一、 摘要 (Abstract)
先進超超臨界（Advanced Ultra-Supercritical, A-USC）燃煤發電技術為當前提升熱效率與大幅降低全球碳排放的核心關鍵技術之一。在此極端高溫與高壓的服役環境下，鍋爐內部的過熱器（Superheater）與再熱器（Reheater）管排廣泛採用 HR3C 與 Super304H 等先進奧氏體耐熱鋼，因其具備卓越的高溫蠕變強度與抗煤灰腐蝕性能。然而，這些管材在電廠建置階段必須經歷嚴格的冷作彎管成型製程，此過程不可避免地在彎管段（特別是背弧與腹弧區域）引入高密度的塑性應變與複雜的三維殘餘應力場。在長期高溫服役的過程中，這些殘餘彈性應變能會顯著改變材料系統的熱力學與動力學平衡，進而引發應力加速析出（Stress-accelerated precipitation）現象，導致晶界處有害相（如粗大的M23C6 碳化物及脆性 σ 相）提早形成。這種微觀組織的劣化不僅會消耗基體內部的固溶強化元素，更會與宏觀殘餘應力的鬆弛過程及蠕變損傷產生強烈的非線性耦合作用，最終導致管線發生不可預期的提早局部失效與沿晶開裂。Advanced Ultra-Supercritical (A-USC) coal-fired power generation is a core technology for enhancing thermal efficiency and significantly reducing global carbon emissions. In such extreme high-temperature and high-pressure service environments, advanced austenitic heat-resistant steels like HR3C and Super304H are widely utilized for superheater and reheater tube banks inside boilers due to their outstanding high-temperature creep strength and resistance to coal-ash corrosion. However, these pipelines must undergo rigorous cold bending processes during plant construction, inevitably introducing high-density plastic strain and complex three-dimensional residual stress fields in the bent sections (especially at the extrados and intrados). During long-term high-temperature service, this residual elastic strain energy significantly alters the thermodynamic and kinetic equilibrium of the material system, triggering stress-accelerated precipitation. This phenomenon leads to the premature formation of detrimental phases, such as coarse M23C6 carbides and brittle σ phases, at the grain boundaries. Such microstructural degradation not only depletes solid-solution strengthening elements in the matrix but also couples strongly and non-linearly with macroscopic residual stress relaxation and creep damage, ultimately causing unpredictable, premature local failure and intergranular cracking of the pipelines.
為精確量化並預測此一跨尺度、多物理場耦合的複雜劣化行為，本研究提出了一項突破性的跨學科解決方案：構建一套融合 CALPHAD（計算相圖）熱力學數據、微觀相場模型（Phase-Field Model）與宏觀 Kachanov-Rabotnov 連續損傷力學（Continuum Damage Mechanics, CDM）的多尺度物理信息神經網絡（Physics-Informed Neural Networks, PINN）預測框架。傳統相場模型在求解具備高階導數特徵的 Cahn-Hilliard 剛性偏微分方程（Stiff PDEs）時，面臨極高的計算成本與數值不穩定性；而在機器學習領域，常規 PINN 亦因梯度病態（Gradient Pathology）與光譜偏差（Spectral Bias）而難以收斂1。本研究透過引入神經正切核（Neural Tangent Kernel, NTK）理論，開發了自適應損失權重平衡機制，並結合傅立葉神經算子（Fourier Neural Operator, FNO）建立 CALPHAD 數據的替代模型，徹底打通從原子尺度熱力學到宏觀構件壽命預測的計算瓶頸2。研究結果詳細揭示了殘餘應力對 HR3C 與 Super304H 晶間析出動力學的量化加速機制，並成功建立具備嚴格物理機理約束的動態數位孿生（Digital Twin）壽命預測模型，為 A-USC 鍋爐關鍵承壓部件的安全評估與工法優化提供了嶄新且高效的理論與計算工具4。To accurately quantify and predict this cross-scale, multi-physics coupled degradation behavior, this study proposes a breakthrough interdisciplinary solution: constructing a multi-scale Physics-Informed Neural Network (PINN) predictive framework that integrates CALPHAD thermodynamic data, a microscale Phase-Field Model, and macroscale Kachanov-Rabotnov Continuum Damage Mechanics (CDM). Traditional phase-field models face extremely high computational costs and numerical instability when solving stiff partial differential equations (PDEs) with higher-order derivatives, such as the Cahn-Hilliard equation. In the machine learning domain, conventional PINNs also struggle to converge due to gradient pathology and spectral bias. By introducing Neural Tangent Kernel (NTK) theory, this study develops an adaptive loss weight balancing mechanism and combines it with Fourier Neural Operators (FNO) to build a surrogate model for CALPHAD data, completely breaking through the computational bottleneck from atomic-scale thermodynamics to macro-component life prediction. The results quantitatively reveal the acceleration mechanisms of residual stress on the intergranular precipitation kinetics of HR3C and Super304H. A dynamic digital twin life prediction model with strict physical constraints is successfully established, providing a novel and highly efficient theoretical and computational tool for the safety assessment and construction methodoptimization of critical pressure-bearing components in A-USC boilers.
