both extremes exist in the same system, so a method that only works for one of them will fail visibly.",[514,994,995,998],{},[38,996,997],{},"Data-rich tasks drift toward AI-led methods; decision-support tasks need physical interpretability."," That is a practical rule for choosing a method — not based on what is fashionable, but on who has to defend the output.",[514,1000,1001],{},"For anyone building city-scale models, the \"how do I choose\" section is the one to read first: physical knowledge availability, data density, and whether the deliverable is a number or a decision.",{"title":635,"searchDepth":636,"depth":636,"links":1003},[1004,1005,1006,1007,1008],{"id":699,"depth":636,"text":700},{"id":716,"depth":636,"text":717},{"id":840,"depth":636,"text":841},{"id":952,"depth":636,"text":953},{"id":982,"depth":636,"text":983},"A survey of physics-informed AI for complex urban systems","How do you put physics into city-scale models? This survey proposes a taxonomy of three paradigms and seven methods, reviews eight urban subsystems, and names what is still missing.","en",{},"\u002Fblog\u002Fen\u002Farxiv-2506-13777v1",{"title":680,"description":1010},"blog\u002Fen\u002Farxiv-2506-13777v1",[673,674,675],"1gwEeKIu-vP0ABuoHpyY9xShz03tEALrPNkeNBiXGXc",{"id":1019,"company":1020,"extension":1031,"meta":1032,"services":1033,"social":1111,"stem":1112,"__hash__":1113},"site\u002Fsite.yml",{"name":1021,"nameEn":1022,"email":1023,"address":1024,"addressEn":1025,"lineUrl":1026,"lineId":1027,"duns":1028,"youtube":1029,"youtubeId":1030},"銡策數位有限公司","Jecision Digital Co., Ltd.","contact@jecision.com","231 新北市新店區北新路二段 262 號 3 樓之 3","3F.-3, No. 262, Sec. 2, Beixin Rd., Xindian Dist., New Taipei City 231, Taiwan","https:\u002F\u002Fline.me\u002FR\u002Fti\u002Fp\u002F@374qnumi","@374qnumi","963905705","https:\u002F\u002Fwww.youtube.com\u002Fshorts\u002F6UE0WTmyjsg","6UE0WTmyjsg","yml",{},[1034,1058,1084],{"key":1035,"zh":1036,"en":1047},"audit",{"title":1037,"tagline":1038,"summary":1039,"points":1040,"deliverable":1044,"detail":1045,"duration":1046},"企業 AI 導入健檢","先看清楚，再花錢","60 分鐘線上訪談，盤點最耗時、最依賴人力的營運環節；我們會告訴你哪些工作現在就能交給 AI、哪些還不成熟。",[1041,1042,1043],"現況盤點：梳理日常營運中最耗時、最依賴人力的環節","痛點診斷：判斷哪些工作現在就能交給 AI，哪些建議先保留人力","可行性評估：依你的業務特性與預算，篩選適合的模型與工具","《企業 AI 導入建議報告書》（PDF）","含 1–2 個「開發成本最低、成效最明顯」的優先項目、技術棧建議、預期節省的人力時間與下一步","訪談 60 分鐘，報告書 3 個工作天內交付",{"title":1048,"tagline":1049,"summary":1050,"points":1051,"deliverable":1055,"detail":1056,"duration":1057},"AI Readiness Audit","Look before you spend","A 60-minute remote audit of the workflows that eat the most time and people-power — and an honest answer on what AI can take over today, and what it cannot.",[1052,1053,1054],"Audit the daily workflows that consume the most time and headcount","Triage which tasks AI can own today, and which should stay human","Shortlist the models and tools that fit your budget and constraints","AI Adoption Recommendation Report (PDF)","One or two highest-ROI entry points, a recommended tech stack, expected hours saved, and the concrete next step","60-minute call, report delivered within 3 business days",{"key":1059,"zh":1060,"en":1071},"prompt",{"title":1061,"tagline":1062,"summary":1063,"points":1064,"deliverable":1068,"detail":1069,"duration":1070},"提示詞工程","讓 AI 穩定產出，而不是碰運氣","針對你的實際任務設計提示詞與多步驟工作流，並做極限測試，確保換了資料、換了人操作也不會走鐘。",[1065,1066,1067],"客製提示詞：ChatGPT、Claude、Gemini、Midjourney 等主流工具的專屬模板","除錯與優化：修掉幻覺、格式跑掉、時好時壞的問題","工作流建置：把複雜任務拆成多步驟提示詞與可重複的 SOP","提示詞文件或設定檔","附使用說明與範例，並提供 1 對 1 教學，確保你與團隊能自己維護","依任務複雜度報價，通常 3–7 個工作天",{"title":1072,"tagline":1073,"summary":1074,"points":1075,"deliverable":1081,"detail":1082,"duration":1083},"Prompt Engineering","Make AI produce reliably, not by luck","Prompt sets and multi-step workflows designed for your actual tasks, stress-tested so they keep working when the data or the operator changes.",[1076,1077,1079],"Custom prompts for ChatGPT, Claude, Gemini, Midjourney and friends",{"Debugging and tuning":1078},"kill hallucinations, broken formats and flaky output",{"Workflow design":1080},"break complex work into repeatable multi-step SOPs","Prompt library or exportable config","Ships with usage notes and examples, plus a 1:1 walkthrough so your team can maintain it","Quoted by complexity, typically 3–7 business days",{"key":1085,"zh":1086,"en":1097},"poc",{"title":1087,"tagline":1088,"summary":1089,"points":1090,"deliverable":1094,"detail":1095,"duration":1096},"極速 PoC","五天，看到能跑的東西","不談架構藍圖，直接做一個能動的雛形。