《Materials & Design》:Atomistic insights into continuous and discontinuous precipitation in high-entropy alloys: Role of interfacial and elastic anisotropy
理解空间结构如何控制瞬态热响应,对于设计用于航空航天、生物医学和能源系统的功能梯度材料(FGMs)至关重要。特别是当材料梯度在径向和周向两个方向独立变化时——如许多中空结构那样——瞬态条件下的控制方程变得极为复杂。这种复杂性因周向变化的对流系数和时空变化的环境温度而进一步加剧,而这些问题在解析FGM研究中很少被涉及。为应对这一挑战,研究人员提出了一个统一的解析-机器学习框架,用于预测非对称梯度空心圆柱中的瞬态热行为。经典的傅里叶-斯特姆-刘维尔(Fourier–Sturm–Liouville)展开被推广至具有周向变化Robin条件的正交各向异性FGMs,产生了Toeplitz耦合的模态约束。所得系统通过投影器-伪逆(projector–pseudoinverse)公式精确求解,得到封闭形式的特征函数级数解。五种梯度架构——幂律、指数、Sigmoid和混合模型——被分析,揭示了从均匀加热到扩散热缓冲的热传递特性。应用实例展示了生物组织消融和核燃料包壳设计。最后,解析解和数值解被用作高保真训练数据,以改进多功能梯度增强的物理信息深度算子网络(gPI-DeepONet)中的算子学习和预测效率。与DeepONet相比,gPI-DeepONet、混合有限元分析(hybrid-FEA)和混合解析(hybrid-analytical)变体的误差分别降低了约、和。
功能梯度材料(FGMs)是具有空间变化特性的先进复合材料,能够在单一结构内实现受控的热力和机械响应。功能梯度的概念深受自然系统启发,其中微观结构变化是对非均匀环境的适应性响应。毛竹(Bamboo)是自然梯度生物复合材料的典型代表,其纤维密度、维管束分布和基质孔隙率沿径向和轴向系统变化,这种梯度结构增强了结构效率和载荷分布。此类仿生设计为工程材料中引入空间依赖的热物理性质提供了明确的物理基础。在热工程中,梯度配置对于承受高且时变热梯度的组件尤为宝贵,这些条件常见于航空航天系统、发电设备和热防护结构。涡轮叶片、火箭喷管、冷却管和压力容器等部件越来越多地采用圆柱形FGM设计,以在非稳态工况下维持性能。这些环境产生瞬态温度场和陡峭热梯度,显著影响热应力演化和结构可靠性,因此精确设计瞬态热传导对于FGM圆柱的可靠设计至关重要。
材料梯度与非稳态热载荷在FGMs中的耦合产生了复杂的瞬态热传导行为。准确预测这些响应对于可靠高效的设计至关重要,这需要能够捕捉材料梯度、几何复杂性和时间演化的模型。因此,已开发出众多计算框架来研究FGMs和多尺度材料中的热传导,包括有限元法(FEM)、有限差分法(FDM)、边界元法(BEM)和近场动力学(peridynamic)公式。尽管这些方法应用广泛,但对于复杂几何、任意域和强时变问题,它们可能变得计算密集。其精度还对网格细化、边界实现和空间变化性质的数值表征敏感,这进一步增加了计算成本。
鉴于纯数值方法的局限性,解析解在理解FGMs中的热传导方面仍具根本重要性。它们提供直接洞察温度演化和热行为的闭式表达式。Amiri Delouei等人开发了一套严格的解析解,用于各种边界条件下FGM圆柱结构中的热传导,主要假设幂律热导率并采用斯特姆-刘维尔理论和傅里叶变换等经典分析工具。其他互补的解析贡献进一步丰富了该领域:Andrade等人推导了完美粘结多层圆柱的二维非对称热传导解;Yaghoobi等人解析研究了非均匀热通量下的FGM圆柱;Ghasemi利用双相滞后框架研究了指数梯度FGM圆柱中的瞬态轴向热传导。
然而,现有解析解依赖于简化的材料梯度定律和规定的可分离边界条件以保持数学可处理性,这限制了其在任意性质梯度、任意时空边界变化和实际FGM应用中典型的不可分离热载荷系统中的应用。随着问题复杂度的增加,这些假设变得不足,促使研究人员采用替代的数据驱动框架。在此背景下,机器学习方法——特别是物理信息神经网络(PINNs)——已成为建模FGMs中热传导的有效工具。PINN将热传导方程等控制物理定律直接嵌入学习过程,确保物理一致的预测。Guo等人使用基于深度配点的PINN结合龙格-库塔时间积分求解FGM材料中的瞬态热传导;Xing等人将PINNs扩展至三维各向异性稳态传热;Jiang等人提出了热传导方程辅助的贝叶斯神经网络,将物理约束作为正则化项以改善有限且含噪声数据下的鲁棒性;Zobeiry等人将PINNs应用于增材制造中的耦合传导-对流传热。
PINNs在建模热传导方面已展现出强大能力,但纯物理驱动的公式常受收敛和优化困难的影响,这些问题在解呈现陡峭梯度、控制物理不完整或训练数据稀疏时尤为突出。此外,PINNs直接逼近状态变量,这限制了其在变化边界条件和材料梯度剖面间的泛化能力。深度算子网络(DeepONets)通过求解底层解算子克服了这一局限,实现对参数输入、功能输入和边界变化的高效泛化。然而,纯数据驱动的DeepONets在训练数据有限时可能违反物理一致性,这在非均质热系统中尤为明显。因此,混合物理信息DeepONets提供了一种自然折中,将算子学习与控制物理约束以及解析或高保真信息相结合。
现有文献揭示了FG材料热建模中未解决的挑战:双向材料非均匀性和周向变化的对流边界条件引起强模态耦合和陡峭空间梯度,这些效应很少在统一的解析或学习框架中被处理。 capturing such behavior requires both analytical structure and gradient-aware learning, motivating physics-informed operator-learning architectures that integrate analytically derived spectral representations or high-fidelity numerical data. In this study, a generalized Fourier–Sturm–Liouville formulation is derived for orthotropic FGMs under angularly varying Robin boundary conditions, yielding a Toeplitz-coupled spectral representation. The analytical structure is embedded into a gradient-enhanced Physics-Informed Deep Operator Network (gPI-DeepONet). Within this framework, gradient-based residuals enhance physics enforcement and improve accuracy while preserving operator-level generalization across functional variations. The proposed framework unifies spectral analysis and operator learning, enabling accurate, physically consistent prediction of spatio-temporal temperature fields for arbitrary material gradation and non-separable thermal loading.