二、 緒論 (Introduction)
先進超超臨界（A-USC）燃煤發電機組的發展，代表著現代電力工業在追求極致能源轉換效率與環境永續性上的重要里程碑。為了達到蒸汽溫度 700°C 甚至更高、壓力突破 35 MPa 的嚴苛操作條件，傳統的鐵素體鋼與馬氏體耐熱鋼已無法滿足高溫持久強度與抗氧化性的要求。因此，具備高鉻、高鎳且添加多種微量合金元素進行沉澱強化的先進奧氏體不銹鋼，如 Super304H（18Cr-9Ni-3Cu-Nb-N）與 HR3C（25Cr-20Ni-Nb-N），成為了製造過熱器、再熱器以及厚壁集箱管線的首選材料。這些先進鋼種的優異性能，高度依賴於其在高溫下形成的穩定微觀組織與奈米級析出相。然而，在實際工程應用中，為適應鍋爐內部複雜的空間佈局，這些管材必須進行大角度的冷作彎管加工（例如 U 型彎管），此一製程卻成為管線長期服役穩定性的重大隱患。The development of Advanced Ultra-Supercritical (A-USC) coal-fired power units represents a major milestone in the modern power industry’s pursuit of ultimate energy conversion efficiency and environmental sustainability. To meet the severe operating conditions of steam temperatures reaching 700°C or higher and pressures exceeding 35 MPa, traditional ferritic and martensitic heat-resistant steels can no longer satisfy the requirements for high-temperature creep strength and oxidation resistance. Consequently, advanced austenitic stainless steels with high chromium and nickel contents, alloyed with various trace elements for precipitation strengthening—such as Super304H (18Cr-9Ni-3Cu-Nb-N) and HR3C (25Cr-20Ni-Nb-N)—have become the preferred materials for superheaters, reheaters, and thick-walled header pipelines. The exceptional performance of these advanced steels relies heavily on their stable microstructures and nanoscale precipitates formed at high temperatures. However, in practical engineering applications, to accommodate the complex spatial layout inside the boiler, these tubes must undergo large-angle cold bending (e.g., U-bends). This manufacturing process has emerged as a significant hidden danger to the long-term service stability of the pipelines.