五個工作天內交付線上可操作的智能體，用你真實的資料驗證。",[1091,1092,1093],"Day 1：15 分鐘對焦，鎖定單一痛點與資料盤點","Day 2–4：模型串接、邏輯設計與雛形開發","Day 5：Demo 展示與 AI 擴展建議書","可線上操作的智能體雛形","交付範圍為概念驗證與展示，不含後續維運或大規模伺服器部署；要進階成完整系統會另行報價","5 個工作天",{"title":1098,"tagline":1099,"summary":1100,"points":1101,"deliverable":1108,"detail":1109,"duration":1110},"Rapid PoC","Five days to something that runs","No architecture decks — a working prototype built in five business days, validated against your real data.",[1102,1104,1106],{"Day 1":1103},"15-minute scoping call, one pain point, data inventory",{"Day 2–4":1105},"model integration, logic design, prototype build",{"Day 5":1107},"live demo plus an AI expansion recommendation","A runnable online agent prototype","Proof of concept and demo scope only; ongoing operations and large-scale deployment are quoted separately","5 business days",[],"site","XTwARfVAOcrXNiKehsM6V6cdkd2HXy4d5yZVrSWrI3Q",[1115,1387],{"id":1116,"title":1117,"author":7,"body":1118,"cover":1374,"coverAlt":1375,"date":1376,"description":1377,"draft":663,"extension":664,"featured":663,"lang":1011,"meta":1378,"navigation":667,"path":1379,"seo":1380,"slug":1381,"stem":1382,"tags":1383,"translationKey":1381,"updated":676,"__hash__":1386},"blog\u002Fblog\u002Fen\u002Farxiv-2510-27688v1.md","CALM, or why token-by-token generation hit a wall",{"type":9,"value":1119,"toc":1366},[1120,1133,1137,1140,1143,1157,1160,1164,1171,1179,1228,1234,1238,1244,1250,1255,1266,1269,1274,1288,1292,1310,1314,1335,1339,1346,1349,1363],[514,1121,1122,1125,1126,688,1128],{},[38,1123,1124],{},"From"," — Tencent WeChat AI × Tsinghua University\n",[38,1127,687],{},[17,1129,1132],{"href":1130,"rel":1131},"https:\u002F\u002Farxiv.org\u002Fhtml\u002F2510.27688v1",[21],"arXiv:2510.27688",[12,1134,1136],{"id":1135},"the-wall","The wall",[514,1138,1139],{},"Every large language model today predicts one token at a time. Each token costs a full forward pass, so cost grows linearly — and with attention, worse — as sequences get longer.",[514,1141,1142],{},"Text representation has already moved from characters to subwords (BPE), but the discrete representation itself is running out of room:",[32,1144,1145,1151],{},[35,1146,1147,1150],{},[38,1148,1149],{},"Vocabulary inflation",": to cover more meaning, vocabularies grew from 32k to 256k. The final layer's compute and memory cost exploded with them.",[35,1152,1153,1156],{},[38,1154,1155],{},"Low information density",": each discrete token carries roughly 15–18 bits. The model has enormous compute and spends it on the labour-intensive job of predicting one low-information token at a time.",[514,1158,1159],{},"CALM's premise: stop picking from a vocabulary, and predict a point in continuous space instead.",[12,1161,1163],{"id":1162},"the-idea","The idea",[514,1165,1166,1167,1170],{},"An autoencoder compresses ",[38,1168,1169],{},"K discrete tokens into one high-dimensional continuous vector",", so sequence length drops from T to T\u002FK and the number of generation steps drops by the same factor.",[514,1172,1173,1174,1178],{},"It trades more ",[1175,1176,1177],"em",{},"semantic bandwidth"," per step for fewer steps.",[81,1180,1181,1193],{},[84,1182,1183],{},[87,1184,1185,1187,1190],{},[90,1186],{},[90,1188,1189],{},"Discrete",[90,1191,1192],{},"CALM (continuous)",[103,1194,1195,1206,1217],{},[87,1196,1197,1200,1203],{},[108,1198,1199],{},"Scaling",[108,1201,1202],{},"vocabulary grows exponentially",[108,1204,1205],{},"vector dimension grows smoothly",[87,1207,1208,1211,1214],{},[108,1209,1210],{},"Objective",[108,1212,1213],{},"pick one of hundreds of thousands",[108,1215,1216],{},"predict coordinates in a continuous space",[87,1218,1219,1222,1225],{},[108,1220,1221],{},"Sequence",[108,1223,1224],{},"long T",[108,1226,1227],{},"short T\u002FK",[1229,1230,1231],"blockquote",{},[514,1232,1233],{},"A useful mental image: conventional generation is searching a warehouse with hundreds of thousands of drawers. CALM draws coordinates.",[12,1235,1237],{"id":1236},"four-pillars","Four pillars",[514,1239,1240,1243],{},[38,1241,1242],{},"1. A high-fidelity autoencoder."," Compresses K tokens into 128+ dimensions at roughly 75M parameters of overhead, with reconstruction accuracy above 99.9%. Without this, everything downstream is garbage.",[514,1245,1246,1249],{},[38,1247,1248],{},"2. An energy-based head, no likelihood."," Instead of softmax over a vocabulary, the model regresses the next target vector directly — a single step, not the dozens or hundreds of denoising iterations a diffusion model needs.",[514,1251,1252],{},[38,1253,1254],{},"3. Three tricks for robustness in continuous space.",[32,1256,1257,1260,1263],{},[35,1258,1259],{},"Variational regularisation, to smooth folds and cliffs in the latent space",[35,1261,1262],{},"KL clipping, to stop dimensions collapsing",[35,1264,1265],{},"Double dropout, for extra noise tolerance",[514,1267,1268],{},"Together they build a smooth, elastic landscape where a small prediction error does not turn into gibberish.",[514,1270,1271],{},[38,1272,1273],{},"4. Evaluation without probabilities.",[32,1275,1276,1282],{},[35,1277,1278,1281],{},[38,1279,1280],{},"BrierLM",": perplexity does not apply when there is no likelihood. CALM uses a Brier score (accuracy minus collision rate) as its yardstick, which correlates strongly and negatively with cross-entropy