2. Analytical formulation of the problemsThis section introduces an analytical model designed to address transient heat conduction phenomena within FGM hollow cylinders. The study examines various schemes of material gradation: (i) power law (-FGM), (ii) parabolic (-FGM), (iii) mixture (M-FGM), (iv) sigmoid (S-FGM), and (v) exponential (E-FGM). Initially, the spatial variability of thermal properties is defined, which subsequently facilitates the formulation of solutions to the governing heat equation under asymmetric conditions.2.1. Materials characterization of FGM hollow cylinderFig. 2 depicts the representation of a FGM hollow cylinder under transient thermal loading. The cylinder is characterized by an asymmetric configuration, with the inner and outer radii identified as and , respectively. A radial gradation is introduced, reflecting the systematic variation of thermal properties in both radial and circumferential directions—an uncommon feature in standard FGM designs. A homogenization approach is employed to represent the gradation, deriving effective properties from the underlying heterogeneous microstructure [51]. In the context of a ceramic-metal FGM, the effective property is represented using the widely accepted Voigt scheme as , where and represent the properties of the metal and ceramic, respectively, and and denote their respective volume fractions, with .Download: Download high-res image (465KB)Download: Download full-size imageFig. 2. Illustration of the two-dimensional cross-section of an asymmetric, infinitely extended functionally graded material (FGM) cylinder: (a) boundary conditions that are invariant with time; (b) boundary conditions that vary with time.Since constituent volume fractions can be tailored to meet specific design requirements, various distribution laws have been proposed, including power-law, exponential, mixing, and sigmoid functions. These commonly employed formulations are shown in Table 1.Table 1. Gradient laws and effective property expressions.Gradient lawEffective propertyGradient parameterRef.Power law[47], [48]Parabolic law[28]Exponential law[49]Mixture model[27], [30]Sigmoid model,[30], [50]2.2. Governing equation and its analytical solutionsThe governing equation for the transient heat conduction problem in polar coordinates () for an asymmetric FGM cylindrical can be formulated as follows:(1)The associated initial and boundary conditions governing the problem are expressed as:(2a)(2b)(2c)(2d)(2e)Here, and denote the radial and circumferential thermal conductivity profiles, respectively. Equations (2a)–(2b) prescribe generalized mixed (Robin-type) thermal boundary conditions at the inner and outer surfaces of the hollow cylinder. The coefficients and describe angularly varying convection at the boundaries, whereas and scale the radial conductive heat flux. The forcing terms and represent externally applied thermal loads that may vary in both space and time, thereby enabling the modeling of transient and circumferentially non-uniform thermal environments. Furthermore, the function denotes the initial temperature distribution.Certainly, obtaining a straightforward solution for Eq. (1), based on the complex nature of the generalized property function shown in Table 1, can be challenging and, in some cases, not achievable. Therefore, a layer-wise method has been adopted in the present study, where the material property functions can be decomposed as . This step-wise approximation preserves the essential effects of spatial grading while enabling analytical tractability, consistent with standard multilayer FGM modeling approaches [52], [53].Accordingly, the following simplifying assumptions are adopted to facilitate analytical tractability.•Temperature-independent homogenized properties are assumed to vary spatially according to the prescribed gradation law. The continuously graded medium is represented using a layer-wise approximation under ideal interfacial contact conditions. Continuity of temperature and heat flux is strictly enforced at each interface, and interfacial thermal resistance is neglected.•Macroscopic two-dimensional Fourier heat conduction is considered in an infinitely long cylinder. Axial heat transfer, internal volumetric heat generation, and microscale effects are neglected. This assumption is appropriate when continuum heat transfer dominates.•The radial and circumferential conductivities are defined independently but assumed equal, . This assumption is consistent with radially graded FGMs, where composition varies along while the local microstructure remains statistically uniform in the – plane. Although analytical solutions are possible for in the constant-coefficient case, equality is adopted here for notational simplicity and streamlined analysis.•Thermal radiation effects are neglected. Angularly dependent convective boundary conditions are considered, which introduce modal coupling in the angular direction and are handled through Fourier-series representation.Under these assumptions and using the above decomposition, Eq. (1) can be written for the layer as(3)and(4)Here, the condition enforces identical functional grading in the radial and circumferential directions, and denotes the layer-wise thermal diffusivity.To simplify the notation, the layer index is fixed throughout the following derivation, and the superscript is therefore omitted from all layer-wise quantities. Unless stated otherwise, all symbols are understood to correspond to the th layer.Here, convection induces circumferential variations (see Fig. 2). Accordingly, the temperature场 is expanded in angular eigenfunctions, as defined in Eq. 5.(5)In this expression, denotes the angular eigenfunction obtained from the circumferential eigenvalue problem, whose analytical form is standard and can be found in the classical reference by Hahn et al. [54].Because the problem is strongly non-homogeneous with time-dependent boundary conditions, classical separation of variables alone is insufficient. As a result, to render the solution analytically tractable, the finite integral transform technique is introduced, as summarized in Table 2.Table 2. Coordinate transformation sequences and Corresponding integral transform pairs.Transformation directionTransformed pairEq.6(a,b)7(a,b)Applying the angular and radial transformations defined in Eqs. (6a) and (7a), respectively, together with Eq. (5), yields the following ordinary differential equations:(8)Here,(9a)(9b)(9c)The eigenvalues and the corresponding eigenfunctions are obtained from the radial Sturm–Liouville eigenvalue problem; whose derivations are provided in Ref. [54].Because , , and depend on the angular coordinate , the solution of Eq. (6) is not straightforward. The -dependence is therefore treated using a complex Fourier representation.(10)Here, and represent the complex Fourier coefficients of the angularly varying material parameters and source terms, respectively, while denotes the modal amplitude of the temperature field.Substituting Eq. (10) into the boundary terms of Eqs.(2a)-(2b) and using Eq. (5) yields the resulting equations after applying standard properties of the complex Fourier series.(11)(12)From Eq. 2, the homogeneous form of the Robin boundary condition is written as(13)(14)The expression for is substituted into Eq. (9a) from Eqs.(11)-(12). The boundary terms are then obtained by combining Eqs.(13), (14).