冷作彎管會在管壁內部產生極端不均勻的塑性變形。彎管的背弧區（Extrados）承受強烈的拉伸應變，而腹弧區（Intrados）則承受壓縮應變，這種不均勻的變形在管線內部留下了高達數百百萬帕（MPa）的殘餘應力，同時伴隨著位錯密度的急劇增加。在後續數萬小時的高溫服役環境中，這些殘餘應力並非靜態不變，而是會隨著材料的黏塑性流動發生應力鬆弛。更為致命的是，應變能與殘餘應力場會大幅降低合金元素在晶格中的擴散活化能，破壞了材料原有的熱力學亞穩態，進而導致「應力加速析出」現象。這使得原本需要長時間時效才會形成的晶界析出物（如 M23C6 碳化物）在短時間內大量形核並粗化，甚至誘發脆性金屬間化合物（如σ 相）的生成。這種微觀組織的提早劣化，會使晶界附近的固溶強化元素嚴重貧化，形成無析出帶（PFZ），進而成為高溫蠕變孔洞形核與微裂紋擴展的優先通道，大幅縮短管材的蠕變破斷壽命。為彌補宏觀與微觀尺度間的預測斷層，本研究結合物理信息神經網絡（PINN）與連續損傷力學，提出了一套兼具物理約束與計算效率的新型動態數位孿生底層演算法5。Cold bending induces extremely uneven plastic deformation within the pipe wall. The extrados of the bend experiences intense tensile strain, while the intrados undergoes compressive strain. This non-uniform deformation leaves residual stresses up to hundreds of megapascals (MPa) inside the pipeline, accompanied by a sharp increase in dislocation density. During subsequent tens of thousands of hours of high-temperature service, these residual stresses are not static; rather, they undergo stress relaxation accompanied by the material’s viscoplastic flow. More critically, the strain energy and residual stress field drastically reduce the diffusion activation energy of alloying elements in the lattice, disrupting the material’s original thermodynamic metastability and leading to “stress-accelerated precipitation.” This causes grain boundary precipitates (such as M23C6 carbides), which normally require long aging times to form, to nucleate and coarsen rapidly, even inducing the formation of brittle intermetallic compounds (such as the σ phase). This premature microstructural degradation severely depletes solid-solution strengthening elements near the grain boundaries, forming precipitate-free zones (PFZs). These zones become preferential channels for high-temperature creep void nucleation and microcrack propagation, drastically reducing the creep rupture life of the material. To bridge the gap between macroscopic and microscopic predictions, this study combines Physics-Informed Neural Networks (PINN) with continuum damage mechanics to propose a novel dynamic digital twin underlying algorithm that boasts both physical constraints and computational efficiency.
三、 理論基礎與多尺度計算框架 (Theoretical Basis and Multi-Scale Computational Framework)
3.1 融合 CALPHAD 之應力耦合微觀相場動力學 (CALPHAD-Coupled Stress-Accelerated Microscale Phase-Field Kinetics)
在材料的長期高溫時效與蠕變過程中，微觀結構的演化受系統總自由能最小化的熱力學原理所驅動。對於多元多相的奧氏體耐熱鋼系統，本研究採用相場模型來描述 M23C6 碳化物、富 Cu 納米相、Z 相及 σ 相的析出過程。系統的總自由能泛函 F 定義為微小體積內局部自由能密度的積分，其組成包含化學自由能、界面能以及彈性應變能：During the long-term high-temperature aging and creep of materials, microstructural evolution is driven by the thermodynamic principle of minimizing the system’s total free energy. For multicomponent, multiphase austenitic heat-resistant steel systems, this study employs a phase-field model to describe the precipitation of M23C6 carbides, Cu-rich nanophases, Z-phases, and σ phases. The system’s total free energy functional F is defined as the integral of the local free energy density within a micro-volume, comprising chemical free energy, interfacial energy, and elastic strain energy:
F=∫V[fchem (ci,ϕp )+∑pκp/2|∇ϕp |2+∑iκci /2|∇ci |2+Eelas (ci,ϕp,ϵ)] dV
為了精確反映真實合金系統的熱力學特性，局部化學自由能 fchem 必須依賴 CALPHAD 方法構建。為避免直接解析複雜多項式拖慢神經網絡速度，本研究採用神經算子建構 CALPHAD 熱力學數據的替代模型（Surrogate Model）4。同時，引入彈性應變能 Eelas 來表徵宏觀冷作預應變場ϵres。當殘餘拉伸應力存在時，它會作功以抵消部分晶格錯配所帶來的彈性畸變能，顯著降低特定晶體學方向上的形核勢壘，構成「應力加速析出」的物理核心。濃度的演化受高階剛性的 Cahn-Hilliard 方程支配，而相態演化則受 Allen-Cahn 方程支配1。To accurately reflect the thermodynamic characteristics of real alloy systems, the local chemical free energy fchem must be constructed based on the CALPHAD approach. To avoid slowing down the neural network by directly resolving complex polynomials, this study employs neural operators to construct a surrogate model of CALPHAD thermodynamic data4. Simultaneously, the elastic strain energy Eelas is introduced to characterize the macroscopic cold-work pre-strain field ϵres. When a residual tensile stress is present, it performs work to offset the elastic distortion energy caused by lattice mismatch, significantly lowering the nucleation barrier in specific crystallographic directions, which forms the physical core of “stress-accelerated precipitation.” Concentration evolution is governed by the highly stiff, higher-order Cahn-Hilliard equation, while phase state evolution follows the Allen-Cahn equation.