loss.",[35,1283,1284,1287],{},[38,1285,1286],{},"Black-box temperature",": with no probability denominator to tune, batch size N becomes the temperature knob. Small N is cold and deterministic, large N is hot and diverse. With rejection sampling, it behaves much like conventional temperature.",[12,1289,1291],{"id":1290},"the-loop","The loop",[1293,1294,1295,1298,1301,1304,1307],"ol",{},[35,1296,1297],{},"Input: discrete tokens",[35,1299,1300],{},"Compress: the encoder turns every K tokens into one vector",[35,1302,1303],{},"Process: the transformer runs attention over a much shorter sequence",[35,1305,1306],{},"Predict: the head emits the next continuous vector in one step",[35,1308,1309],{},"Restore: the decoder expands it back into K tokens",[12,1311,1313],{"id":1312},"results","Results",[32,1315,1316,1319,1325,1332],{},[35,1317,1318],{},"With K=4: training compute down 44%, inference compute down 34%, generation steps from T to T\u002F4",[35,1320,1321,1324],{},[38,1322,1323],{},"Larger but cheaper",": 371M parameters (slightly larger than the baseline) with equal or better quality and less total compute",[35,1326,1327,1328,1331],{},"A new scaling axis: ",[38,1329,1330],{},"information density",". As semantic bandwidth grows, CALM reaches a higher Brier score at lower compute — pushing past the old compute–quality frontier",[35,1333,1334],{},"An unusual training curve: performance climbs sharply after ~200k steps and then overtakes conventional models. Learning continuous structure seems to need a long runway before it pays off",[12,1336,1338],{"id":1337},"my-take","My take",[514,1340,1341,1342,1345],{},"The framing is the contribution: ",[38,1343,1344],{},"the vocabulary was never the point",". Discrete tokens are a compression artifact of text, and once you accept that, predicting a continuous vector is the obvious next move.",[514,1347,1348],{},"Two things to watch before believing it replaces anything:",[1293,1350,1351,1357],{},[35,1352,1353,1356],{},[38,1354,1355],{},"Reconstruction fidelity is a hard ceiling."," 99.9% sounds great until you multiply it across a long document.",[35,1358,1359,1362],{},[38,1360,1361],{},"Sampling control moves to a different knob."," Batch-size-as-temperature works, but it means your serving stack has to be redesigned around it.",[514,1364,1365],{},"For deployment on constrained hardware, a 4× reduction in generation steps is the most interesting number here — that is the kind of win that changes what fits on a device.",{"title":635,"searchDepth":636,"depth":636,"links":1367},[1368,1369,1370,1371,1372,1373],{"id":1135,"depth":636,"text":1136},{"id":1162,"depth":636,"text":1163},{"id":1236,"depth":636,"text":1237},{"id":1290,"depth":636,"text":1291},{"id":1312,"depth":636,"text":1313},{"id":1337,"depth":636,"text":1338},"\u002Fimages\u002Fblog\u002F2510.27688v1.webp","Continuous autoregressive language models (CALM)","2026-08-06","CALM compresses K discrete tokens into one continuous vector, cutting generation steps to 1\u002FK. Notes on the autoencoder, the likelihood-free head, and what it actually buys you.",{},"\u002Fblog\u002Fen\u002Farxiv-2510-27688v1",{"title":1117,"description":1377},"arxiv-2510-27688v1","blog\u002Fen\u002Farxiv-2510-27688v1",[673,1384,1385],"llm","efficiency","6E96cFhns3BbhpqUak9RrqhAZpia6m0Go_yTYpgqN2g",{"id":1388,"title":1389,"author":7,"body":1390,"cover":2753,"coverAlt":2754,"date":2755,"description":2756,"draft":663,"extension":664,"featured":663,"lang":1011,"meta":2757,"navigation":667,"path":2758,"seo":2759,"slug":2760,"stem":2761,"tags":2762,"translationKey":2760,"updated":676,"__hash__":2765},"blog\u002Fblog\u002Fen\u002Farxiv-2005-03448v3.md","Discovering governing equations from scarce data with PiDL",{"type":9,"value":1391,"toc":2745},[1392,1408,1412,1419,1422,1433,1439,1443,1446,1449,1453,1456,2189,2194,2205,2210,2232,2299,2301,2374,2377,2380,2711,2715,2721,2727,2733,2735,2742],[514,1393,1394,688,1396,1401,1403,1404,1407],{},[38,1395,687],{},[17,1397,1400],{"href":1398,"rel":1399},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2005.03448",[21],"arXiv:2005.03448",[38,1402,695],{}," — Physics-informed learning of governing equations from scarce data\n",[38,1405,1406],{},"Authors"," — Zhao Chen¹, Yang Liu², Hao Sun¹˒³\n¹ Northeastern University, Civil & Environmental Engineering\n² Northeastern University, Mechanical & Industrial Engineering\n³ MIT, Civil & Environmental Engineering",[12,1409,1411],{"id":1410},"abstract-in-short","Abstract, in short",[514,1413,1414,1415,1418],{},"Recovering the governing equations of a complex physical system from data is what turns observations into a model. This paper proposes PiDL, a physics-informed deep learning architecture that discovers the governing PDEs of nonlinear spatio-temporal systems from ",[38,1416,1417],{},"scarce and noisy"," data.",[514,1420,1421],{},"It fuses four things: the representation power of deep neural networks, physical