(15)If the coefficients and are independent of , Eq. (10) reduces by setting and assigning , , and , leading to the following formulations:(16)The formulation in Eq. (16) reduces to a simplified form that removes the angular dependence while retaining the essential material parameters.Again, setting and in Eq. (15) eliminates the diagonal terms, leaving only the convolution sums asThese terms represent mode-coupling effects that complicate the resulting scalar ODEs when is retained. Accordingly, in the corrected pipeline, the convolution sums are extracted from the scalar ODEs and imposed as a small matrix system with a Toeplitz structure in [55].Within this framework, the radial traces are expressed in terms of the modal coefficients, as defined in Eqs. (10)-(12), and summarized in Eq. 17.(17)The same representation applies at .To proceed, we define the diagonal trace operators , , , and are defined in Eq.(17), whose diagonal entries are , , , and , respectively. Similarly, the convolution structure arising from Eqs. (11)–(12) is encoded in Toeplitz matrices, as given in Eq. (18).(18)From Eq. (18), the boundary relations defined in Eqs. (11)-(12) can be written in the compact form(19)Accordingly, from Eq. (8), the interior modal equations remain uncoupled as(20)Combining Eqs. (19) and (20), the governing differential–algebraic system can be written in the compact form given in Eq. (21).(21)Where,Since the algebraic constraint matrix (see in Eq. (21)) is not guaranteed to be square or invertible, a generalized inverse is employed to construct a consistent particular solution. Choose any Moore–Penrose pseudoinverse of [56]. Then, from Eq.(21b), a particular solution for is(22)Then the homogeneous solution can be defined by(23)Substituting Eq. (23) into Eq. (21a) transforms the constrained system into a homogeneous problem, which is solved using the variation-of-constants method, leading to Eq. (24).(24)Substituting into Eq. (23), we find the complete solution of (25)Here, denotes the vector of integration constants determined from the initial condition, subject to the algebraic consistency requirement . The projection operator extracts the dynamically admissible component of the solution by projecting onto the null space , thereby ensuring compatibility with the algebraic constraints.If the are constant parameters then the governing equation (Eq. 8) is just a scalar and whose solution is defined by Eq.(26)(26)Which is simplified version of Eq. (25). Here, is given by Eq. (16)Using the inverse transforms defined in Eqs. (6b) and (7b), the final solution obtained from Eq. (25) can be expressed in the following vector form:(27)Here,For completeness, the corresponding scalar representation is given by(28)For brevity, the layer index was previously suppressed. We now reinstate the superscript and enforce continuity at the interface , .(29a)(29b)3. Hybrid gradient-enhanced physics-informed DeepONet (gPI-DeepONet) formulationThe preceding analytical solution accurately describes the temperature distribution in a two-dimensional asymmetric FGM cylinder. However, its extension to more complex configurations remains challenging. In addition, rapid spatial variations in material properties render the governing partial differential equation on the plane increasingly stiff, leading to sharp solution gradients. In such generalized settings, an analytical solution alone becomes difficult to obtain and inefficient to deploy. Moreover, design and optimization tasks often require repeated evaluations under varying and uncertain material and boundary inputs, for which direct analytical evaluation becomes impractical. To improve solution fidelity under such conditions, we employ a gradient-enhanced gPI-DeepONet (see Fig. 3), in which gradient information is incorporated into the loss function.Download: Download high-res image (723KB)Download: Download full-size imageFig. 3. Schematic of the proposed stacked hybrid analytical and gradient-enhanced multi-parametric gPI-DeepONet framework for two-dimensional asymmetric heat conduction analysis in FGMs. The framework combines analytically generated synthetic data with physics-informed learning to model heat transfer in heterogeneous domains with spatially varying material properties.This augmentation examines whether gradient-based regularization, together with analytical solutions or high-fidelity numerical data when necessary, improves accuracy. Accordingly, a hybrid training framework is adopted by integrating physics-based and data-driven loss terms. The resulting performance is assessed in the upcoming section.For notational brevity, the governing heat Eq.(1) can be expressed in the compact operator form(30)where denotes a symbolic operator representing the simplified form of the heat equation given in (1). The operator can be recast as(31)Here, denote the physical input functions supplied to the branch networks, defining the operator mapping . We approximate the solution operator with a stacked hybrid gPI-DeepONet, , where collects the trainable parameters. The approach follows the operator-learning framework of Lu et al. [57]. In the present formulation, the input functions are organized into three branch inputs asThe hybrid network output is determined by the branch and trunk networks. More precisely, the expression for the network prediction of a function evaluated at a spatial coordinate can be given by(32)In Eq. (32), the trunk network maps the spatial–temporal coordinate to the basis functions , while each branch network produces a vector of coefficients from its corresponding input. The product represents element-wise multiplication across the three branch outputs (Hadamard product), which modulates the spatial basis learned by the trunk. The final output is obtained by summing over , yielding the predicted temperature at .Then the data-driven loss for hybrid gPI-DeepONet can be obtained from the MSE loss as below(33)Where M is some sample input function. The physics-driven loss can be written as(34)Here,Moreover, the residual expressions associated with the loss functions in Eq. (34) are summarized in Table 3.Table 3. PI-based residuals used for the hybrid DeepONet.Residue / VariablesPI-based Residuals of DeepONetTo evaluate the proposed gPI-DeepONet, we consider the configuration shown in Fig. 2(a). The input profiles , , and (see Table 4) are prescribed by the underlying physical problem. The training datasets are generated by sampling these boundary-condition functions from Gaussian random fields (GRFs), which enables the construction of diverse yet physically consistent realizations. The time-dependent heat-flux distribution applied on the outer surface is modeled as a two-dimensional GRF defined over and . To enforce angular periodicity and ensure smooth temporal evolution, a separable spatio-temporal covariance kernel is employed [58],(35)This choice yields smooth yet nontrivial variations in both the angular and temporal directions. Each realization of is discretized on an equispaced grid of points. The resulting matrix is flattened into a vector of length and supplied to the first branch network. The circumferentially varying convection coefficient and the boundary temperature distribution are modeled as angular GRFs. These are obtained from the same covariance structure by setting in Eq. (36). Each realization is discretized on equispaced angular points and encoded as a -dimensional vector. These vectors serve as inputs to the second and third branch networks, respectively. Each input function is represented by 100 realizations generated from the Gaussian random field. The trunk network employs