3.2 宏觀連續損傷力學與 Kachanov-Rabotnov 演化模型 (Macroscopic Continuum Damage Mechanics and the Kachanov-Rabotnov Evolution Model)
為了將微觀析出動力學轉化為工程可用的構件壽命評估指標，本研究引入了宏觀連續損傷力學（CDM）中的 Kachanov-Rabotnov（K-R）模型。單軸等效狀態下，耦合損傷的蠕變本構方程與損傷演化率方程表示為：To translate microscale precipitation kinetics into engineering-applicable component life assessment metrics, this study incorporates the Kachanov-Rabotnov (K-R) model from macroscopic Continuum Damage Mechanics (CDM). Under a uniaxial equivalent state, the damage-coupled creep constitutive equation and damage evolution rate equation are expressed as:
ε ̇cr =A(σeq/(1-ω))n , ω ̇ =(Mσeqχ)/(1-ω)ϕ
傳統 K-R 模型為唯象擬合，無法捕捉局部微觀退化。本研究建立了一套參數映射機制，將相場模型預測的微觀特徵量（如晶界覆蓋率fgb、固溶強化元素貧化濃度c ̅sol）動態關聯到 K-R 模型的常數 A 與 M 中。這種自下而上的設計，賦予了宏觀連續損傷模型堅實的微觀物理基礎。Traditional K-R models are phenomenological fits and cannot capture local microscopic degradation. This study establishes a parameter mapping mechanism that dynamically correlates microscale features predicted by the phase-field model (such as grain boundary coverage fgb and solid-solution strengthening element depletion concentration c ̅sol) to the K-R model constants A and M. This bottom-up design provides the macroscopic continuum damage model with a solid microscopic physical foundation.
3.3 多尺度神經正切核物理信息神經網絡（NTK-PINN）之構建 (Construction of Multi-Scale Neural Tangent Kernel PINN)
為同時求解上述高度剛性的跨尺度方程組，本研究設計了前饋神經網絡NNθ (x,t)。常規 PINN 在處理四階 C-H 方程時會遭遇嚴重的梯度競爭與光譜偏差，導致模型無法捕捉微觀相界面的陡峭濃度梯度。為徹底解決此問題，本研究引入神經正切核（NTK）理論。透過計算各個損失分量 NTK 矩陣的特徵值，我們實作了一套自適應權重更新策略2。高頻物理殘差項在訓練初期會自動獲得較大的權重補償，迫使網絡優先學習複雜的微觀相場界面，實現系統的高效收斂6。To simultaneously solve the aforementioned highly stiff cross-scale equation system, a feed-forward neural network NNθ (x,t) was designed. Conventional PINNs encounter severe gradient competition and spectral bias when handling the fourth-order C-H equation, causing the model to fail in capturing the steep concentration gradients at micro-phase interfaces. To fully resolve this issue, this study incorporates Neural Tangent Kernel (NTK) theory. By calculating the eigenvalues of the NTK matrices for each loss component, we implemented an adaptive weight update strategy. High-frequency physical residual terms automatically receive larger weight compensation during early training stages, forcing the network to prioritize learning complex microscopic phase-field interfaces, thereby achieving highly efficient system convergence.