embedding, automatic differentiation, and sparse regression. Three goals:",[1293,1423,1424,1427,1430],{},[35,1425,1426],{},"Approximate the solution of the system state",[35,1428,1429],{},"Compute the relevant derivative terms",[35,1431,1432],{},"Identify which terms and coefficients actually form the PDE",[514,1434,1435,1438],{},[38,1436,1437],{},"Keywords"," — governing equation discovery, physics-informed deep learning, PDEs",[12,1440,1442],{"id":1441},"why-the-problem-is-hard","Why the problem is hard",[514,1444,1445],{},"Classical PDE modelling relies on first principles — conservation laws and the like. But for many complex systems, the analytical form is simply unknown.",[514,1447,1448],{},"Data-driven methods such as SINDy are limited by data quality and quantity, and numerical differentiation is notoriously sensitive to noise. Ask for derivatives from noisy, sparse measurements and the noise gets amplified into the coefficients you are trying to identify.",[12,1450,1452],{"id":1451},"the-pidl-framework","The PiDL framework",[514,1454,1455],{},"The loss combines data fit, physics residual, and sparsity:",[1457,1458,1461],"span",{"className":1459},[1460],"katex-display",[1457,1462,1465,1582],{"className":1463},[1464],"katex",[1457,1466,1469],{"className":1467},[1468],"katex-mathml",[1470,1471,1474],"math",{"xmlns":1472,"display":1473},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML","block",[1475,1476,1477,1577],"semantics",{},[1478,1479,1480,1485,1490,1493,1497,1501,1504,1507,1527,1530,1533,1548,1550,1553],"mrow",{},[1481,1482,1484],"mi",{"mathvariant":1483},"script","L",[1486,1487,1489],"mo",{"stretchy":1488},"false","(",[1481,1491,1492],{},"θ",[1486,1494,1496],{"separator":1495},"true",",",[1481,1498,1500],{"mathvariant":1499},"normal","Λ",[1486,1502,1503],{"stretchy":1488},")",[1486,1505,1506],{},"=",[1508,1509,1510,1523],"munder",{},[1508,1511,1512,1520],{},[1513,1514,1515,1517],"msub",{},[1481,1516,1484],{"mathvariant":1483},[1481,1518,1519],{},"d",[1486,1521,1522],{"stretchy":1495},"⏟",[1524,1525,1526],"mtext",{},"data",[1486,1528,1529],{},"+",[1481,1531,1532],{},"α",[1508,1534,1535,1545],{},[1508,1536,1537,1543],{},[1513,1538,1539,1541],{},[1481,1540,1484],{"mathvariant":1483},[1481,1542,514],{},[1486,1544,1522],{"stretchy":1495},[1524,1546,1547],{},"physics",[1486,1549,1529],{},[1481,1551,1552],{},"β",[1508,1554,1555,1574],{},[1508,1556,1557,1572],{},[1478,1558,1559,1562,1564],{},[1481,1560,1561],{"mathvariant":1499},"∥",[1481,1563,1500],{"mathvariant":1499},[1513,1565,1566,1568],{},[1481,1567,1561],{"mathvariant":1499},[1569,1570,1571],"mn",{},"0",[1486,1573,1522],{"stretchy":1495},[1524,1575,1576],{},"sparsity",[1578,1579,1581],"annotation",{"encoding":1580},"application\u002Fx-tex","\\mathcal{L}(\\theta, \\Lambda) = \\underbrace{\\mathcal{L}_d}_{\\text{data}} + \\alpha \\underbrace{\\mathcal{L}_p}_{\\text{physics}} + \\beta \\underbrace{\\|\\Lambda\\|_0}_{\\text{sparsity}}",[1457,1583,1586,1636,1844,2019],{"className":1584,"ariaHidden":1495},[1585],"katex-html",[1457,1587,1590,1595,1600,1604,1609,1613,1618,1621,1625,1629,1633],{"className":1588},[1589],"base",[1457,1591],{"className":1592,"style":1594},[1593],"strut","height:1em;vertical-align:-0.25em;",[1457,1596,1484],{"className":1597},[1598,1599],"mord","mathcal",[1457,1601,1489],{"className":1602},[1603],"mopen",[1457,1605,1492],{"className":1606,"style":1608},[1598,1607],"mathnormal","margin-right:0.0278em;",[1457,1610,1496],{"className":1611},[1612],"mpunct",[1457,1614],{"className":1615,"style":1617},[1616],"mspace","margin-right:0.1667em;",[1457,1619,1500],{"className":1620},[1598],[1457,1622,1503],{"className":1623},[1624],"mclose",[1457,1626],{"className":1627,"style":1628},[1616],"margin-right:0.2778em;",[1457,1630,1506],{"className":1631},[1632],"mrel",[1457,1634],{"className":1635,"style":1628},[1616],[1457,1637,1639,1643,1833,1837,1841],{"className":1638},[1589],[1457,1640],{"className":1641,"style":1642},[1593],"height:2.1674em;vertical-align:-1.4841em;",[1457,1644,1647],{"className":1645},[1646,1508],"minner",[1457,1648,1652,1824],{"className":1649},[1650,1651],"vlist-t","vlist-t2",[1457,1653,1656,1821],{"className":1654},[1655],"vlist-r",[1457,1657,1661,1686],{"className":1658,"style":1660},[1659],"vlist","height:0.6833em;",[1457,1662,1664,1669],{"style":1663},"top:-1.5159em;",[1457,1665],{"className":1666,"style":1668},[1667],"pstrut","height:3em;",[1457,1670,1676],{"className":1671},[1672,1673,1674,1675],"sizing","reset-size6","size3","mtight",[1457,1677,1679],{"className":1678},[1598,1675],[1457,1680,1683],{"className":1681},[1598,1682,1675],"text",[1457,1684,1526],{"className":1685},[1598,1675],[1457,1687,1689,1692],{"style":1688},"top:-3em;",[1457,1690],{"className":1691,"style":1668},[1667],[1457,1693,1695],{"className":1694},[1646,1508],[1457,1696,1698,1812],{"className":1697},[1650,1651],[1457,1699,1701,1809],{"className":1700},[1655],[1457,1702,1704,1754],{"className":1703,"style":1660},[1659],[1457,1705,1709,1712],{"className":1706,"style":1708},[1707],"svg-align","top:-2.202em;",[1457,1710],{"className":1711,"style":1668},[1667],[1457,1713,1717,1734,1744],{"className":1714,"style":1716},[1715],"stretchy","height:0.548em;min-width:1.6em;",[1457,1718,1722],{"className":1719,"style":1721},[1720],"brace-left","height:0.548em;",[1723,1724,1730],"svg",{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1729},"http:\u002F\u002Fwww.w3.org\u002F2000\u002Fsvg","400em","0.548em","0 