spatio-temporal collocation points generated via Latin hypercube sampling (LHS) to ensure uniform domain coverage [59]. To further stabilize optimization, analytical solution fields obtained from Eqs. (28)–(29) using Maple are incorporated as an auxiliary data-based loss. These labeled data enhance the network’s physical consistency, leading to improved accuracy and generalization to previously unseen functional inputs. To effectively leverage this information, the model is trained for optimization steps using a hybrid Adam–L-BFGS strategy with an initial learning rate of . The loss weights are chosen empirically to balance the relative contributions of the PDE, gradient, boundary, and initial-condition residuals, ensuring stable training and consistent convergence (see Fig. 5). Based on a sensitivity analysis of network depth and width (see Fig. 4), all branch and trunk networks are implemented with 9 hidden layers and neurons per layer. This configuration provides a balanced trade-off between prediction accuracy and training stability. The corresponding training convergence is quantified by the evolution of the error with respect to training epochs (Fig. 5). The error is initially large due to insufficient training. As optimization proceeds, it decreases steadily and converges to a minimal value, indicating stable and effective learning. All computations were performed on a single NVIDIA Tesla V100 GPU, with a wall-clock training time of approximately 10 h for the reported configuration. Once trained, the model predicts the temperature field for previously unseen functional inputs in about . In contrast, a conventional COMSOL solver requires approximately for the same setup using an extra-fine mesh with around grid points. It is very important to mention that the same architectural and training configuration is also employed for another case shown in Fig. 2(b).Table 4. Boundary conditions, associated coefficients, and parameters for time independent B.C.s.Types of B.C.sCoefficientsParametersDirichlet (inner surface)?Convection (outer top surface)??? (given in Ref.[62])Neumann (outer bottom surface)?Download: Download high-res image (86KB)Download: Download full-size imageFig. 4. Combined error versus Network depth (with neurons per layer), showing a systematic reduction in error with increasing depth and saturation at larger depths.Download: Download high-res image (191KB)Download: Download full-size imageFig. 5. The error as a function of training epochs is shown. The error is Initially large due to insufficient training, decreases steadily as learning progresses, and converges to a minimal value upon stabilization.Direct experimental data are unavailable due to the geometric complexity and spatio-temporal boundary variations considered. Moreover, for such generalized configurations, closed-form analytical solutions are not always attainable. In this context, high-fidelity numerical simulations serve as a reliable surrogate for physical measurements and provide a principled basis for data-assisted learning. Accordingly, two controlled sources of supervision are employed: COMSOL-based finite-element (FEA) solutions and analytical solution fields derived from Eqs. (28)–(29) for cases where analytical expressions remain tractable. Fig. 6 compares the error across five approaches: DeepONet, PI-DeepONet, gPI-DeepONet, gPI-DeepONet (Hybrid–FEA), and gPI-DeepONet (Hybrid–Analytical). The purely data-driven DeepONet exhibits the largest error, as it relies solely on data correlations and lacks explicit physical constraints, which limits extrapolation to unseen functional inputs. Enforcing the governing equations in PI-DeepONet significantly reduces the error by restricting the solution space to physically admissible states. The gradient-enhanced gPI-DeepONet further improves accuracy by incorporating derivative-based constraints that enhance local consistency, leading to an error reduction of approximately relative to DeepONet. Additional improvements are obtained through external supervision. In the hybrid FEA variant, high-fidelity numerical solutions provide global solution-level guidance, reducing accumulated model bias and achieving an error reduction of about . The lowest error is obtained with gPI-DeepONet (Hybrid), where analytical solutions supply exact, noise-free supervision, most effectively reinforcing the learned operator and yielding an error reduction of approximately relative to the baseline DeepONet. Along with the error bars, Fig. 7 shows the corresponding radial and angular temperature distributions compared against the analytical solutions. The same trend is observed: the purely data-driven DeepONet exhibits larger deviations and stronger spatial variation, whereas the physics-informed and hybrid models progressively reduce these variations. The gPI-DeepONet (Hybrid) shows the smallest deviations in both radial and angular profiles, indicating the highest consistency with the reference solution.Download: Download high-res image (118KB)Download: Download full-size imageFig. 6. Model performance comparison. The error across DeepONet variants shows a systematic reduction as physical constraints and data-based supervision are incorporated, with the hybrid gPI-DeepONet achieving the lowest error.Download: Download high-res image (255KB)Download: Download full-size imageFig. 7. Temperature distribution at for Case–1. (a) radial and (b) angular temperature profiles comparing DeepONet variants with the exact solution. The hybrid gPI-DeepONet exhibits the smallest deviation in both cases.4. Numerical examples and discussionsThis section articulates the findings of the current study and their implications. Initially, the convergence behavior of the analytical solutions was scrutinized to ensure the stability of the solution. A verification study was subsequently conducted by comparing the present results with those documented in prior research [22]. Following this, two illustrative cases were examined. Case 1 addresses the time-independent scenario, whereas Case 2 considers the time-dependent case. Together, these examples demonstrate the versatility and effectiveness of the proposed method in handling diverse boundary-condition configurations. Ultimately, the proposed analytical solution is compared with the machine-learning framework gPI-DeepONet, showing excellent agreement with the analytical results.4.1. Convergence studyTheoretical modeling suggests that the accuracy of the solution is substantially affected by the number of terms retained in the Fourier series expansion presented in (27), (28). While theoretically, an infinite number of terms could yield the precise solution, practical applications are constrained by computational limitations. Therefore, the Fourier series must be truncated to a finite number of terms, requiring an optimal balance between accuracy and computational efficiency. Fig. 8 shows the variation of error with increasing truncation levels of and over the time interval – for Case–1, evaluated at the spatial location . The corresponding model details are provided in Section 4.5. As evident from Fig. 8, the error attains its minimum when the retained series terms lie within the ranges – and –, respectively.Download: Download high-res image (241KB)Download: Download full-size imageFig. 