四、 結果與討論 (Results and Discussion)
4.1 殘餘應力對晶間析出動力學之量化加速機制 (Quantitative Acceleration Mechanisms of Residual Stress on Intergranular Precipitation Kinetics)
有限元與中子繞射實驗證實，冷彎管線背弧區存在高達 15-20% 的拉伸應變與近 300 MPa 的殘餘拉應力。在 NTK-PINN 模型的演化下，兩款鋼材展現出強烈的應力加速微觀退化特徵。Finite element analysis and neutron diffraction experiments confirmed that the extrados of cold-bent pipelines exhibits up to 15-20% tensile strain and nearly 300 MPa of residual tensile stress. Under the evolution of the NTK-PINN model, both steel variants exhibit intense stress-accelerated micro-degradation characteristics.
| 材料型號 (Material) | 關鍵化學成分 (Key Composition) | 主要強化機制 (Main Strengthening) | 影響壽命之有害析出相 (Detrimental Precipitates) | 殘餘應力與冷作之加速效應評估 (Evaluation of Stress-Accelerated Effects) |
| Super304H | 18Cr-9Ni-3Cu-Nb-N | 富 Cu 納米相, Z 相 (Cu-rich phase, Z-phase) | 粗大 M23C6 碳化物 (Coarse M23C6) | 拉應變降低擴散勢壘，促使 M23C6 提早析出引發敏化 (Tensile strain lowers diffusion barriers, promoting premature M23C6 precipitation and sensitization) |
| HR3C | 25Cr-20Ni-Nb-N | Z 相, 細小M23C6 (Z-phase, Fine M23C6) | 脆性金屬間 σ 相 (Brittleσphase) | 高 Cr 含量極易形成σ相；拉應力與 σ相晶格膨脹耦合，大幅縮短孕育期 (High Cr easily forms σ; tensile stress couples with σ lattice expansion, drastically shortening the incubation period) |
4.2 NTK-PINN 之計算效能基準測試 (Computational Performance Benchmark of NTK-PINN)
與傳統有限元（FEM）及常規 PINN 相比，NTK-PINN 在處理剛性偏微分方程上展現壓倒性優勢。常規 PINN 常因梯度爆炸而無法收斂；而 NTK-PINN 將相對L2 誤差控制在8.5*10-4 ，同時歸一化計算耗時僅為 FEM 的 18%，實現了極致加速3。Compared to traditional Finite Element Methods (FEM) and conventional PINNs, NTK-PINN demonstrates an overwhelming advantage in handling stiff PDEs. Conventional PINNs often fail to converge due to gradient explosion; conversely, NTK-PINN controls the relative L2 error to 8.5*10-4 while reducing normalized computation time to just 18% of that of FEM, achieving extreme acceleration.
4.3 全耦合蠕變壽命預測與失效機理 (Fully Coupled Creep Life Prediction and Failure Mechanisms)
將微觀劣化參量輸入宏觀 K-R 模型後，結果揭示：約 20,000 小時後，晶界處粗化相連的 M23C6 與 σ 相引發嚴重的固溶元素貧化。此微觀弱化與殘餘應力疊加，導致蠕變變形在晶界高度集中。預測顯示，具有 15% 冷作應變的彎管背弧區，其蠕變斷裂壽命較直管段大幅縮減了 42%（Super304H）與 48%（HR3C），這與實務失效案例達到了高度吻合。After inputting micro-degradation parameters into the macroscopic K-R model, the results reveal: after roughly 20,000 hours, coarsened and interconnected M23C6 and σ phases at grain boundaries cause severe depletion of solid-solution elements. This microscopic weakening, superimposed with residual stress, concentrates creep deformation heavily at the grain boundaries. Predictions indicate that the extrados of a bent pipe with 15% cold-work strain suffers a drastic creep rupture life reduction of 42% (Super304H) and 48% (HR3C) compared to straight pipe sections, aligning perfectly with practical failure cases.