0 400000 548","xMinYMin slice",[1731,1732],"path",{"d":1733},"M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13\n 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688\n 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7\n-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z",[1457,1735,1738],{"className":1736,"style":1721},[1737],"brace-center",[1723,1739,1741],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1740},"xMidYMin slice",[1731,1742],{"d":1743},"M199572 214\nc100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14\n 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3\n 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0\n-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z",[1457,1745,1748],{"className":1746,"style":1721},[1747],"brace-right",[1723,1749,1751],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1750},"xMaxYMin slice",[1731,1752],{"d":1753},"M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3\n 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237\n-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z",[1457,1755,1756,1759],{"style":1688},[1457,1757],{"className":1758,"style":1668},[1667],[1457,1760,1762],{"className":1761},[1598],[1457,1763,1765,1768],{"className":1764},[1598],[1457,1766,1484],{"className":1767},[1598,1599],[1457,1769,1772],{"className":1770},[1771],"msupsub",[1457,1773,1775,1800],{"className":1774},[1650,1651],[1457,1776,1778,1795],{"className":1777},[1655],[1457,1779,1782],{"className":1780,"style":1781},[1659],"height:0.3361em;",[1457,1783,1785,1789],{"style":1784},"top:-2.55em;margin-left:0em;margin-right:0.05em;",[1457,1786],{"className":1787,"style":1788},[1667],"height:2.7em;",[1457,1790,1792],{"className":1791},[1672,1673,1674,1675],[1457,1793,1519],{"className":1794},[1598,1607,1675],[1457,1796,1799],{"className":1797},[1798],"vlist-s","​",[1457,1801,1803],{"className":1802},[1655],[1457,1804,1807],{"className":1805,"style":1806},[1659],"height:0.15em;",[1457,1808],{},[1457,1810,1799],{"className":1811},[1798],[1457,1813,1815],{"className":1814},[1655],[1457,1816,1819],{"className":1817,"style":1818},[1659],"height:0.798em;",[1457,1820],{},[1457,1822,1799],{"className":1823},[1798],[1457,1825,1827],{"className":1826},[1655],[1457,1828,1831],{"className":1829,"style":1830},[1659],"height:1.4841em;",[1457,1832],{},[1457,1834],{"className":1835,"style":1836},[1616],"margin-right:0.2222em;",[1457,1838,1529],{"className":1839},[1840],"mbin",[1457,1842],{"className":1843,"style":1836},[1616],[1457,1845,1847,1851,1855,1858,2010,2013,2016],{"className":1846},[1589],[1457,1848],{"className":1849,"style":1850},[1593],"height:2.4397em;vertical-align:-1.7563em;",[1457,1852,1532],{"className":1853,"style":1854},[1598,1607],"margin-right:0.0037em;",[1457,1856],{"className":1857,"style":1617},[1616],[1457,1859,1861],{"className":1860},[1646,1508],[1457,1862,1864,2001],{"className":1863},[1650,1651],[1457,1865,1867,1998],{"className":1866},[1655],[1457,1868,1870,1888],{"className":1869,"style":1660},[1659],[1457,1871,1873,1876],{"style":1872},"top:-1.3798em;",[1457,1874],{"className":1875,"style":1668},[1667],[1457,1877,1879],{"className":1878},[1672,1673,1674,1675],[1457,1880,1882],{"className":1881},[1598,1675],[1457,1883,1885],{"className":1884},[1598,1682,1675],[1457,1886,1547],{"className":1887},[1598,1675],[1457,1889,1890,1893],{"style":1688},[1457,1891],{"className":1892,"style":1668},[1667],[1457,1894,1896],{"className":1895},[1646,1508],[1457,1897,1899,1989],{"className":1898},[1650,1651],[1457,1900,1902,1986],{"className":1901},[1655],[1457,1903,1905,1936],{"className":1904,"style":1660},[1659],[1457,1906,1909,1912],{"className":1907,"style":1908},[1707],"top:-2.0659em;",[1457,1910],{"className":1911,"style":1668},[1667],[1457,1913,1915,1922,1929],{"className":1914,"style":1716},[1715],[1457,1916,1918],{"className":1917,"style":1721},[1720],[1723,1919,1920],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1729},[1731,1921],{"d":1733},[1457,1923,1925],{"className":1924,"style":1721},[1737],[1723,1926,1927],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1740},[1731,1928],{"d":1743},[1457,1930,1932],{"className":1931,"style":1721},[1747],[1723,1933,1934],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1750},[1731,1935],{"d":1753},[1457,1937,1938,1941],{"style":1688},[1457