8. The convergence study of the proposed framework for and . The error surface is plotted as a function of the Fourier expansion terms . A systematic reduction in error is observed with increasing Fourier parameters, demonstrating the stability and convergence of the analytical solution.Beyond these thresholds, the error is not influenced by higher-order terms, thus validating the convergence of the solution.4.2. Effect of number of layers on temperature convergenceThe influence of layer count on the temperature response for Case–1 is examined in Fig. 9. Fig. 9(a) shows the response at and , while Fig. 9(b) corresponds to at . In both cases, the number of layers is varied from 5 to 30. The corresponding problem setup and model details are provided in Section 4.5.Download: Download high-res image (170KB)Download: Download full-size imageFig. 9. Effect of layer count on the temperature response for Case–1: (a) M-FGM at ; and (b) E-FGM at . In both cases, the results converge for layer numbers exceeding approximately 20.Increasing the number of layers enhances accuracy by allowing a finer representation of spatial heterogeneity and more precise evaluation of the field variables. A higher layer匀层 count better resolves material-property variations and reduces discretization errors, resulting in a smoother and more accurate temperature field. The results additionally indicate that, beyond approximately 20 layers, the predicted temperature becomes effectively insensitive to further mesh refinement. The solution thus exhibits convergence, and the variation in temperature with additional layers becomes negligible. This behavior indicates that numerical saturation has been achieved and that the adopted discretization is sufficiently fine to resolve the dominant thermal gradients and material inhomogeneities. Consequently, further increasing the number of layers yields no practically significant gain in accuracy. Instead, it only increases computational cost, confirming the robustness of the analytical formulation at moderate discretization levels.4.3. Model verifications against analytical solutionsThe unsteady problem is complicated by generalized property functions, which necessitate linearization to obtain an analytical solution. Moreover, analytical solutions for generalized two-dimensional orthotropic FGM cylinders under unsteady heat conduction have not been reported in the literature. As a result, direct validation of the present formulation against existing studies is not feasible. Nevertheless, recent studies explore a layer-wise 2D homogeneous model for heat conduction with both time-dependent and time-independent boundary conditions [22]. For the purpose of verification, we substitute the orthotropic medium in the linearized unsteady problem with a finite, layer-wise isotropic surrogate, aligned with the methodology in ref.[22]. The problem consider a three-layer isotropic cylindrical configuration with , , , , and , , . The outer surface is subjected to a prescribed heat flux , the remaining surface is thermally insulated, and the initial temperature is zero throughout the domain. This setup enables a juxtaposition of the present solution with the reference, with a side-by-side evaluation illustrated in Fig. 10. Specifically, Fig. 10(a) displays the time-independent boundary-condition scenario (), while Fig. 10(b) investigates the time-dependent scenario (). In both cases, the predicted results agree closely with the benchmark of Ref. [22], with discrepancies within 4% (Fig. 10(c), thereby validating the mathematical model.Download: Download high-res image (512KB)Download: Download full-size imageFig. 10. The current study compares the time-varying temperature distribution with Ref. [22]: a) time-independent heat flux; b) time-dependent heat flux; and c) representative pointwise error comparison at for both time-independent and time-dependent cases.4.4. Model verifications against numerical solutionsAn asymmetric, two-dimensional transient thermal model was developed using the commercial finite element software COMSOL Multiphysics [60]. The computational domain consists of a hollow cylindrical annulus with nondimensional inner and outer radii of 1 and 2, respectively. A spatially varying convective boundary condition, represented in Fourier form, is imposed on the upper half of the outer boundary, with the angular dependence of given in Table 1. The lower half of the outer boundary is subjected to a circumferentially nonuniform heat flux. In contrast, an angularly varying prescribed temperature is imposed along the inner boundary, as shown in Figs. 1(a) and 9(a).A mesh-independence analysis was conducted to confirm that the numerical solution is insensitive to spatial discretization (see Fig. 9(b)). The simulations were performed using the COMSOL Multiphysics heat-transfer module with two-dimensional unstructured triangular meshes, ranging from extremely coarse to extra-fine resolutions. Convergence was evaluated at and , at for the E-FGM and at for the M-FGM. With systematic refinement, the temperature fields converged monotonically, and further refinement beyond the “fine” mesh produced negligible changes. Fig. 9(c) and (d) compare the radial and circumferential temperature distributions from the analytical solution (Eqs. (28)–(29)) with the corresponding finite element results. Excellent agreement is observed between the analytical and numerical predictions over the entire domain. The discrepancies remain within approximately (see Fig. 11(e)), thereby validating both the analytical formulation and the numerical implementation.Download: Download high-res image (775KB)Download: Download full-size imageFig. 11. (a) mesh configuration and applied boundary conditions for Case 1; (b) convergence behavior with increasing Element count; (c) radial temperature distribution; (d) circumferential temperature distribution compared with COMSOL Multiphysics results; and (e) pointwise relative error distribution for the M-FGM along the radial and angular direction. Overall, the analytical and finite-element solutions show excellent agreement, with errors below across the domain.4.5. Case 1: time-independent boundary conditionsTo evaluate practical reliability, we consider a generalized configuration consisting of a FG hollow cylinder subjected to asymmetric and unsteady thermal loading. This configuration is analogous to thermal ablation scenarios, marked by spatial heterogeneity and non-uniform boundary conditions due to localized energy deposition and perfusion. The present study does not aim to simulate biomedical outcomes. Rather, it is conceptually motivated by the laser–tissue interaction literature [61], which reports circumferentially varying thermal inputs, transient internal temperature evolution, and convective effects. These elements align closely with the thermal features modeled herein and provide a relevant analogue for the formulation of asymmetric boundary conditions. Moreover, reported non-uniform collagen architecture and directional heat reactivity motivate the use of a heterogeneous FGM framework. This correspondence reinforces the physical relevance of the selected material gradation and boundary loading strategy. For illustration, we examine an asymmetric hollow cylinder with inner radius and outer radius . The geometric configuration and boundary conditions are shown in Fig. 2(a) and the corresponding boundary terms and parameters are summarized in Table 4. The inner surface () is prescribed with a circumferentially varying temperature. The outer表面 is divided into两个区域:上部段()受到空间变化的对流,而下部段()暴露于周向变化的热通量。这些条件与上述热消融背景一致。