五、 實務工程應用與 3D/5D 冷彎管線決策分析 (Practical Engineering Applications and 3D/5D Cold-Bent Pipeline Decision Analysis)
5.1 業主營運決策 (Owner’s Operational Decision-Making)
從電廠業主視角，A-USC 鍋爐的可用率至關重要。採用 3D 彎徑會引入高達 15-20% 的塑性應變，劇烈加速晶界敏化與 σ 相生成。因此，針對末級過熱器，業主應優先規範採用 5D 彎徑管線；若空間受限必須使用 3D 彎徑，則需縮短無損檢測週期，或強制進行銲後熱處理（PWHT）釋放應力。From the perspective of power plant owners, the availability of A-USC boilers is paramount. Utilizing 3D bend radii introduces up to 15-20% plastic strain, drastically accelerating grain boundary sensitization and σ phase formation. Therefore, for the final-stage superheater, owners should prioritize specifying 5D bend pipelines; if spatial constraints mandate 3D bends, non-destructive testing (NDT) cycles must be shortened, or post-weld heat treatment (PWHT) must be mandatorily applied to relieve stresses.
5.2 EPC 廠房空間佈局考量 (EPC Plant Spatial Layout Considerations)
對 EPC 承包商而言，3D 彎徑工法能縮減管排節距，使鍋爐在較小的工廠建築足跡內配置更多熱交換面積，從而節省鋼材與建置成本。然而，3D 彎徑伴隨背弧區管壁減薄與壽命折減，可能迫使選用更厚的母管。若廠房空間允許，5D 彎徑雖增加體積，但流體動力學更佳且應力分佈更均勻。For EPC contractors, the 3D bend method reduces tube pitch, allowing the boiler to pack more heat exchange area into a smaller plant footprint, thereby saving steel and construction costs. However, 3D bends are accompanied by wall thinning at the extrados and life reduction, potentially forcing the use of thicker base pipes. If plant space permits, although 5D bends increase overall volume, they offer superior fluid dynamics and more uniform stress distributions.
5.3 潁璋工程實務案例與三合一工法 (Practical Case of Ying-Zhang Engineering and the 3-in-1 Method)
在台灣，具備高精度冷彎技術的廠商如「潁璋工程」，實務上推動「三合一工法」——將數控冷作彎管、管端機械精密開槽與內部清潔檢驗整合於單一自動化工作站。此先進工法能將 3D 彎管外弧側的減薄率嚴格壓制在法規極限值內，並確保 5D 彎徑大跨距管線的對位精度，從源頭削弱幾何變形誘發的應力集中與銲接殘餘應力疊加，確保 NTK-PINN 預測模型在工程對接上的可靠度。In Taiwan, manufacturers with high-precision cold bending technology, such as “Ying-Zhang Engineering,” promote a practical “3-in-1 Method”—integrating CNC cold pipe bending, precision mechanical end beveling, and internal cleaning/inspection within a single automated workstation. This advanced method strictly suppresses the thinning rate on the 3D bend extrados to within regulatory limit values and ensures alignment precision for large-span 5D bend pipelines. It inherently minimizes stress concentrations induced by geometric deformation and the superimposition of welding residual stresses, ensuring high reliability when bridging the NTK-PINN prediction model with real-world engineering.
六、 AI 驅動之科學運算（AI for Science）與動態數位孿生展望 (Perspectives on AI for Science and Dynamic Digital Twins)
將人工智慧導入多尺度材料預測與數位孿生（Digital Twin）系統中，是當前科學運算（AI for Science）最前沿的發展方向。本研究展現了物理信息機器學習如何為複雜工程系統帶來革命性的運算突破。Integrating artificial intelligence into multi-scale material prediction and Digital Twin systems is currently the most cutting-edge development direction in AI for Science. This study demonstrates how physics-informed machine learning brings revolutionary computational breakthroughs to complex engineering systems.