,1939],{"className":1940,"style":1668},[1667],[1457,1942,1944],{"className":1943},[1598],[1457,1945,1947,1950],{"className":1946},[1598],[1457,1948,1484],{"className":1949},[1598,1599],[1457,1951,1953],{"className":1952},[1771],[1457,1954,1956,1977],{"className":1955},[1650,1651],[1457,1957,1959,1974],{"className":1958},[1655],[1457,1960,1963],{"className":1961,"style":1962},[1659],"height:0.1514em;",[1457,1964,1965,1968],{"style":1784},[1457,1966],{"className":1967,"style":1788},[1667],[1457,1969,1971],{"className":1970},[1672,1673,1674,1675],[1457,1972,514],{"className":1973},[1598,1607,1675],[1457,1975,1799],{"className":1976},[1798],[1457,1978,1980],{"className":1979},[1655],[1457,1981,1984],{"className":1982,"style":1983},[1659],"height:0.2861em;",[1457,1985],{},[1457,1987,1799],{"className":1988},[1798],[1457,1990,1992],{"className":1991},[1655],[1457,1993,1996],{"className":1994,"style":1995},[1659],"height:0.9341em;",[1457,1997],{},[1457,1999,1799],{"className":2000},[1798],[1457,2002,2004],{"className":2003},[1655],[1457,2005,2008],{"className":2006,"style":2007},[1659],"height:1.7563em;",[1457,2009],{},[1457,2011],{"className":2012,"style":1836},[1616],[1457,2014,1529],{"className":2015},[1840],[1457,2017],{"className":2018,"style":1836},[1616],[1457,2020,2022,2026,2030,2033],{"className":2021},[1589],[1457,2023],{"className":2024,"style":2025},[1593],"height:2.4516em;vertical-align:-1.7016em;",[1457,2027,1552],{"className":2028,"style":2029},[1598,1607],"margin-right:0.0528em;",[1457,2031],{"className":2032,"style":1617},[1616],[1457,2034,2036],{"className":2035},[1646,1508],[1457,2037,2039,2180],{"className":2038},[1650,1651],[1457,2040,2042,2177],{"className":2041},[1655],[1457,2043,2046,2064],{"className":2044,"style":2045},[1659],"height:0.75em;",[1457,2047,2049,2052],{"style":2048},"top:-1.4345em;",[1457,2050],{"className":2051,"style":1668},[1667],[1457,2053,2055],{"className":2054},[1672,1673,1674,1675],[1457,2056,2058],{"className":2057},[1598,1675],[1457,2059,2061],{"className":2060},[1598,1682,1675],[1457,2062,1576],{"className":2063},[1598,1675],[1457,2065,2066,2069],{"style":1688},[1457,2067],{"className":2068,"style":1668},[1667],[1457,2070,2072],{"className":2071},[1646,1508],[1457,2073,2075,2168],{"className":2074},[1650,1651],[1457,2076,2078,2165],{"className":2077},[1655],[1457,2079,2081,2112],{"className":2080,"style":2045},[1659],[1457,2082,2085,2088],{"className":2083,"style":2084},[1707],"top:-2.102em;",[1457,2086],{"className":2087,"style":1668},[1667],[1457,2089,2091,2098,2105],{"className":2090,"style":1716},[1715],[1457,2092,2094],{"className":2093,"style":1721},[1720],[1723,2095,2096],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1729},[1731,2097],{"d":1733},[1457,2099,2101],{"className":2100,"style":1721},[1737],[1723,2102,2103],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1740},[1731,2104],{"d":1743},[1457,2106,2108],{"className":2107,"style":1721},[1747],[1723,2109,2110],{"xmlns":1725,"width":1726,"height":1727,"viewBox":1728,"preserveAspectRatio":1750},[1731,2111],{"d":1753},[1457,2113,2114,2117],{"style":1688},[1457,2115],{"className":2116,"style":1668},[1667],[1457,2118,2120,2124],{"className":2119},[1598],[1457,2121,2123],{"className":2122},[1598],"∥Λ",[1457,2125,2127,2130],{"className":2126},[1598],[1457,2128,1561],{"className":2129},[1598],[1457,2131,2133],{"className":2132},[1771],[1457,2134,2136,2157],{"className":2135},[1650,1651],[1457,2137,2139,2154],{"className":2138},[1655],[1457,2140,2143],{"className":2141,"style":2142},[1659],"height:0.3011em;",[1457,2144,2145,2148],{"style":1784},[1457,2146],{"className":2147,"style":1788},[1667],[1457,2149,2151],{"className":2150},[1672,1673,1674,1675],[1457,2152,1571],{"className":2153},[1598,1675],[1457,2155,1799],{"className":2156},[1798],[1457,2158,2160],{"className":2159},[1655],[1457,2161,2163],{"className":2162,"style":1806},[1659],[1457,2164],{},[1457,2166,1799],{"className":2167},[1798],[1457,2169,2171],{"className":2170},[1655],[1457,2172,2175],{"className":2173,"style":2174},[1659],"height:0.898em;",[1457,2176],{},[1457,2178,1799],{"className":2179},[1798],[1457,2181,2183],{"className":2182},[1655],[1457,2184,2187],{"className":2185,"style":2186},[1659],"height:1.7016em;",[1457,2188],{},[514,2190,2191],{},[38,2192,2193],{},"How the pieces fit together",[32,2195,2196,2199,2202],{},[35,2197,2198],{},"A DNN learns the representation of the solution",[35,2200,2201],{},"Automatic differentiation computes derivatives accurately",[35,2203,2204],{},"ℓ₀ sparse regression picks out the governing terms",[514,2206,2207],{},[38,2208,2209],{},"Architecture",[32,2211,2212,2218],{},[35,2213,2214,2217],{},[1175,2215,2216],{},"Single dataset",": the DNN directly models the solution function",[35,2219,2220,2223,2224],{},[1175,2221,2222],{},"Multiple datasets",": a \"root-branch\" DNN handles different initial\u002Fboundary