4.5.1. Time–temperature evolution in asymmetric FGM cylindersThis section examines the transient thermal response of asymmetric FGM cylinders, as illustrated in Fig. 12(b). At the initial stage, the temperature field remains close to its baseline minimum. As heat conduction proceeds, thermal energy diffuses throughout the domain, leading to a gradual rise in temperature. A steady state is eventually reached, characterized by spatially stabilized temperature distributions. Among the configurations considered, the M-FGM exhibits the slowest overall temperature increase. This behavior is directly linked to its non-monotonic thermal conductivity profile (Fig. 12(a)). As the radial coordinate increases, the conductivity of the M-FGM initially increases and subsequently decreases. While this evolution accelerates heat diffusion at early times, it progressively impedes it at later stages. Consequently, a radially regulated thermal resistance develops, delaying the attainment of thermal equilibrium, as observed in Fig. 12(b). At the same time, the enhanced effective conductivity at larger radii improves long-term heat transfer and reduces internal temperature gradients. Such a moderated transient response is advantageous for controlled-heating applications, as it reduces thermal shock, material degradation, and structural failure under non-steady thermal loading. In contrast, the P1-FGM displays the most rapid temperature rise and reaches steady state significantly earlier than the other configurations. This response arises from its spatially uniform increase in thermal conductivity, density, and heat capacity, which collectively enhance the thermal diffusivity across the domain. The resulting reduction in thermal resistance promotes efficient radial and circumferential energy redistribution, leading to rapid homogenization of the temperature field. The E-FGM, P2-FGM, and S-FGM configurations also exhibit relatively fast transient responses, although their temperature profiles tend to converge at later times. The early dominance of the P1-FGM highlights its superior transient efficiency. This characteristic is particularly beneficial for thermal ablation applications, where rapid and spatially predictable heating is essential to achieve precise thermal damage while minimizing collateral effects. The combination of a rapid thermal response and controlled spatial temperature distribution suggests that P1-FGM may be suitable for ablative technologies. Such applications include minimally invasive cancer therapies and localized tissue coagulation.Download: Download high-res image (251KB)Download: Download full-size imageFig. 12. (a) the spatial variation of material properties along the and direction, where thermal conductivities ( and ) are shown as a representative example for simplicity among the considered properties, such as specific heat () and density (); (b) Time–temperature distribution for various FGMs in the – plane at .4.5.2. Spatial temperature distributions for various FGMsFig. 13 provides a comparison of the spatio-temporal temperature fields for P1-, P2-, E-, S-, and M-FGMs under conditions of asymmetric heat conduction. At early times (; Fig. 13(a–b)), the radial temperature profiles are largely concave with a minimum near the core. This behavior results from inward diffusion from the isothermal inner wall and heating at the outer wall due to mixed convection and prescribed heat flux. At this stage, the M-FGM exhibits the most attenuated temperature peaks and the smoothest radial gradients. This indicates a more rapid and spatially uniform heat distribution, which is attributed to its higher effective thermal conductivity. The non-monotonic radial conductivity profile of M-FGM enhances lateral heat spreading, thereby suppressing the development of pronounced circumferential temperature variations at early times. An exception is observed for the E-FGM, which develops localized angular maxima at approximately , , and . These hot spots originate from angularly non-uniform convection at the outer surface. At early times, thermal diffusion is insufficient to smooth these localized surface inputs, allowing circumferential non-uniformities to persist. As time progresses (Fig. 13(c–h)), the thermal diffusion length, , approaches the wall-to-wall length scale. Circumferential temperature variations then become more pronounced and, in some regions, exhibit abrupt changes. These sharp angular transitions arise from non-uniform outer-wall convection combined with radial and angular varying thermal properties of the FGMs. Regions with higher local convective resistance retain heat longer, whereas neighboring regions with higher effective diffusivity redistribute heat more efficiently. This contrast in local thermal response produces steep circumferential temperature gradients and localized hot spots. P1- and P2-FGMs progressively accumulate thermal energy near the outer boundary (Fig. 13(g)). As a result, they reach the highest late-time temperatures and develop the steepest radial gradients. In contrast, E- and S-FGMs maintain moderate gradients due to their smoother conductivity variation. The M-FGM remains comparatively uniform with reduced peak values. This behavior underscores its strong homogenizing effect, even under asymmetric boundary forcing. Late-time angular sections (Fig. 13(h)) reveal clear circumferential asymmetry and distinct hot spots. These features are most pronounced in the E-FGM and P1-FGM, indicating preferential heat-flow paths dictated by the gradation law and amplified by non-uniform convection. From a practical standpoint, M-FGM is advantageous for applications requiring spatially uniform heating, whereas P-type gradations favor heat retention near the heated boundary. E-FGM can be tailored for angle-selective heating. From a mechanistic perspective, the response is governed by two coupled effects. First, angularly non-uniform boundary forcing introduces localized thermal loading. Second, radial stratification of thermal conductivity and diffusivity controls how this heat is redistributed within the domain. In P1- and P2-FGMs, this interaction promotes prolonged heat retention near the outer wall, leading to temperatures relevant to thermal ablation by . In contrast, the elevated effective conductivity of M-FGM homogenizes the thermal field and suppresses excessive temperature overshoot. Overall, the coupled FGM–BC architecture acts as a spatial regulator of heat flow. It enables material-mediated control over radial dose accumulation and angular concentration, features that may be relevant to thermal ablation and selective, minimally invasive thermal therapy. The contour maps in Fig. 14, Fig. 15, Fig. 16, Fig. 17 provide a concise visualization of these mechanisms and their temporal evolution.Download: Download high-res image (1MB)Download: Download full-size imageFig. 13. Spatial temperature distribution in – plane: (a) – at , 0.9b) – at , (c) – at , (d) – at , (e) – at , (f) – at , (g) – at , (h) – at ; where – and – plots are taken at and , respectively. The result illustrates that M-FGM suppresses temperature localization, while P- and E-type gradations amplify circumferential hot spots under asymmetric heating.Download: Download high-res image (275KB)Download: Download full-size imageFig. 14. Contour plot of temperature distribution of P2-FGM in the – plane: (a) , (b) , (c) , (d) , and (e) . With increasing time, the -FGM Develops enhanced circumferential asymmetry and localized temperature intensification near the outer wall.Download: Download high-res image (271KB)Download: Download full-size imageFig. 