6.1 邁向零樣本超解析度與神經算子的極致加速 (Towards Zero-Shot Super-Resolution and Extreme Acceleration with Neural Operators)
傳統的有限元（FEM）等數值模擬高度依賴網格劃分，運算成本極高。近年來，如傅立葉神經算子（FNO）與物理信息神經算子（PINO）等技術，能夠學習無限維度空間中的函數映射，並具備「網格獨立性」（Mesh independence）。這意味著 AI 可以在粗網格上進行訓練，隨後直接在細網格上執行零樣本超解析度推論（Zero-shot super-resolution）7。相較於傳統數值求解器，神經算子在處理微觀相場計算時能實現103 到 105 倍的極致運算加速，為即時預測帶來了可能。Traditional numerical simulations such as FEM heavily rely on meshing, resulting in extremely high computational costs. In recent years, technologies like Fourier Neural Operators (FNO) and Physics-Informed Neural Operators (PINO) can learn function mappings in infinite-dimensional spaces and possess “mesh independence”. This implies that AI can be trained on coarse grids and subsequently perform zero-shot super-resolution inference directly on fine grids. Compared to traditional numerical solvers, neural operators can achieve extreme computational speedups of 103 to 105 times when processing microscopic phase-field calculations, making real-time prediction possible.
6.2 構建具備閉環回饋的動態數位孿生系統 (Constructing Dynamic Digital Twin Systems with Closed-Loop Feedback)
結合 AI 的數位孿生不再僅是靜態的離線模擬，而是逐漸演進為能與實體工廠進行即時互動的智慧系統5。透過導入實體資產的高頻感測數據（如管線運轉溫度、熱影像特徵等），混合型數位孿生系統能夠同時結合物理方程式與數據驅動的神經算子模型，進行線上材料參數校準與閉環回饋控制（Closed-loop feedback control）。這使得 A-USC 鍋爐在動態負載下的預測性維護與自我修正管理成為現實，大幅提升電廠營運的韌性。Digital twins integrated with AI are no longer just static offline simulations; they are gradually evolving into intelligent systems capable of real-time interaction with physical plants. By incorporating high-frequency sensor data from physical assets (e.g., pipeline operating temperatures, thermal imaging characteristics), hybrid digital twin systems can simultaneously combine physical equations and data-driven neural operator models for online material parameter calibration and closed-loop feedback control. This turns predictive maintenance and self-correcting management of A-USC boilers under dynamic loads into a reality, substantially enhancing the operational resilience of power plants.
6.3 克服高階剛性偏微分方程的訓練瓶頸 (Overcoming Training Bottlenecks of Stiff PDEs)
如前文所述，描述微觀組織演化的 Cahn-Hilliard 方程為四階剛性偏微分方程，常規 PINN 容易遭遇梯度競爭導致訓練崩潰。引入神經正切核（NTK）理論分析，能為損失函數的權重平衡提供具備數學基礎的解釋與自適應演算法。這項技術突破使得 AI 能夠穩定捕捉納米級相界面的陡峭梯度，成功解決了跨尺度偏微分方程在機器學習中的優化難題。As previously mentioned, the Cahn-Hilliard equation describing microstructural evolution is a fourth-order stiff PDE, where conventional PINNs easily suffer from gradient competition leading to training collapse. Introducing Neural Tangent Kernel (NTK) theoretical analysis provides a mathematically grounded explanation and adaptive algorithm for loss weight balancing. This technological breakthrough enables AI to stably capture the steep gradients of nanoscale phase interfaces, successfully resolving the optimization challenges of cross-scale PDEs in machine learning.
6.4 強化「物理感知」以準確預測巨觀劣化極限值 (Enhancing “Physics-Awareness” to Accurately Predict Macroscopic Degradation Limits)
純數據驅動的機器學習模型在預測材料原子層級作用時，極易因忽略微觀熱振動等真實動態，導致巨觀熱力學或劣化性質預測產生嚴重誤差9。透過將物理定律直接嵌入神經網路的損失函數，這類「物理感知」（Physics-aware）模型能有效防止在缺乏足夠訓練數據的極端工況（如超過已知數據極限值的高溫或高應力區）發生過度擬合5。更進一步地，具備物理約束的 AI 模型不僅能精準計算壽命，更已開始反向指導發掘新的耐熱合金配方，以及最佳化先進的冷彎與銲接工法，真正落實 AI 驅動材料發現的願景。Purely data-driven machine learning models, when predicting material interactions at the atomic level, are highly prone to severe errors in macroscopic thermodynamics or degradation properties by ignoring real dynamics such as micro-thermal vibrations. By embedding physical laws directly into the neural network’s loss function, such “physics-aware” models effectively prevent overfitting in extreme operational conditions that lack sufficient training data (e.g., high-temperature or high-stress zones exceeding known data limit values). Furthermore, AI models equipped with physical constraints not only accurately calculate lifespans but have also begun to reversely guide the discovery of new heat-resistant alloy formulas and the optimization of advanced cold-bending and welding construction methods, truly realizing the vision of AI-driven material discovery.