conditions\n",[32,2225,2226,2229],{},[35,2227,2228],{},"a shared root network captures common features",[35,2230,2231],{},"independent branch networks adapt to each condition",[514,2233,2234,2237,2238,2268,2269,2298],{},[38,2235,2236],{},"Optimisation (ADO)"," — alternating direction optimisation trains the network parameters ",[1457,2239,2241,2255],{"className":2240},[1464],[1457,2242,2244],{"className":2243},[1468],[1470,2245,2246],{"xmlns":1472},[1475,2247,2248,2252],{},[1478,2249,2250],{},[1481,2251,1492],{},[1578,2253,2254],{"encoding":1580},"\\theta",[1457,2256,2258],{"className":2257,"ariaHidden":1495},[1585],[1457,2259,2261,2265],{"className":2260},[1589],[1457,2262],{"className":2263,"style":2264},[1593],"height:0.6944em;",[1457,2266,1492],{"className":2267,"style":1608},[1598,1607]," and the PDE coefficients ",[1457,2270,2272,2286],{"className":2271},[1464],[1457,2273,2275],{"className":2274},[1468],[1470,2276,2277],{"xmlns":1472},[1475,2278,2279,2283],{},[1478,2280,2281],{},[1481,2282,1500],{"mathvariant":1499},[1578,2284,2285],{"encoding":1580},"\\Lambda",[1457,2287,2289],{"className":2288,"ariaHidden":1495},[1585],[1457,2290,2292,2295],{"className":2291},[1589],[1457,2293],{"className":2294,"style":1660},[1593],[1457,2296,1500],{"className":2297},[1598]," in turn, which sidesteps the NP-hard ℓ₀ problem.",[12,2300,1313],{"id":1312},[81,2302,2303,2319],{},[84,2304,2305],{},[87,2306,2307,2310,2313,2316],{},[90,2308,2309],{},"PDE",[90,2311,2312],{},"Noise",[90,2314,2315],{},"Coefficient error",[90,2317,2318],{},"Data used",[103,2320,2321,2335,2348,2361],{},[87,2322,2323,2326,2329,2332],{},[108,2324,2325],{},"Burgers",[108,2327,2328],{},"10%",[108,2330,2331],{},"1.15%",[108,2333,2334],{},"1.95%",[87,2336,2337,2340,2342,2345],{},[108,2338,2339],{},"Kuramoto–Sivashinsky (chaotic)",[108,2341,2328],{},[108,2343,2344],{},"0.71%",[108,2346,2347],{},"12.3%",[87,2349,2350,2353,2355,2358],{},[108,2351,2352],{},"Navier–Stokes",[108,2354,2328],{},[108,2356,2357],{},"1.40%",[108,2359,2360],{},"0.22%",[87,2362,2363,2366,2368,2371],{},[108,2364,2365],{},"Schrödinger",[108,2367,2328],{},[108,2369,2370],{},"2.31%",[108,2372,2373],{},"37.5%",[514,2375,2376],{},"Multiple-dataset settings were tested on Burgers with small-viscosity shock waves and on the FitzHugh–Nagumo neuron model.",[514,2378,2379],{},"The real-world test is the interesting one: on a cell migration and proliferation experiment, the method recovered the Fisher–Kolmogorov model,",[1457,2381,2383],{"className":2382},[1460],[1457,2384,2386,2450],{"className":2385},[1464],[1457,2387,2389],{"className":2388},[1468],[1470,2390,2391],{"xmlns":1472,"display":1473},[1475,2392,2393,2447],{},[1478,2394,2395,2403,2405,2408,2419,2421,2429,2431,2433,2440],{},[1513,2396,2397,2400],{},[1481,2398,2399],{},"ρ",[1481,2401,2402],{},"t",[1486,2404,1506],{},[1481,2406,2407],{},"γ",[1513,2409,2410,2412],{},[1481,2411,2399],{},[1478,2413,2414,2417],{},[1481,2415,2416],{},"x",[1481,2418,2416],{},[1486,2420,1529],{},[1513,2422,2423,2426],{},[1481,2424,2425],{},"λ",[1569,2427,2428],{},"1",[1481,2430,2399],{},[1486,2432,1529],{},[1513,2434,2435,2437],{},[1481,2436,2425],{},[1569,2438,2439],{},"2",[2441,2442,2443,2445],"msup",{},[1481,2444,2399],{},[1569,2446,2439],{},[1578,2448,2449],{"encoding":1580},"\\rho_t = \\gamma \\rho_{xx} + \\lambda_1 \\rho + \\lambda_2 \\rho^2",[1457,2451,2453,2510,2574,2633],{"className":2452,"ariaHidden":1495},[1585],[1457,2454,2456,2460,2501,2504,2507],{"className":2455},[1589],[1457,2457],{"className":2458,"style":2459},[1593],"height:0.625em;vertical-align:-0.1944em;",[1457,2461,2463,2466],{"className":2462},[1598],[1457,2464,2399],{"className":2465},[1598,1607],[1457,2467,2469],{"className":2468},[1771],[1457,2470,2472,2493],{"className":2471},[1650,1651],[1457,2473,2475,2490],{"className":2474},[1655],[1457,2476,2479],{"className":2477,"style":2478},[1659],"height:0.2806em;",[1457,2480,2481,2484],{"style":1784},[1457,2482],{"className":2483,"style":1788},[1667],[1457,2485,2487],{"className":2486},[1672,1673,1674,1675],[1457,2488,2402],{"className":2489},[1598,1607,1675],[1457,2491,1799],{"className":2492},[1798],[1457,2494,2496],{"className":2495},[1655],[1457,2497,2499],{"className":2498,"style":1806},[1659],[1457,2500],{},[1457,2502],{"className":2503,"style":1628},[1616],[1457,2505,1506],{"className":2506},[1632],[1457,2508],{"className":2509,"style":1628},[1616],[1457,2511,2513,2517,2521,2565,2568,2571],{"className":2512},[1589],[1457,2514],{"className":2515,"style":2516},[1593],"height:0.7778em;vertical-align:-0.1944em;",[1457,2518,2407],{"className":2519,"style":2520},[1598,1607],"margin-right:0.0556em;",[1457,2522,2524,2527],{"className":2523},[1598],[1457,2525,2399],{"className":2526},[1598,1607],[1457,2528,2530],{"className":2529},[1771],[1457,2531,2533,2557],{"className":2532},[1650,1651],[1457,2534,2536,2554],{"className":2535},[1655],[1457,2537,2539],{"className