15. Contour plot of temperature distribution of E-FGM in the – plane: (a) , (b) , (c) , (d) , and (e) . The E-FGM exhibits time-dependent redistribution of thermal energy along preferred angular paths, reflecting the bi-directional influence of exponential gradation under asymmetric loading.Download: Download high-res image (283KB)Download: Download full-size imageFig. 16. Contour plot of temperature distribution of M-FGM in the – plane: (a) , (b) , (c) , (d) , and (e) . M-FGM gradation promotes efficient radial–circumferential heat redistribution, preventing localized temperature build-up even at late times.Download: Download high-res image (270KB)Download: Download full-size imageFig. 17. Contour plot of temperature distribution of S-FGM in the – plane: (a) , (b) , (c) , (d) , and (e) . S-FGM provides intermediate thermal regulation, bridging localized and homogenized heat-transfer behaviors.4.6. Case 2: time-dependent boundary conditionsThis investigation considers a functionally graded hollow cylinder under asymmetric and time-dependent thermal boundary conditions. Such conditions are relevant to advanced nuclear components, including fuel cladding, reactor piping, and core structural elements [24], [63], [64], [65]. This study adopts a conceptual and parametric formulation that remains physically representative of these conditions. The approach obtained here allows systematic analysis of material gradation, circumferential asymmetry, and transient thermal effects. The objective is to capture general thermal response behavior rather than reproduce a specific reactor design. Consequently, direct validation against an actual nuclear configuration is not pursued, as this would require design-dependent material properties and operating conditions beyond the scope of this work.In practical nuclear systems, the inner surfaces of cylindrical components are subjected to strong convective cooling due to high-velocity coolant flow. In contrast, the outer surfaces experience spatially non-uniform heating caused by anisotropic neutron flux and gamma irradiation. During transient events, such as power ramps or loss-of-coolant accidents (LOCAs), these thermal boundary conditions become strongly time-dependent and directionally biased. Such behavior introduces pronounced coupling across both spatial and temporal scales. To capture these effects, the present framework employs parabolically distributed, time-varying heat fluxes together with angularly dependent convection. This formulation represents asymmetric and transient heat transfer while remaining computationally feasible. The functional gradation of thermal properties is motivated by emerging composite cladding concepts, including SiC–Zr and U3Si2-based designs. These architectures aim to enhance heat dissipation, suppress thermal hot spots, and mitigate stress localization under extreme operating conditions. By combining anisotropic material gradation with transient boundary excitation, the proposed model provides a physically realistic yet analytically manageable platform. It supports the evaluation of thermal performance, informs material selection, and aids the structural design of next-generation nuclear systems. For clarity, the geometric configuration and boundary conditions are illustrated in Fig. 2(b), while representative boundary terms and parameter values are summarized in Table 5.Table 5. Boundary conditions, associated coefficients, and parameters for time dependent B,C,s.Types of B.C.sCoefficientsParametersDirichlet (inner surface) (given in Ref.[62])Convection (outer top surface)NoneNeumann (outer bottom surface)?4.6.1. Time-temperature evolution in asymmetric FGM cylinders with time-dependent boundary conditionsFig. 18 illustrates the spatiotemporal thermal response of the P1-FGM under periodic thermal loading. The response exhibits a clear oscillatory character that intensifies with increasing radial distance. Such behavior is consistent with cyclic temperature fluctuations observed in nuclear cladding and reactor piping during load-following operations or transient power excursions. At the outer radius (), the pronounced amplification of temperature oscillations indicates heightened thermal sensitivity of the boundary layer directly exposed to heat flux. This region is therefore particularly vulnerable to thermal fatigue and creep damage. In contrast, the response at is significantly damped. This attenuation highlights the intrinsic buffering capability of the graded architecture, wherein material inhomogeneity suppresses oscillatory extremes and stabilizes the thermal field. The coexistence of outer-surface amplification and inner-domain attenuation captures the essential role of functional gradation in redistributing transient thermal loads. Consequently, FGMs emerge as promising candidates for next-generation nuclear claddings, where durability under cyclic thermal conditions is a critical requirement.Download: Download high-res image (232KB)Download: Download full-size imageFig. 18. Time-temperature distributions of P1-FGM at various locations on plane and at different radial positions. P1-FGM shows increasing thermal amplitude with radius under transient excitation.4.6.2. Spatial temperature distribution for various FGMs in time-dependent boundary conditionsFig. 19 depicts the spatiotemporal thermal response of cylindrical nuclear components under irregular, time-dependent thermal loads. In the initial period ( s; Fig. 19(a)), all FGMs exhibit a slight increase in temperature near the inner surface, followed by a more pronounced gradient approaching the outer wall. This profile exemplifies the interplay between energy extraction by the coolant at the bore and the persistence of external heating. Among the materials examined, M-FGM exhibits the lowest temperature field, acting as a passive thermal buffer, whereas P2-FGM shows greater heat-absorption capacity, which is beneficial for regulated thermal-storage applications. During the transient phase (– s; Figs. 19(c–f)), P1-FGM exceeds the other gradation types, displaying higher radial temperatures combined with distinct circumferential maxima. These characteristics resemble localized flux hotspots that may arise along cladding or piping segments during power maneuvers. At later times (; Figs. 19(g–h)), P1-FGM reaches temperatures of approximately , confirming its resilience under sustained, non-uniform loading. In contrast, M-FGM continues to limit the overall temperature rise, reinforcing its effectiveness as an insulating or sacrificial layer. Collectively, these findings suggest that FGMs with steeper radial gradation (P1- and P2-types) are better suited for regions requiring active heat absorption and localization during transient conditions. In contrast, diffusively graded materials such as M-FGM provide superior thermal stability where limiting temperature rise is critical. This material-specific behavior supports the zoning of claddings, liners, or shields using tailored FGMs to manage spatially and temporally varying reactor heat loads more safely and efficiently. Lastly, the temporal development of temperature fields in the – plane is illustrated in the contour plots of Figs. 20(a–e). These visualizations complement the quantitative results by showcasing the progression of thermal penetration and localized heating over time, offering a concise and comprehensive depiction of the transient behavior.Download: Download high-res image (1020KB)Download: Download full-size imageFig. 