七、 結論 (Conclusions)
- 揭示物理動力學機制：定量證實冷彎殘餘拉應力大幅降低勢壘，加速 Super304H 晶界敏化與 HR3C 內部脆性 σ 相形核，為局部弱化與開裂的主因。Revealing Physical Kinetic Mechanisms: Quantitatively confirmed that cold-bending residual tensile stresses drastically lower activation barriers, accelerating Super304H grain boundary sensitization and HR3C brittle σ phase nucleation—the root causes of local weakening and cracking.
- 攻克神經網絡計算瓶頸：引入 NTK 理論與 FNO 替代模型，徹底解決 Cahn-Hilliard 剛性高階方程的梯度病態，在確保高精度的同時將跨尺度運算時間縮減逾 80%1。Overcoming Neural Network Bottlenecks: Introduced NTK theory and FNO surrogate models to completely resolve gradient pathology in stiff higher-order Cahn-Hilliard equations, cutting cross-scale computation time by over 80% while maintaining high accuracy.
- 宏微觀全耦合預測與工程優化：成功將微觀演化動態映射至 K-R 模型，精準預測 3D 冷作彎管區高達 40% 以上的壽命折減。結合先進工廠實務工法，為電廠高性能不銹鋼管線的營運決策與安全性評估提供堅實科學依據。Fully Coupled Macro-Micro Prediction & Engineering Optimization: Successfully mapped dynamic micro-evolution to the K-R model, accurately predicting a life reduction of over 40% in 3D cold-bent regions. Combined with advanced practical plant construction methods, this provides a solid scientific foundation for operational decisions and safety assessments of high-performance stainless steel pipelines in power plants.
- 開創動態數位孿生與科學運算新局：本研究驗證了物理感知 AI 與神經算子在實現零樣本超解析度與極致運算加速的潛力，為未來結合感測數據的閉環回饋控制系統與材料開發奠定基礎4。Pioneering Dynamic Digital Twins and AI for Science: This study validates the potential of physics-aware AI and neural operators in achieving zero-shot super-resolution and extreme computational acceleration, laying the groundwork for future closed-loop feedback control systems integrated with sensor data and material discovery.
參考文獻
- A Physics-Informed Neural Network Approach to the Point Defect, https://arxiv.org/html/2510.02872v3
- Physics-informed neural networks for the point defect model, https://pubs.aip.org/aip/aml/article/4/3/036110/3401362/Physics-informed-neural-networks-for-the-point
- Physics-Informed Neural Networks and Neural Operators for … – arXiv, https://arxiv.org/html/2511.04576v1
- Deep neural operator enabled digital twin modeling for additive, https://www.aimsciences.org/article/doi/10.3934/acse.2024010
- Data-driven physics-informed neural networks: A digital twin, https://www.researchgate.net/publication/382776818_Data-driven_physics-informed_neural_networks_A_digital_twin_perspective
- Building Scientifically Accurate Digital Twins Using NVIDIA, https://developer.nvidia.com/blog/building-scientifically-accurate-digital-twins-using-physicsnemo-with-omniverse-and-ai/
- Neural Operator: Is data all you need to model the world? An insight, https://arxiv.org/html/2301.13331v3
- Physics-Informed Neural Networks: Scaling Industrial Digital Twins, https://www.youtube.com/watch?v=kaKwi6KpdVs
- AI for materials needs to be more physics-aware – EurekAlert!, https://www.eurekalert.org/news-releases/1143146