":2538,"style":1962},[1659],[1457,2540,2541,2544],{"style":1784},[1457,2542],{"className":2543,"style":1788},[1667],[1457,2545,2547],{"className":2546},[1672,1673,1674,1675],[1457,2548,2550],{"className":2549},[1598,1675],[1457,2551,2553],{"className":2552},[1598,1607,1675],"xx",[1457,2555,1799],{"className":2556},[1798],[1457,2558,2560],{"className":2559},[1655],[1457,2561,2563],{"className":2562,"style":1806},[1659],[1457,2564],{},[1457,2566],{"className":2567,"style":1836},[1616],[1457,2569,1529],{"className":2570},[1840],[1457,2572],{"className":2573,"style":1836},[1616],[1457,2575,2577,2581,2621,2624,2627,2630],{"className":2576},[1589],[1457,2578],{"className":2579,"style":2580},[1593],"height:0.8889em;vertical-align:-0.1944em;",[1457,2582,2584,2587],{"className":2583},[1598],[1457,2585,2425],{"className":2586},[1598,1607],[1457,2588,2590],{"className":2589},[1771],[1457,2591,2593,2613],{"className":2592},[1650,1651],[1457,2594,2596,2610],{"className":2595},[1655],[1457,2597,2599],{"className":2598,"style":2142},[1659],[1457,2600,2601,2604],{"style":1784},[1457,2602],{"className":2603,"style":1788},[1667],[1457,2605,2607],{"className":2606},[1672,1673,1674,1675],[1457,2608,2428],{"className":2609},[1598,1675],[1457,2611,1799],{"className":2612},[1798],[1457,2614,2616],{"className":2615},[1655],[1457,2617,2619],{"className":2618,"style":1806},[1659],[1457,2620],{},[1457,2622,2399],{"className":2623},[1598,1607],[1457,2625],{"className":2626,"style":1836},[1616],[1457,2628,1529],{"className":2629},[1840],[1457,2631],{"className":2632,"style":1836},[1616],[1457,2634,2636,2640,2680],{"className":2635},[1589],[1457,2637],{"className":2638,"style":2639},[1593],"height:1.0585em;vertical-align:-0.1944em;",[1457,2641,2643,2646],{"className":2642},[1598],[1457,2644,2425],{"className":2645},[1598,1607],[1457,2647,2649],{"className":2648},[1771],[1457,2650,2652,2672],{"className":2651},[1650,1651],[1457,2653,2655,2669],{"className":2654},[1655],[1457,2656,2658],{"className":2657,"style":2142},[1659],[1457,2659,2660,2663],{"style":1784},[1457,2661],{"className":2662,"style":1788},[1667],[1457,2664,2666],{"className":2665},[1672,1673,1674,1675],[1457,2667,2439],{"className":2668},[1598,1675],[1457,2670,1799],{"className":2671},[1798],[1457,2673,2675],{"className":2674},[1655],[1457,2676,2678],{"className":2677,"style":1806},[1659],[1457,2679],{},[1457,2681,2683,2686],{"className":2682},[1598],[1457,2684,2399],{"className":2685},[1598,1607],[1457,2687,2689],{"className":2688},[1771],[1457,2690,2692],{"className":2691},[1650],[1457,2693,2695],{"className":2694},[1655],[1457,2696,2699],{"className":2697,"style":2698},[1659],"height:0.8641em;",[1457,2700,2702,2705],{"style":2701},"top:-3.113em;margin-right:0.05em;",[1457,2703],{"className":2704,"style":1788},[1667],[1457,2706,2708],{"className":2707},[1672,1673,1674,1675],[1457,2709,2439],{"className":2710},[1598,1675],[12,2712,2714],{"id":2713},"why-it-matters","Why it matters",[514,2716,2717,2720],{},[38,2718,2719],{},"It breaks the data wall."," Accurate discovery under 10% noise with less than 1% of the original grid is the headline result.",[514,2722,2723,2726],{},[38,2724,2725],{},"It stays interpretable."," The DNN provides a global approximation, while the sparse PDE constrains the network.",[514,2728,2729,2732],{},[38,2730,2731],{},"It generalises to expensive domains"," — fluid mechanics, biological migration, climate prediction — where you can never collect a dense dataset.",[12,2734,1338],{"id":1337},[514,2736,2737,2738,2741],{},"The valuable idea here is not \"use a neural network\". It is the ",[38,2739,2740],{},"separation of concerns",": let the network own the function, let sparse regression own the equation, and let AD bridge the two. Each part is something you can reason about and check.",[514,2743,2744],{},"The weakness is the same one every physics-informed method shares: the library of candidate terms is a human decision. If the true governing term is not in the candidate set, no amount of optimisation will find it.",{"title":635,"searchDepth":636,"depth":636,"links":2746},[2747,2748,2749,2750,2751,2752],{"id":1410,"depth":636,"text":1411},{"id":1441,"depth":636,"text":1442},{"id":1451,"depth":636,"text":1452},{"id":1312,"depth":636,"text":1313},{"id":2713,"depth":636,"text":2714},{"id":1337,"depth":636,"text":1338},"\u002Fimages\u002Fblog\u002F2005.03448v3.webp","Physics-informed learning of governing equations from scarce data","2025-07-09","PiDL recovers the PDEs behind nonlinear spatio-temporal systems from sparse, noisy observations. Notes on its architecture, training strategy, and results on Burgers and Navier–Stokes.",{},"\u002Fblog\u002Fen\u002Farxiv-2005-03448v3",{"title":1389,"description":2756},"arxiv-2005-03448v3","blog\u002Fen\u002Farxiv-2005-03448v3",[673,2763,2764],"physics-informed","pde","RIokw_sIVtYhgrTpZy8_PrCcQPPaMadK3X9n8I9ivck",1792775653514]