19. Spatial temperature distribution in – plane: (a) – at , (b) – at , (c) – at , (d) – at , (e) – at , (f) – at , (g) – at , (h) – at , (i) – at , and (j) – at ; where – and – plots are taken at and , respectively. Overall, P1- and P2-FGMs exhibit enhanced transient heat accumulation. E-FGM promotes angular redistribution, S-FGM provides a moderated thermal response, and M-FGM consistently suppresses peak temperature rise.Download: Download high-res image (281KB)Download: Download full-size imageFig. 19. (Continued...)Download: Download high-res image (699KB)Download: Download full-size imageFig. 20. Contour plot of temperature distribution of -,-,E-,M- and S-FGM under time dependent boundary conditions in the – plane: (a) , (b) , (c) , (d) , and (e) . Distinct gradation profiles lead to markedly different temporal growth patterns, revealing material-dependent delays and accelerations in temperature evolution.4.7. Comparison of analytical and gPI-DeepONet resultsThe accuracy and reliability of the gPI-DeepONet are evaluated through a comparative analysis of its results with analytical solutions depicted in Fig. 21. These analytical solutions serve as a rigorous benchmark for evaluating the gPI-DeepONet’s ability to capture essential physics and reproduce exact solutions in simple geometric settings. Consequently, such a comparison is critical for evaluating the model’s performance and validating its applicability. Moreover, the comparison results demonstrate the effectiveness of the gPI-DeepONet in scenarios where analytical solutions are available. This validation establishes a foundation for its application to more complex problems for which such solutions do not exist.Download: Download high-res image (673KB)Download: Download full-size imageFig. 21. Temperature distributions in the P1-FGM cylindrical domain at two time instances: (a) and (b) . Left panels show analytical solutions, and right panels display gPI-DeepONet predictions. The close agreement, supported by the low pointwise errors shown in the right panels, Demonstrates accurate prediction of transient thermal behavior in graded composites.5. Future outlooks on experimental integration of gPI-DeepONetIn the present study, gPI-DeepONet is trained on noise-free analytical or high-fidelity numerical data under physics-informed enforcement of the governing heat-conduction equation. This establishes a stable operator-learning framework under controlled conditions. The next step is to extend the formulation to learn the solution operator directly from experimental measurements while preserving strict physics-based constraints. In future work, gPI-DeepONet will be integrated with transient temperature measurements obtained at discrete space–time locations . Embedded micro-thermocouples will deliver pointwise temperature data within the domain and at its boundaries. Infrared thermography will provide time-resolved surface temperature fields . These measurements will be fused with the physics-constrained operator learner by augmenting the training objective with a measurement likelihood term. The existing physics-informed losses (PDE residual and BC/IC residuals) will therefore be complemented by additional data-consistency terms [66].(36)Here, denotes the gPI-DeepONet mean prediction, and represents the measurement uncertainty. A Gaussian noise model will be adopted, as calibrated sensor errors are typically dominated by random fluctuations that are well approximated by a normal distribution. Accordingly, the variance may be treated as a constant estimated from repeated thermocouple measurements under a homoscedastic assumption. Alternatively, it may be modeled as a space–time-dependent term to represent spatially varying uncertainty in infrared thermography. In this case, the heteroscedastic behavior will be implemented through a weighted likelihood formulation. If required, epistemic uncertainty will be quantified using Monte Carlo dropout or Bayesian last-layer inference, enabling predictive intervals in regions with sparse sensing.Validation will proceed in three stages. First, hold-out testing will be performed by excluding a subset of thermocouple or infrared data during training and using these measurements for prediction assessment. Second, predictions will be compared with independent FEM simulations under identical boundary and initial conditions. Third, quantitative metrics will be reported, including RMSE, relative error, and maximum deviation over the held-out measurements. In addition, uncertainty coverage will be evaluated by measuring the fraction of held-out observations that fall within the predicted confidence intervals. Through this extension, gPI-DeepONet further consolidates its hybrid physics–data structure by integrating experimental measurements within the governing-equation framework. This strengthens predictive robustness and supports reliable parameter identification under realistic experimental settings.6. ConclusionsThis investigation introduces a unified analytical and physics-informed machine-learning framework (gPI-DeepONet) for transient heat conduction in asymmetrically graded cylinders. The problem is challenging due to the coupling of spatial inhomogeneity and time-dependent boundary conditions. Using a Fourier–Sturm–Liouville spectral formulation, an exact closed-form solution is derived to resolve radial and circumferential property variations in orthotropic FGMs. To clarify the role of spatial gradation, five representative architectures are examined: power-law (P1, P2), exponential (E), sigmoid (S), and mixture (M). Each exhibits distinct thermal behavior, revealing previously unreported phenomena. The M-FGM architecture exhibits delayed and diffused thermal propagation, making it suitable for thermal protection in nuclear fuel cladding. In contrast, P-type FGMs display rapid and spatially uniform heating, which is advantageous for localized thermal ablation. The alternative gradation types—, E, and S—facilitate tunable transitions between these two extremes, providing flexible design methodologies tailored to application-specific performance. Two case studies demonstrate the framework’s predictive capability. The first investigates non-uniform heating in biological tissue, highlighting controlled hot-spot localization. The second analyzes transient thermal loading in nuclear fuel rods, showing how material gradation reshapes temperature fields and moderates thermal gradients. The analytical solution is further integrated with a gPI-DeepONet, enabling efficient operator-level prediction across complex domains while preserving strict physical consistency. The resulting hybrid framework establishes a scalable and design-oriented platform for transient thermal analysis in advanced materials systems relevant to biomedical, aerospace, and energy applications.Future developments will relax the assumption of temperature-independent properties and homogeneous layers. This will enable temperature-dependent behavior and intra-layer power-law gradation. Uncertainty quantification will be incorporated to assess material and boundary variability. The framework will also be extended to fully three-dimensional formulations.CRediT authorship contribution statementPalash Das: Writing – review & editing, Writing – original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Md Ashraful Islam: Writing – review & editing, Supervision. Fahim Rahman Rijvi: Writing – review & editing, Data curation. Mohammad Rafiqur Rashid: Writing – review & editing.